From 31b27926b955002751377909fe9cd82dd9415112 Mon Sep 17 00:00:00 2001
From: mariosasko <mariosasko777@gmail.com>
Date: Wed, 31 Mar 2021 02:59:15 +0200
Subject: [PATCH 01/26] Fix references in colabs

---
 docs/source/scripts/convert_doc_to_notebooks.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/docs/source/scripts/convert_doc_to_notebooks.py b/docs/source/scripts/convert_doc_to_notebooks.py
index 4ec1608d..f518d3e2 100644
--- a/docs/source/scripts/convert_doc_to_notebooks.py
+++ b/docs/source/scripts/convert_doc_to_notebooks.py
@@ -272,7 +272,7 @@ def convert_math(text):
 def convert_anchor(text):
     """ Convert text to an anchor that can be used in the notebook."""
     anchor_name = _re_anchor_section.search(text).groups()[0]
-    return f"<a id='{anchor_name}'></a>"
+    return f"<a name='{anchor_name}' id='{anchor_name}'></a>"
 
 
 ###################################

From c7f23d6b66cd45082e8d04870621119e66e3ba80 Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Wed, 31 Mar 2021 16:56:14 +0200
Subject: [PATCH 02/26] Rework examples

---
 docs/source/examples/tfidf_example.rst | 105 ++++++++++++
 examples/dataset_example.py            |  25 ---
 examples/experiment_example.py         | 150 -----------------
 examples/ner_example.py                | 213 -------------------------
 examples/ner_pipeline.py               | 191 ----------------------
 examples/notebooks/img/hook.png        | Bin 8521 -> 0 bytes
 examples/vectors_example.py            |  26 ---
 7 files changed, 105 insertions(+), 605 deletions(-)
 create mode 100644 docs/source/examples/tfidf_example.rst
 delete mode 100644 examples/dataset_example.py
 delete mode 100644 examples/experiment_example.py
 delete mode 100644 examples/ner_example.py
 delete mode 100644 examples/ner_pipeline.py
 delete mode 100644 examples/notebooks/img/hook.png
 delete mode 100644 examples/vectors_example.py

diff --git a/docs/source/examples/tfidf_example.rst b/docs/source/examples/tfidf_example.rst
new file mode 100644
index 00000000..0d904c68
--- /dev/null
+++ b/docs/source/examples/tfidf_example.rst
@@ -0,0 +1,105 @@
+
+TFIDF + scikit-learn SVM
+=========================
+
+In this example, we will cover a once popular family of models -- support vector machines (SVMs) with TF-IDF representations. As a simple example, we will analyse binary classification on the Stanford sentiment treebank (SST) dataset.
+
+First, we will implement a minimalistic example without much additional preprocessing. Since we're using TFIDF representation of our dataset, it is smart to limit the size of the vocabulary as each word needs to be present in the instance TFIDF representation. Let's load the SST dataset and convert it into a single batch:
+
+.. code-block:: python
+
+  >>> from podium import Vocab, Field, LabelField
+  >>> from podium.datasets import SST
+  >>> vocab = Vocab(max_size=5000, specials=())
+  >>> text = Field(name='text', numericalizer=vocab, disable_batch_matrix=True)
+  >>> label = LabelField(name='label')
+  >>> fields = {'text': text, 'label': label}
+  >>> train, dev, test = SST.get_dataset_splits(fields=fields)
+  >>> train.finalize_fields()
+  >>> x, y = train.batch(add_padding=True)
+
+We have now loaded our dataset, finalized its Fields and obtained it as a batch of input and target data. What we need to do next is define the TF-IDF vectorization for each instance in the dataset. This is done by using our :class:`podium.vectorizers.TfIdfVectorizer`, which adapts the ``scikit-learn`` vectorizer to the Podium input data.
+
+.. code-block:: python
+  
+  >>> from podium.vectorizers import TfIdfVectorizer
+  >>> tfidf_vectorizer = TFIdfVectorizer()
+  >>> tfidf_vectorizer.fit(train, field=train.field('text'))
+  >>> tfidf_x = tfidf_vectorizer.transform(x.text)
+  >>> print(type(tfidf_batch), tfidf_batch.shape)
+  <class 'scipy.sparse.csr.csr_matrix'> (6920, 5000)
+
+We have transformed the train dataset to a sparse matrix containing TF-IDF values for each word in the vocabulary in each instance. What is left to do now is to train our classification model:
+
+.. code-block:: python
+
+  >>> from sklearn.svm import LinearSVC
+  >>> from sklearn.metrics import accuracy_score
+  >>> # Train the SVM on the training set
+  >>> svm = LinearSVC()
+  >>> svm.fit(tfidf_batch, y.label.ravel())
+  >>> # Obtain accuracy on the train set
+  >>> y_hat = svm.predict(tfidf_batch)
+  >>> acc = accuracy_score(y_hat, y.label.ravel())
+  >>> print(f"Accuracy on train set: {acc:.4f}")
+  0.9597
+
+And for a more accurate performance evaluation of our model we turn to the test set:
+
+.. code-block:: python
+
+  >>> test_x, test_y = test.batch()
+  >>> tfidf_test_batch = tfidf_vectorizer.transform(test_x.text)
+  >>> y_test_hat = svm.predict(tfidf_test_batch)
+  >>> acc = accuracy_score(y_test_hat, test_y.label.ravel())
+  >>> print(f"Accuracy on test set: {acc:.4f}")
+  Accuracy on test set: 0.7946
+
+Our basic unigram TF-IDF linear SVM performs pretty well on the SST dataset, reaching accuracy of almost ``0.8``. While this example encapsulates the basics of using Podium with ``scikit-learn``\s ``SVM``, we will delve a bit deeper and consider some additional preprocessing.
+
+Using ngram features
+---------------------
+
+We have only considered basic unigram features for our model, and this is somewhat prohibitive. Apart from that, we have not implemented any preprocessing for our dataset, and our dataset is cased -- which might be detrimental for the performance of our model since we agressively trim the vocabulary size.
+
+We will do two things: (1) implement a pre-tokenization hook to lowercase our data, which in our case is fine as we are using the case-insensitive ``str.split`` as a tokenizer, and (2) implement ngram extraction as a post-tokenization hook. For a more detailed overview of hooks and how to use them, check :ref:`fields`. We will first implement our lowercase hook:
+
+.. code-block:: python
+
+  >>> def lowercase(raw):
+  ...   """Lowercases the input string"""
+  ...   return raw.lower()
+
+And then implement flexible ngram extraction where the ``n`` is an interval using ``nltk``\s ``ngrams`` function:
+
+.. code-block:: python
+  
+  >>> from ntlk import ngrams
+  >>> class NGramHook:
+  ...   # Transforms a sequence of unigrams into a sequence of
+  ...   # [min_n, max_n]-grams
+  ...   def __init__(self, min_n, max_n):
+  ...     self.min_n = min_n
+  ...     self.max_n = max_n
+  ...   def __call__(self, raw, tokenized):
+  ...     tokenized_ngrams = []
+  ...     for n in range(self.min_n, self.max_n+1):
+  ...        tokenized_ngrams.extend(ngrams(tokenized, n))
+  ...     return raw, tokenized_ngrams
+
+We will now incorporate these two hooks into our text input Field:
+
+.. code-block:: python
+
+  >>> ngram_hook = NGramHook(1,3)
+  >>> vocab = Vocab(max_size=5000, specials=())
+  >>> text = Field(name='text', numericalizer=vocab, 
+                   disable_batch_matrix=True,
+                   pretokenization_hooks=[lowercase],
+                   posttokenization_hooks=[ngram_hook]
+                   )
+  >>> label = LabelField(name='label')
+  >>> fields = {'text': text, 'label': label}
+  >>> train, dev, test = SST.get_dataset_splits(fields=fields)
+  >>> train.finalize_fields()
+
diff --git a/examples/dataset_example.py b/examples/dataset_example.py
deleted file mode 100644
index 5a51b2f6..00000000
--- a/examples/dataset_example.py
+++ /dev/null
@@ -1,25 +0,0 @@
-"""
-Example how to use simple PauzaHR dataset.
-"""
-import dill
-
-from podium.datasets import SST
-from podium.storage import LargeResource
-
-
-if __name__ == "__main__":
-    # for large resource settings see
-    # https://github.com/mttk/podium/wiki/Large-resources
-    LargeResource.BASE_RESOURCE_DIR = "downloaded_datasets"
-
-    train_set, _, test_set = SST.get_dataset_splits()
-    print(f"train set {len(train_set)}")
-    print(f"test set {len(test_set)}")
-
-    # save dataset
-    with open("dataset.pkl", "wb") as f:
-        dill.dump(train_set, f)
-
-    # load dataset
-    with open("dataset.pkl", "rb") as f:
-        loaded_train_set = dill.load(f)
diff --git a/examples/experiment_example.py b/examples/experiment_example.py
deleted file mode 100644
index 1a64b8e3..00000000
--- a/examples/experiment_example.py
+++ /dev/null
@@ -1,150 +0,0 @@
-"""
-Example how to use model on simple PauzaHR dataset using the Experiment class.
-"""
-import os
-
-import numpy as np
-from sklearn.metrics import accuracy_score
-from sklearn.preprocessing import StandardScaler
-
-from podium import Field, LabelField, Vocab
-from podium.datasets import ExampleFormat, Iterator, PauzaHRDataset
-from podium.experimental.model_selection import grid_search
-from podium.experimental.models import (
-    Experiment,
-    FeatureTransformer,
-    SklearnTensorTransformerWrapper,
-)
-from podium.experimental.models.impl.fc_model import ScikitMLPClassifier
-from podium.experimental.models.impl.simple_trainers import SimpleTrainer
-from podium.experimental.pipeline import Pipeline
-from podium.experimental.validation import k_fold_classification_metrics
-from podium.storage import LargeResource
-from podium.vectorizers.impl import NlplVectorizer
-
-
-def numericalize_pauza_rating(rating):
-    """
-    Function numericalizes pauza_hr dataset rating field.
-    """
-    label = round(float(rating) * 2) - 1
-    return label
-
-
-def basic_pauza_hr_fields():
-    """
-    Function returns pauza-hr fields used for classification.
-    """
-    rating = LabelField(name="Rating", pretokenize_hooks=[numericalize_pauza_rating])
-
-    text = Field(
-        name="Text",
-        numericalizer=Vocab(),
-        tokenizer="split",
-        keep_raw=False,
-        fixed_length=100,
-    )
-
-    return {"Text": text, "Rating": rating}
-
-
-def label_transform_fun(y_batch):
-    return y_batch.Rating.ravel()
-
-
-def experiment_example():
-    """
-    Example of setting up and using the Experiment class.
-    """
-
-    LargeResource.BASE_RESOURCE_DIR = "downloaded_datasets"
-
-    fields = basic_pauza_hr_fields()
-    train_dataset, test_dataset = PauzaHRDataset.get_train_test_dataset(fields)
-
-    num_of_classes = len(train_dataset.field_dict["Rating"].vocab.itos)
-
-    vector_cache_path = os.path.join(
-        LargeResource.BASE_RESOURCE_DIR, "experimet_example_nlpl_cache.txt"
-    )
-
-    vectorizer = NlplVectorizer(cache_path=vector_cache_path)
-    vectorizer.load_vocab(vocab=fields["Text"].vocab)
-    embedding_matrix = vectorizer.get_embedding_matrix(fields["Text"].vocab)
-
-    def feature_transform_fn(x_batch):
-        """
-        Batch transform function that returns a mean of embedding vectors for
-        every token in an Example.
-        """
-        x_tensor = np.take(embedding_matrix, x_batch.Text.astype(int), axis=0)
-        x = np.mean(x_tensor, axis=1)
-        return x
-
-    trainer = SimpleTrainer()
-
-    tensor_transformer = SklearnTensorTransformerWrapper(StandardScaler())
-    feature_transformer = FeatureTransformer(feature_transform_fn, tensor_transformer)
-
-    experiment = Experiment(
-        ScikitMLPClassifier,
-        trainer=trainer,
-        feature_transformer=feature_transformer,
-        label_transform_fn=label_transform_fun,
-    )
-
-    _, model_params, train_params = grid_search(
-        experiment,
-        test_dataset,
-        accuracy_score,
-        model_param_grid={
-            "classes": ([i for i in range(num_of_classes)],),
-            "hidden_layer_sizes": [(10,), (10, 10), (100,)],
-        },
-        trainer_param_grid={
-            "max_epoch": [2, 3, 4],
-            "iterator": [Iterator(batch_size=32), Iterator(batch_size=64)],
-        },
-    )
-
-    experiment.set_default_model_args(**model_params)
-    experiment.set_default_trainer_args(**train_params)
-
-    accuracy, precision, recall, f1 = k_fold_classification_metrics(
-        experiment, test_dataset, 5, average="macro"
-    )
-
-    print(
-        f"Accuracy = {accuracy}\n"
-        f"Precision = {precision}\n"
-        f"Recall = {recall}\n"
-        f"F1 score = {f1}"
-    )
-
-    experiment.fit(train_dataset)
-
-    dataset_fields = {"Text": train_dataset.field_dict["Text"]}
-
-    pipeline = Pipeline(
-        fields=dataset_fields,
-        example_format=ExampleFormat.XML,
-        model=experiment.model,
-        feature_transformer=feature_transformer,
-    )
-
-    example_good = (
-        "<Example><Text>Izvrstan, ogroman Zagrebački, "
-        "dostava na vrijeme, ljubazno osoblje ...</Text></Example>"
-    )
-    prediction = pipeline.predict_raw(example_good)
-    print(f"Good example score: {prediction}")
-
-    example_bad = (
-        "<Example><Text>Hrana kasnila, dostavljac neljubazan, " "uzas...</Text></Example>"
-    )
-    prediction = pipeline.predict_raw(example_bad)
-    print(f"Bad example score: {prediction}")
-
-
-if __name__ == "__main__":
-    experiment_example()
diff --git a/examples/ner_example.py b/examples/ner_example.py
deleted file mode 100644
index c73e48a1..00000000
--- a/examples/ner_example.py
+++ /dev/null
@@ -1,213 +0,0 @@
-"""
-Example how to use BLCC model on Croatian NER dataset for NER task.
-"""
-
-import pickle
-import sys
-from collections import namedtuple
-from functools import partial
-
-import numpy as np
-from sklearn.metrics import f1_score
-
-from podium import BucketIterator, Field, Vocab
-from podium.datasets.impl.croatian_ner_dataset import CroatianNERDataset
-from podium.experimental.models import FeatureTransformer
-from podium.experimental.models.impl.blcc_model import BLCCModel
-from podium.experimental.models.impl.simple_trainers import SimpleTrainer
-from podium.storage.resources.large_resource import LargeResource
-from podium.vectorizers.vectorizer import BasicVectorStorage
-from podium.vocab import PAD
-
-
-Inputs = namedtuple("Inputs", ["tokens", "casing"])
-# using the same label set as original CroNER
-label_mapping = {
-    "Organization": "Organization",
-    "Person": "Person",
-    "Location": "Location",
-    "Date": "Date",
-    "Time": "Time",
-    "Money": "Money",
-    "Percent": "Percent",
-    "Etnic": "Etnic",
-    # remapped or unused types
-    "PersonPossessive": "Person",
-    "Product": None,
-    "OrganizationAsLocation": "Organization",
-    "LocationAsOrganization": "Location",
-}
-
-
-def label_mapper_hook(data, tokens):
-    """
-    Function maps the labels to a reduced set.
-    """
-    new_tokens = []
-    for token in tokens:
-        if token == "O":
-            new_tokens.append("O")
-            continue
-
-        prefix, value = token.split("-")
-
-        mapped_token = label_mapping[value]
-        if not mapped_token:
-            new_tokens.append("O")
-        else:
-            new_tokens.append(prefix + "-" + mapped_token)
-
-    return data, new_tokens
-
-
-def casing_mapper_hook(data, tokens):
-    """
-    Hook for generating the casing feature from the tokenized text.
-    """
-    tokens_casing = []
-
-    for token in tokens:
-        token_casing = "other"
-        if token.isdigit():
-            token_casing = "numeric"
-        elif token.islower():
-            token_casing = "lowercase"
-        elif token.isupper():
-            token_casing = "uppercase"
-        elif token[0].isupper():
-            token_casing = "initial_uppercase"
-
-        tokens_casing.append(token_casing)
-
-    return data, tokens_casing
-
-
-def feature_extraction_fn(x_batch, embedding_matrix):
-    """
-    Function transforms iterator batches to a form acceptable by the model.
-    """
-    tokens_numericalized = x_batch.tokens.astype(int)
-    casing_numericalized = x_batch.casing.astype(int)
-    X = [np.take(embedding_matrix, tokens_numericalized, axis=0), casing_numericalized]
-    return X
-
-
-def label_transform_fun(y_batch):
-    return y_batch.labels.astype(int)
-
-
-def example_word_count(example):
-    """
-    Function returns the number of tokens in an Example.
-    """
-    return len(example.tokens[1])
-
-
-def ner_croatian_blcc_example(fields, dataset, feature_transform):
-    """
-    Example of training the BLCCModel with Croatian NER dataset.
-    """
-    output_size = len(fields["labels"].vocab.itos)
-    casing_feature_size = len(fields["inputs"].casing.vocab.itos)
-
-    train_set, test_set = dataset.split(split_ratio=0.8)
-
-    train_iter = BucketIterator(batch_size=32, sort_key=example_word_count)
-
-    model = BLCCModel(
-        **{
-            BLCCModel.OUTPUT_SIZE: output_size,
-            BLCCModel.CLASSIFIER: "CRF",
-            BLCCModel.EMBEDDING_SIZE: 300,
-            BLCCModel.LSTM_SIZE: (100, 100),
-            BLCCModel.DROPOUT: (0.25, 0.25),
-            BLCCModel.FEATURE_NAMES: ("casing",),
-            BLCCModel.FEATURE_INPUT_SIZES: (casing_feature_size,),
-            # set to a high value because of a tensorflow-cpu bug
-            BLCCModel.FEATURE_OUTPUT_SIZES: (30,),
-        }
-    )
-    trainer = SimpleTrainer()
-    feature_transformer = FeatureTransformer(feature_transform)
-
-    print("Training started")
-    trainer.train(
-        model=model,
-        dataset=train_set,
-        feature_transformer=feature_transformer,
-        iterator=train_iter,
-        label_transform_fun=label_transform_fun,
-        max_epoch=1,
-    )
-    print("Training finished")
-
-    X_test_batch, y_test_batch = test_set[:32].batch()
-    X_test = feature_transformer.transform(X_test_batch)
-    y_test = label_transform_fun(y_test_batch)
-
-    prediction = model.predict(X=X_test)[BLCCModel.PREDICTION_KEY]
-    # pickle for later use
-    pickle.dump(model, open("ner_model.pkl", "wb"))
-
-    pad_symbol = fields["labels"].vocab.padding_index()
-    prediction_filtered, y_test_filtered = filter_out_padding(
-        pad_symbol, prediction, y_test
-    )
-
-    print("Expected:")
-    print(y_test_filtered)
-
-    print("Actual:")
-    print(prediction_filtered)
-
-    f1 = f1_score(y_test_filtered, prediction_filtered, average="weighted")
-    info_msg = f"F1: {f1}"
-    print(info_msg)
-
-
-def filter_out_padding(pad_symbol, prediction, y_test):
-    """
-    Filters out padding from the predictiopytn and test arrays.
-
-    The resulting arrays are flattened.
-    """
-    indices_to_leave = np.where(np.ravel(y_test) != pad_symbol)
-    y_test_filtered = np.ravel(y_test)[indices_to_leave]
-    prediction_filtered = np.ravel(prediction)[indices_to_leave]
-    return prediction_filtered, y_test_filtered
-
-
-def ner_dataset_classification_fields():
-    """
-    Function creates fields to use with the Croatian NER dataset on NER task.
-    """
-    tokens = Field(name="tokens", numericalizer=Vocab(), tokenizer=None)
-    casing = Field(name="casing", numericalizer=Vocab(specials=(PAD(),)), tokenizer=None)
-    labels = Field(
-        name="labels",
-        is_target=True,
-        numericalizer=Vocab(specials=(PAD(),)),
-        tokenizer=None,
-    )
-
-    casing.add_posttokenize_hook(casing_mapper_hook)
-    labels.add_posttokenize_hook(label_mapper_hook)
-
-    return {"inputs": Inputs(tokens, casing), "labels": labels}
-
-
-if __name__ == "__main__":
-    vectors_path = sys.argv[1]
-    LargeResource.BASE_RESOURCE_DIR = "downloaded_datasets"
-
-    fields = ner_dataset_classification_fields()
-    dataset = CroatianNERDataset.get_dataset(fields=fields)
-
-    vocab = fields["inputs"].tokens.vocab
-    embedding_matrix = BasicVectorStorage(path=vectors_path).load_vocab(vocab)
-
-    feature_transform = partial(feature_extraction_fn, embedding_matrix=embedding_matrix)
-
-    ner_croatian_blcc_example(
-        fields=fields, dataset=dataset, feature_transform=feature_transform
-    )
diff --git a/examples/ner_pipeline.py b/examples/ner_pipeline.py
deleted file mode 100644
index 9d58e12e..00000000
--- a/examples/ner_pipeline.py
+++ /dev/null
@@ -1,191 +0,0 @@
-"""
-Example how to use BLCC model on Croatian NER dataset for NER task.
-"""
-
-import os
-import pickle
-import sys
-import time
-from functools import partial
-
-from ner_example import (
-    example_word_count,
-    feature_extraction_fn,
-    label_transform_fun,
-    ner_dataset_classification_fields,
-)
-
-from podium.dataload.ner_croatian import convert_sequence_to_entities
-from podium.datasets.impl.croatian_ner_dataset import CroatianNERDataset
-from podium.datasets.iterator import BucketIterator
-from podium.models.impl.blcc_model import BLCCModel
-from podium.models.impl.simple_trainers import SimpleTrainer
-from podium.pipeline import Pipeline
-from podium.storage import ExampleFormat
-from podium.storage.resources.large_resource import LargeResource
-from podium.storage.vectorizers.vectorizer import BasicVectorStorage
-
-
-class CroatianNER(Pipeline):
-    """
-    Pipeline used to train named entity recognition for Croatian.
-
-    It is designed to work on the croopinion dataset, but makes no assumptions
-    on the underlying data, except that the data is tokenized before hand and
-    labeled with BIO tags.
-    """
-
-    def __init__(self, vector_path):
-        """
-        Creates a new CroatianNER pipeline. Initializes fields and croopinion
-        dataset. Defines feature transform (word vector lookup) and output
-        transform (mapping labels back to BIO labels). Uses the ```BLCCModel```
-        (BiLSTM + linear chain CRF) Expects examples in DICT field format.
-
-        Parameters
-        ----------
-        vector_path : str
-            Path to Croatian word vector file
-
-        Raises
-        ------
-        ValueError
-            If vector_path is None or is not a valid file path.
-        """
-        if vector_path is None or not os.path.exists(vector_path):
-            raise ValueError(
-                f"Provided path {vector_path} is None or does not exist. "
-                "Path to word Croatian vectors must be defined. "
-                "You can use fastText vectors available at "
-                "https://fasttext.cc/docs/en/crawl-vectors.html"
-            )
-
-        self.fields = ner_dataset_classification_fields()
-        self.dataset = CroatianNERDataset.get_dataset(fields=self.fields)
-        self.feature_transform = self._define_feature_transform(vector_path)
-        self.output_transform_fn = partial(
-            CroatianNER.map_iterable, mapping=self.fields["labels"].vocab.itos
-        )
-
-        super().__init__(
-            fields={
-                # tokens is a feature field
-                "tokens": self.fields["inputs"].tokens,
-                # casing is a feature field
-                "casing": self.fields["inputs"].casing,
-                # labels is a target field
-                "labels": self.fields["labels"],
-            },
-            example_format=ExampleFormat.DICT,
-            model=BLCCModel,
-            feature_transformer=self.feature_transform,
-            output_transform_fn=self.output_transform_fn,
-            label_transform_fn=label_transform_fun,
-        )
-
-    def fit(
-        self,
-        dataset,
-        model_kwargs=None,
-        trainer_kwargs=None,
-        feature_transformer=None,
-        trainer=None,
-    ):
-        """
-        Fits the CroatianNER pipeline on a dataset using provided parameters.
-
-        Parameters
-        ----------
-        dataset : Dataset
-            Dataset to train on
-
-        model_kwargs : dict
-            model keyword argument dict forwarded to the model instance
-
-        trainer_kwargs : dict
-            trainer keyword argument dict forwarded to the trainer instance
-
-        trainer : AbstractTrainer
-            trainer instance (must inherit from `AbstractTrainer`)
-        """
-
-        if model_kwargs is None:
-            model_kwargs = self._define_model_params()
-            print(f"Using default model parameters {model_kwargs}")
-
-        if trainer_kwargs is None:
-            # use bucket iterator to minimize padding in batch
-            iterator = BucketIterator(batch_size=32, sort_key=example_word_count)
-            trainer_kwargs = {"max_epoch": 10, "iterator": iterator}
-            print(f"Using default trainer parameters {trainer_kwargs}")
-
-        trainer = SimpleTrainer() if trainer is None else trainer
-
-        start = time.time()
-        print("Starting training")
-        super().fit(
-            dataset=dataset,
-            model_kwargs=model_kwargs,
-            trainer_kwargs=trainer_kwargs,
-            trainer=trainer,
-        )
-        print(f"Training took {time.time() - start} seconds")
-
-    def predict_raw(self, raw_example, tokenizer=str.split):
-        """
-        Predicts target Fields for raw_example.
-
-        Parameters
-        ----------
-        raw_example : str
-            Sentence in Croatian.
-
-        Returns
-        ------
-        list (dict)
-            List of dicts of recognized named entities,
-            where each dict has keys:
-            name, type, start, end.
-        """
-        tokenized_text = tokenizer(raw_example)
-        example_to_predict = {"tokens": tokenized_text, "casing": tokenized_text}
-        tags = super().predict_raw(example_to_predict)
-        return convert_sequence_to_entities(tags, tokenized_text)
-
-    def _define_feature_transform(self, vector_path):
-        vectorizer = BasicVectorStorage(path=vector_path)
-        embedding_matrix = vectorizer.load_vocab(vocab=self.fields["inputs"].tokens.vocab)
-        return partial(feature_extraction_fn, embedding_matrix=embedding_matrix)
-
-    def _define_model_params(self):
-        output_size = len(self.fields["labels"].vocab.itos)
-        casing_feature_size = len(self.fields["inputs"].casing.vocab.itos)
-        return {
-            BLCCModel.OUTPUT_SIZE: output_size,
-            BLCCModel.CLASSIFIER: "CRF",
-            BLCCModel.EMBEDDING_SIZE: 300,
-            BLCCModel.LSTM_SIZE: (20, 20),
-            BLCCModel.DROPOUT: (0.25, 0.25),
-            BLCCModel.FEATURE_NAMES: ("casing",),
-            BLCCModel.FEATURE_INPUT_SIZES: (casing_feature_size,),
-            # set to a high value because of a tensorflow-cpu bug
-            BLCCModel.FEATURE_OUTPUT_SIZES: (30,),
-        }
-
-    @staticmethod
-    def map_iterable(iterable, mapping):
-        return [mapping[i] for i in iterable]
-
-
-if __name__ == "__main__":
-    model_path = "ner_pipeline_entire_model.pkl"
-    LargeResource.BASE_RESOURCE_DIR = "downloaded_datasets"
-
-    ner_pipeline = CroatianNER(sys.argv[1])
-    ner_pipeline.fit(ner_pipeline.dataset)
-
-    pickle.dump(ner_pipeline, open(model_path, "wb"))
-
-    loaded_ner = pickle.load(open(model_path, "rb"))
-    text = "U Hrvatskoj državi žive mala bića . Velika bića žive u Graškogradu ."
-    print(loaded_ner.predict_raw(text))
diff --git a/examples/notebooks/img/hook.png b/examples/notebooks/img/hook.png
deleted file mode 100644
index 1d79b306c746e72cfa690fe3be9538eb73b57934..0000000000000000000000000000000000000000
GIT binary patch
literal 0
HcmV?d00001

literal 8521
zcmeHLcTkhtw*L?m3jzuXQtiT_>!C<T1XQGWP(U#h5fLd7NQe|E$Ah4hSSbPtMGO|2
z3P=xv;1LN3NTh@y5h4TvA#@UW-;SL7X3o61^XATbf4q0i3@Pi|dzIh%?X}n9maVnL
z=8aMtAqd)h{MgaoAc+4v_*=gYmRRL<euF=Dpv)XmcF0R8j9Y*QWO5OC!DHufUpG&W
z-#pwd2K)c%p$|bj8OM*BoW9mGMyd6avA_{o_r?c=)m?R^b;^T+QWa0WOm*t^$bMP*
z{$fyS-hrJ90(<e1IhWPMWPS7OgEs6M-O!i#(I+_a#^%Bk9p^h%_6F$KAt>AhC_!Uv
zTd1Qc-;xKnLQep8=I84t0zVP>_d`ITagQYg3Dx;H%NNXBNH|DXd9USDpm-)hkYwzQ
zZ7b)+M52U^S%$@+NbGFmgWscs;kPxi@S7-H(|!v7bwDDG1^>cxJ?bw!e+qd7`_};x
z(1oGoZ8y&3hoB#L{-<!j!SIy;L^PJkpQoR*1PS=>zKUL2z{f&?;^XYV+9!+FuZeA%
zwOBqTjg_bQrzSp+ds1>6*o|BgT(A%8zSUHQc;FzyxkqMQ9%Nk)**3;|C=#>a7Z%*}
zZef3}W5f<DTarPxmje`vof`662-n9oxht_p?9$Z+{jWI4f(Pr_9Jgf)>p}mE7QEjl
z6-gLcv;Qu)@LsgU$N|HB5FDKPuzKNO^I6CK`TUDk`m^!X{P1Co)M#9=mY_TF2RD0Z
zbIszXqisZh#`t}t4xOcCW9ss5^F^G^79bNk{ISC?_%@po6`^FMk1LOo0cTwDN_uT#
z)9kUD+~#ZgTSh<x#_4e=$b3u;E*n0??)5Z%6t#oFln@1lATr?HH7oy`LxV^Sv&CW?
zsYvuv=T->vj%QQm19YO0=p}aHGYg}YPAx+;)b&pDJ4CD)<Z$OPlB7Tqc^3k!+Ro>=
zX%7<?r&}qLs1dCN7J?v)?_%$UNHzxBy}>>m^dC3?MWWRQg(3Z4Np_Ey)ma|uqr$`+
z-6n#N+y~i~@ofI0bwRK*4_k)OvPs%E>s56jk&A+|(6v~4(g1!en}2dA@A;cqj!3te
zpsNS?b*9^3rFue{4|@LchA{NvqE_LG+>^OJcIEukXm)w25Y$HoRnc)52@{gY7AIQi
z_j8A-5n|X1`dm`irLP150OPmG*97(x1*7VU%Os;|qtWoF$*Q;+!!Gko{;NpyKEB9M
zhFg`vc#8Zxb)$0n(#fpwq=o96^M>3AWst@qmJT&h4SA?3xI%An==Ma}ZMDL|;tZBS
zWttu|H>?UHbA3pNBKR(N>xxfWjk$hL{6yuf`x-!Ut?{&YddTb@o!V_9iqrT6R01f_
z#G^w!CN3`mPgZT>=G@KS0<tRlUApX6inVrXGb?`T19k(DeR(Z|IchXfAnWqnuam3a
zv>UBW;O7xTSC{HD=^cR|35C_c1aO`F6nwa?a5Ik|SSv%*FtHr=k81#=%BRuC8F_Xs
zh{kwDYgR!U<G@Ry`yIwQ@3gS4ka2ebu<k~efrWm|gw}is{?ejcp9=2~_?%ArD%1Na
z7*V(s7epJ^JG@<yrgjL#5ot`#yRkW}g?4g+!aU5gZ57t_D!96~Ky7=iea~U)j&3PF
zz`+fn$*l8D3`sQBWe)|#^m*(p0N~mG?9Z?o2aO_(yeM69(xu7WyTMz4xL*GyTnsjT
zS~3lLdJsRmSIz=cy>KoJtN6YnmSZwJ7c5JqC5NAY-W1xX@IhBy@DXQnbyh5gJD#w)
z82!20#vJYfVRq5xZPo1+gXjep8cW_lxpk#sZnhe~6=F6E>q1wN6io^;G+f#j)7N)L
zrJC@|H){7vzNdpkOF6s-#_5hy`;ug!S_Uf!vLD{>4-ua^pZlF!94HL@eyLY1P#bu0
z0|3F;o<ixAuph*eXpdee@qzSA*Cn#ltYP0ozG`?w#1HK`B*T?wCt9Ufnj4ZY_@h6;
z8Ix;(-I%oXlHi8CrLLgWkMoM+VizKA`NCVvtD`-EP*+_b3|&b|3xmwc#eFHezJtao
zwvmrS7lFgj{7OCl)Et8ux-PPWCus3;CbNcSj@D3kVc75r(*voIO4g9iH|p@Ic(y5C
z3nWRwpxrKWB5}Ic5_`Irtq7_TbXS$c2L}V)PeA$eLR=?1vcwv4Oy>vjE@G!Cs_kX#
z=4a<Fq^pJTN~){!-5S{>42CNA-pQ|i>eQ8FLjzvbYqT|rR}+k`ZMk0*_MoT4MhHL>
zf4IFvHAE`7a8szaR~g4lh}}rwQQ(RAs2(EE!Jzh`mNMp}gyA-xjv3nsxv7aff-Kw4
zjpxohI?ur^cu%=iOo((JBpmrR?7{{p^0x+8>#v;s^-s_Ja2f)e|Jg<JU*10drb3W*
zh!;Cd3fBjh*cC6~*?5g?<nnmqF>0~UkM9}pj!|F!_#U3CWAtiOJT8eA1nhrMm!}7q
zn+cgy_yGFRaNP%mN8yiC7|@p*^Tk*2gA%pTty315vt6IV<f&e>szUtjstVCn6>I;l
zV%J|PT>qxx$X_ame^c=fFCtgH_=gIApC3BfYJZUvgpVuiQClsyyQ)BipBG02bJc(2
zZL3yq|J~}E(|-lM{_jEm-1JY;i!?41fW$L50@&c9U<MKmQ-jA9I1$Et&j8jit!xEg
zlEAcz<YgG9)&J0@`2+@z<o(0O)n*U<n#88v?BM6CRwv`&hB2YZnbV|3^B12Qe<lr+
z95&7|$#ZF<smm#L+7*@UuwLCi8Rh8GuqUC$D9++lo!uG037Gc-Ia%Ry*aIwi`Ir_2
z!Ee=6lWpx-ft^j|=m9q_`Ki2<2jCr%VroFKWtvHhj+7uelKXKj3{xc`0LUAcJc9Ix
z3>5h(A!DXlGpzfD<Vdf*1VE<f+za?ou6Jig$tYcf6X0nt4#bm1w#dP|%tc(#X9>3w
zMrQ(i=-^V{B|bWyFa1F_UvzX8b8n#;RK}v4evMEwqq8@Z<G%{@=7z=W%B=Gh*45?p
z@)|$#^2Ife!05DlOi25nVmM`_y(+z$=hDy8MXzoY%EI|6V7^2z#!W8$!|uVGAr)DD
z$1xcpJ!g@AT+h|Dau=?DLJ)9?W^;wB4u));8SX-lCc{kw@(dKYYyRU_NWI+_SsySa
z3%QDp!A%M8V*FpI?5>YX;m#5!*+6}d5dBN(N1rwcxS5QC@0*xq-=$Dq1DA4B9=^w6
zBzcqdizax^u#ej8FLng*BIjzxi2j-ktJR<N81$#}=hn_EMA(3o#GCtWa4hStB9(-O
z`<8&)HCALku)qj(Uh2ObZp74<eDpC+lAVT0SKJBG-kAMTz=v3Sb`4POK4qSH$o6nh
zn{=}f?Bvd5WpZZgFxvO&4%bX@Dep`>kG-?i?gf19JM3z;8tixU1>4GBr;^wSxt9_~
z?{c{y8g(@M4}=2fX}*^C1BQel912{N?X7MGo~Ae+d|{eF+bdVR4(Ka8PtGDIJ&JkM
z!a$7&Cx%MBwwDHL{U#1((48z>q{c4lw&v!pgWbQACX2ecm#h}jSI1N544c?EEz@dh
zL(U5AlM?ZO8~Z4~4UoBtLwuO$)3cv_oC!LuEKZ<yEI+)xE$uw5N-yb=;I&61`hzmS
z2r^*pEe(v^ZnE07^AX_2&y=U=OrW21><Fay2PBpW>hb#AkR+uOK`NR`?`PsNyvcab
zx_oykDDKbiuG&7*ca{@;T#sJf*!vDNqmlg?p8Cle7^b@st-PkJ4<H<ZMt^#}i%RRd
zq8&$*3s^ovDCWseOdh0_<DC2qpH#|Ze?V0<GbPS|)BGJl(wL)2x;)%R-y?hw$H*N5
z;oJW4(VbaNi;GHaVd#GAHLxM&>jH!KzV(^Cjw8i<<E$l>!;K@Hj@&=1<1R~NP)7<3
z3J>AhDZPLP4|03WP%66Dn-}8J7_L=32(tnv_RBd1_<bBkhcJU5vL1L<aeF@StBYfm
z`H0_kk4&&UlxT|hhGSP=aK1$_+gtlYQX^*Nn()<@&qau~UrYp3a2+4%dau6yF?2bU
zq*E(S2LYcbdmmPFMXOLmWeJ7Gp6P83>#<W3A+|~S3#MFB&qa(=J`yzj^eOnVGM0xP
zO|f_W)ojT_LD$Rh$vbH$Um0gkA?%>Xq|)?*y@vQCv*C`LIL+bMh0ZKEoROXCKNY&Y
zbwL4F$=YyoKglj-MU(B-f#_XPT2ANIAApE$z^{TFIZ|+&nKTW{Mwz~GFi2nQ%=rvX
zh|r0-@2{T3*wu~>Civ?K2Os$Lp7`*ABrtUlMQ4iRaM>+0-}h*xkt=ttSgjYUq(efF
zyxaKpkJ0fxr7_;0?nhwh+}dl)D>I7ms-%v6LJ^FwodpJhu9L9$li38eTvJMKF(>CX
z8#VHlfwM-=$wbz%Wl+#!h-%2a-$!rJbFjaVmfmd0q~rM2H^LL=vBU~T()jhxa<4vx
z^civbpPQORU^2Rj7pOkcs!5+kh1@l~N)a%3NmlxmMW5><KtxFx%7LKET+u12I@4S#
zFk(=5(4=v}4~zU-Y`GVb+ypbXuarzH@3{6l1LYqsZ^2_cQQ_SeuR%*Z?d@(vI`xb+
zRtzryIHX}&`TcrgY8<?hfVWFF1uT>-Q4N_5373XfM!Y|+J<bQ^48kJ`=WJ~CMlL0F
z=_dVgGMn8-4<sSI>!%v82DS1@%ED$jsnx1C7g6l>rLyuCh5j{&W9hXk<uEbgKNFIm
zNTVsX&A1UfnS@9z%6l#+U*mX&sbL30uY?`!rd1`XMQq|4gin;B$g|NlYa@#f4hj>S
z){Lvy$cRq`waX5VXIisgAqVNp4FpgsT}QK$8q*j@`mLq0o&}cF@AM2+m1}3(JaOT>
z*d;C>A@!gF6TtyhTy(g6b&%6GY#D{A%77_Ry&|&q-NF*y$X}Q!4)bN8;A@hf=xk|G
z*~6eub7M~mhuBt?!y{X+<WlZTm~p~v!j1rD&!Zd0gSI$4$|(xdDn)dCHY8^-169V$
z*>Ur!%!NOAmtvC!M#g1!Q<%4}BeN8@^nl4ZYmZ1FudS(Wo<D0is|aS(fqJ*(Y`?O$
zGBPU5Z;ab&RQuWv$0XIbe0MsDr-+Gd0mj}q4dm)oA|*Il5^f89V_gSkpSP-*EwF9-
zb(RtsswD33Yhlttdv2q$4Aa#mz|pgWhyuI0`#W74=e_q&FPug^nAk2itSuK82<n2q
z@*1k}Ug*AfB!Q(P&(I`M5Z$$#h71mlo3_)YwlACWTyrx?32Ji@QE0y}P9I5}pklw0
zCdz`E?H)Mrhc3Z1f)$Bjr=@)*7|5IA49X+I&IqZS{;5$qf}o~8aeXn^{G6}p!o8rb
zG;zjA$z$*AO7CTc>iZ6F%}I3lE)pMH%!;~WWj-X(m||lS73Ajg<btM>7nNwvB>|J=
zUQKja5CAfEx<ug8?Pe`Gqo6|HIciU(etW)CPTEZQ0$r_weC4PHJUxQCXq<<S9@<m0
z^Z2P{pAW5%d0yb}hQm3?Z}_H(NOy7C2-HnWz(eW2=*6VZDUabv(4d(&4;<WWfN;uS
zPCB1WIL?yP1PjF_N1PCk$@cEkJw-BJm_s@|q0DFuYlwIvzAM;~nO@n?gMmL5wP%8M
z|96hU&4pXJ;NBw^?KViLBBV8qpG&B-%oz(c?*TrlkJ(ahsg0zLFsBSpdXYJ82SE$q
zDsvY%sE<fyrjFEqYM@0U0)eIYL!FEwnNVt})VqNE!+;0-#Yepk7>wshhpIL2R>rhY
zvSfKlxQ)`jt7IV-{|em9u2Wgg(L;U}wyhH{H?bRk<2ABrwuP0ph53z*C2qkP3vYi2
zMWPomb%~4nAM@s1&}J~jM$b4h^9c03%GAF4U){>x`yF|h6<x_A=ui`y*w;?me<QVL
zd}*yzN2&+{ryM<B5Gd1xy7{G;DtUJsf2O||s2Z*vS#URK_kw_!@l~2yeu9MO?oK~m
z7~hq;o~w3skU|XV6e30zhQRX7*N(N_XIHf1*~0;f@2P(b&4)0=Kn5Yx4YR*4J7el_
z%jHsSy*6a5OOQ$b8-6`tThgw`0g;Eqd~)8lbzEBmmxvkp{6!Z$NxATb>~>ff|CSp!
z*_=Qh)x-_I+{F4pAqg6@R2YQ&#=QX<6YAXIrwlr*pRb<?{J%nA8E>Go7epjFy;7p&
h^M5Kf^<xm<I%oi`ym{-H2~ptsaWm_q1xH+e{|{pQD?I=J

diff --git a/examples/vectors_example.py b/examples/vectors_example.py
deleted file mode 100644
index 77916266..00000000
--- a/examples/vectors_example.py
+++ /dev/null
@@ -1,26 +0,0 @@
-"""
-Module contains example on how to use vectorizer class by using GloVe concrete
-vectors class.
-"""
-from podium import Vocab
-from podium.storage import LargeResource
-from podium.vectorizers import GloVe
-
-
-if __name__ == "__main__":
-    LargeResource.BASE_RESOURCE_DIR = "downloaded_datasets"
-    # we use vocab so that we don't need to load all vectors
-    vocab = Vocab.from_itos(["frog", "load", "lizard", "company", "city"])
-
-    vectorizer = GloVe()
-    vectorizer.load_vocab(vocab=vocab)
-
-    # token to vector
-    token = "frog"
-    vector = vectorizer.token_to_vector(token=token)
-    print(token, vector)
-
-    # embedding matrix
-    embedding = vectorizer.get_embedding_matrix(vocab)
-    print(vocab.itos)
-    print(embedding)

From ecd9c265f260063876020a4beb87a83fd10acdc7 Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Wed, 31 Mar 2021 21:03:29 +0200
Subject: [PATCH 03/26] Finalize tfidf example

---
 docs/source/examples/tfidf_example.rst | 46 +++++++++++++++++++++++---
 docs/source/index.rst                  | 11 ++++--
 2 files changed, 51 insertions(+), 6 deletions(-)

diff --git a/docs/source/examples/tfidf_example.rst b/docs/source/examples/tfidf_example.rst
index 0d904c68..bc62b3a4 100644
--- a/docs/source/examples/tfidf_example.rst
+++ b/docs/source/examples/tfidf_example.rst
@@ -41,8 +41,8 @@ We have transformed the train dataset to a sparse matrix containing TF-IDF value
   >>> # Obtain accuracy on the train set
   >>> y_hat = svm.predict(tfidf_batch)
   >>> acc = accuracy_score(y_hat, y.label.ravel())
-  >>> print(f"Accuracy on train set: {acc:.4f}")
-  0.9597
+  >>> print(f"Accuracy on the train set: {acc:.4f}")
+  Accuracy on train set: 0.9597
 
 And for a more accurate performance evaluation of our model we turn to the test set:
 
@@ -52,8 +52,8 @@ And for a more accurate performance evaluation of our model we turn to the test
   >>> tfidf_test_batch = tfidf_vectorizer.transform(test_x.text)
   >>> y_test_hat = svm.predict(tfidf_test_batch)
   >>> acc = accuracy_score(y_test_hat, test_y.label.ravel())
-  >>> print(f"Accuracy on test set: {acc:.4f}")
-  Accuracy on test set: 0.7946
+  >>> print(f"Accuracy on the test set: {acc:.4f}")
+  Accuracy on the test set: 0.7946
 
 Our basic unigram TF-IDF linear SVM performs pretty well on the SST dataset, reaching accuracy of almost ``0.8``. While this example encapsulates the basics of using Podium with ``scikit-learn``\s ``SVM``, we will delve a bit deeper and consider some additional preprocessing.
 
@@ -91,6 +91,7 @@ We will now incorporate these two hooks into our text input Field:
 
 .. code-block:: python
 
+  >>> # Use [1-3]grams, inclusive
   >>> ngram_hook = NGramHook(1,3)
   >>> vocab = Vocab(max_size=5000, specials=())
   >>> text = Field(name='text', numericalizer=vocab, 
@@ -102,4 +103,41 @@ We will now incorporate these two hooks into our text input Field:
   >>> fields = {'text': text, 'label': label}
   >>> train, dev, test = SST.get_dataset_splits(fields=fields)
   >>> train.finalize_fields()
+  >>> print(text.vocab.itos[40:50])
+  [('at',), ('from',), ('one',), ('have',), ('I',), ('like',), ('his',), ('in', 'the'), ('all',), ("'",)]
 
+We can see that our new Vocab now contains tuples as its tokens -- as long as an item in a sequence is hashable, we can represent it as part of a Vocab! We can see that one 2-gram ``('in', 'the')`` has made its way into the 50 most frequent tokens.
+
+As before, we need to train the TFIDF vectorizer and apply it to our data (which now includes 1-, 2- and 3-grams):
+
+.. code-block:: python
+
+  >>> dataset_batch = train.batch(add_padding=True)
+  >>> tfidf_vectorizer = TfIdfVectorizer()
+  >>> tfidf_vectorizer.fit(train, field=train.field('text'))
+  >>> tfidf_batch = tfidf_vectorizer.transform(dataset_batch.text)
+  >>> print(type(tfidf_batch), tfidf_batch.shape)
+  <class 'scipy.sparse.csr.csr_matrix'> (6920, 5000)
+
+We can now train our SVM classification model and evaluate it on the train and test set:
+
+.. code-block:: python
+
+  >>> svm = LinearSVC()
+  >>> text, label = dataset_batch
+  >>> svm.fit(tfidf_batch, label.ravel())
+  >>> # Compute accuracy on the train set
+  >>> y_hat = svm.predict(tfidf_batch)
+  >>> acc = accuracy_score(y_hat, label.ravel())
+  >>> print(f"Accuracy on the train set: {acc:.4f}")
+  Accuracy on the train set: 0.9575
+  >>>
+  >>> # Compute accuracy on the test set
+  >>> test_text, test_label = test.batch(add_padding=True)
+  >>> tfidf_test_batch = tfidf_vectorizer.transform(test_text)
+  >>> y_test_hat = svm.predict(tfidf_test_batch)
+  >>> acc = accuracy_score(y_test_hat, test_label.ravel())
+  >>> print(f"Accuracy on the test set: {acc:.4f}")
+  Accuracy on the test set: 0.7743
+
+Sadly, our new model didn't perform better than our initial one on the train set, but there are many avenues we can try further, such as tuning the hyperparameters of the LinearSVC model on the development set or filtering out stop words and punctuation. We encourage you to open this example in Colab and try some things yourself!
diff --git a/docs/source/index.rst b/docs/source/index.rst
index 6ad3bfe9..d237452f 100644
--- a/docs/source/index.rst
+++ b/docs/source/index.rst
@@ -13,11 +13,12 @@ The goal of Podium is to be **lightweight**, in terms of code and dependencies,
 
 Contents
 ---------------------------------
-The documentation is organized in four parts:
+The documentation is organized in five parts:
 
 - **Quickstart**: an quick preview of the library,
 - **Walkthrough**: a description of how the basics work,
-- **In-depth overview**: examples of advanced usage of Podium,
+- **In-depth overview**: advanced usage options,
+- **Examples**: full stand-alone examples of NLP models using Podium,
 - **Core package Reference**: the documentation of methods and classes in Podium.
 
 
@@ -36,6 +37,12 @@ The documentation is organized in four parts:
     advanced
     coming_soon
 
+.. toctree::
+    :maxdepth: 2
+    :caption: Full examples
+
+    examples/tfidf_example.rst
+
 .. toctree::
     :maxdepth: 2
     :caption: Preprocessing Tools

From f6726e49cf2271603eef3da70c080aa6cfb0d834 Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Thu, 1 Apr 2021 02:12:16 +0200
Subject: [PATCH 04/26] Fix shuffling cost

---
 docs/source/advanced.rst                     | 8 ++++----
 docs/source/examples/pytorch_rnn_example.rst | 4 ++++
 docs/source/index.rst                        | 1 +
 docs/source/walkthrough.rst                  | 2 +-
 podium/datasets/iterator.py                  | 6 ++++--
 5 files changed, 14 insertions(+), 7 deletions(-)
 create mode 100644 docs/source/examples/pytorch_rnn_example.rst

diff --git a/docs/source/advanced.rst b/docs/source/advanced.rst
index ee77b13d..d0815964 100644
--- a/docs/source/advanced.rst
+++ b/docs/source/advanced.rst
@@ -448,18 +448,18 @@ The ``bucket_sort_key`` function defines how the instances in the dataset should
   For Iterator, padding = 148141 out of 281696 = 52.588961149608096%
   For BucketIterator, padding = 2125 out of 135680 = 1.5661851415094339%
 
-As we can see, the difference between using a regular Iterator and a BucketIterator is massive. Not only do we reduce the amount of padding, we have reduced the total amount of tokens processed by about 50%. The SST dataset, however, is a relatively small dataset so this experiment might be a bit biased. Let's take a look at the same statistics for the :class:`podium.datasets.impl.IMDB` dataset. After changing the highligted data loading line in the first snippet to:
+As we can see, the difference between using a regular Iterator and a BucketIterator is massive. Not only do we reduce the amount of padding, we have reduced the total amount of tokens processed by about 50%. The SST dataset, however, is a relatively small dataset so this experiment might be a bit biased. Let's take a look at the same statistics for the :class:`podium.datasets.impl.IMDB` dataset. After changing the data loading line in the first snippet to:
 
 .. code-block:: rest
 
-  train, test = IMDB.get_dataset_splits(fields=fields)
+  >>> train, test = IMDB.get_dataset_splits(fields=fields)
 
 And re-running the code, we obtain the following, still significant improvement:
 
 .. code-block:: rest
 
-  For Iterator, padding = 13569936 out of 19414616 = 69.89546432440385%
-  For BucketIterator, padding = 259800 out of 6104480 = 4.255890755641758%
+  For Iterator, padding = 13569936 out of 19414616 = 69.89%
+  For BucketIterator, padding = 259800 out of 6104480 = 4.25%
 
 Generally, using bucketing when iterating over your NLP dataset is preferred and will save you quite a bit of processing time.
 
diff --git a/docs/source/examples/pytorch_rnn_example.rst b/docs/source/examples/pytorch_rnn_example.rst
new file mode 100644
index 00000000..fa7cc074
--- /dev/null
+++ b/docs/source/examples/pytorch_rnn_example.rst
@@ -0,0 +1,4 @@
+Pytorch RNN classifier
+=======================
+
+In this example, we will cover a simple RNN-based classifier model implemented in Pytorch. We will
\ No newline at end of file
diff --git a/docs/source/index.rst b/docs/source/index.rst
index d237452f..22b7b571 100644
--- a/docs/source/index.rst
+++ b/docs/source/index.rst
@@ -42,6 +42,7 @@ The documentation is organized in five parts:
     :caption: Full examples
 
     examples/tfidf_example.rst
+    examples/pytorch_rnn_example.rst
 
 .. toctree::
     :maxdepth: 2
diff --git a/docs/source/walkthrough.rst b/docs/source/walkthrough.rst
index 360531db..d4f70d18 100644
--- a/docs/source/walkthrough.rst
+++ b/docs/source/walkthrough.rst
@@ -113,7 +113,7 @@ Datasets from 🤗 need to either (1) be wrapped them in :class:`podium.datasets
   })}
 
 .. note::
-  Conversion from features to Fields is **automatically inferred** by default. This is a process which can be error prone, many assumptions have to be made. Nevertheless, it will work for basic use-cases.
+  Conversion from 🤗 dataset features to Fields is **automatically inferred** by default. This is a process which can be error prone. Nevertheless, it will work for basic use-cases.
   In general, we recommend you set the ``fields`` argument of ``from_dataset_dict``.
 
 When we load a 🤗 dataset, we internally perform automatic Field type inference and create Fields. While we expect these Fields to work in most cases, we recommend you try constructing your own.
diff --git a/podium/datasets/iterator.py b/podium/datasets/iterator.py
index c26a60d4..1a81b151 100644
--- a/podium/datasets/iterator.py
+++ b/podium/datasets/iterator.py
@@ -307,13 +307,15 @@ def __iter__(self) -> PythonIterator[Tuple[NamedTuple, NamedTuple]]:
             self._shuffler_state = self.get_internal_random_state()
             self._shuffler.shuffle(indices)
 
-        data = self._dataset[indices]
+        # This is extremely memory intensive for each iteration
+        # data = self._dataset[indices]
 
         # If iteration was stopped, continue where we left off
         start = self.iterations * self.batch_size
 
         for i in range(start, len(data), self.batch_size):
-            batch_instances = data[i : i + self.batch_size]
+            batch_indices = indices[i : i + self.batch_size]
+            batch_instances = data[batch_indices]
 
             if self._sort_key is not None:
                 batch_instances = batch_instances.sorted(key=self._sort_key)

From de5d083e2a69f7585e1043984745e1da01f5f9e6 Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Thu, 1 Apr 2021 02:16:17 +0200
Subject: [PATCH 05/26] Fix shuffling cost

---
 podium/datasets/iterator.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/podium/datasets/iterator.py b/podium/datasets/iterator.py
index 1a81b151..0319821d 100644
--- a/podium/datasets/iterator.py
+++ b/podium/datasets/iterator.py
@@ -315,7 +315,7 @@ def __iter__(self) -> PythonIterator[Tuple[NamedTuple, NamedTuple]]:
 
         for i in range(start, len(data), self.batch_size):
             batch_indices = indices[i : i + self.batch_size]
-            batch_instances = data[batch_indices]
+            batch_instances = self._dataset[batch_indices]
 
             if self._sort_key is not None:
                 batch_instances = batch_instances.sorted(key=self._sort_key)

From 91237895ea55410a38db331dde3df3f10624e4a3 Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Thu, 1 Apr 2021 11:08:14 +0200
Subject: [PATCH 06/26] Fix shuffling cost+

---
 podium/datasets/hf.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/podium/datasets/hf.py b/podium/datasets/hf.py
index 93b36cc8..c072cdfb 100644
--- a/podium/datasets/hf.py
+++ b/podium/datasets/hf.py
@@ -180,6 +180,7 @@ def _get_examples(self):
         yield from self
 
     def __getitem__(self, i):
+        print(f"Called with {i}")
         raw_examples = self.dataset[i]
 
         # Index or slice

From 78757a38ec36749850874a48796570e2fa522f0b Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Thu, 1 Apr 2021 11:19:38 +0200
Subject: [PATCH 07/26] Fix shuffling cost+

---
 podium/datasets/iterator.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/podium/datasets/iterator.py b/podium/datasets/iterator.py
index 0319821d..1747c273 100644
--- a/podium/datasets/iterator.py
+++ b/podium/datasets/iterator.py
@@ -313,7 +313,7 @@ def __iter__(self) -> PythonIterator[Tuple[NamedTuple, NamedTuple]]:
         # If iteration was stopped, continue where we left off
         start = self.iterations * self.batch_size
 
-        for i in range(start, len(data), self.batch_size):
+        for i in range(start, len(self._dataset), self.batch_size):
             batch_indices = indices[i : i + self.batch_size]
             batch_instances = self._dataset[batch_indices]
 

From 1bd5887345b2ae9cbcaf83da409b5b6506c078ac Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Thu, 1 Apr 2021 12:34:40 +0200
Subject: [PATCH 08/26] Fix shuffling cost+

---
 podium/datasets/hf.py       | 1 -
 podium/datasets/iterator.py | 3 ---
 2 files changed, 4 deletions(-)

diff --git a/podium/datasets/hf.py b/podium/datasets/hf.py
index c072cdfb..93b36cc8 100644
--- a/podium/datasets/hf.py
+++ b/podium/datasets/hf.py
@@ -180,7 +180,6 @@ def _get_examples(self):
         yield from self
 
     def __getitem__(self, i):
-        print(f"Called with {i}")
         raw_examples = self.dataset[i]
 
         # Index or slice
diff --git a/podium/datasets/iterator.py b/podium/datasets/iterator.py
index 1747c273..6cce1a7f 100644
--- a/podium/datasets/iterator.py
+++ b/podium/datasets/iterator.py
@@ -307,9 +307,6 @@ def __iter__(self) -> PythonIterator[Tuple[NamedTuple, NamedTuple]]:
             self._shuffler_state = self.get_internal_random_state()
             self._shuffler.shuffle(indices)
 
-        # This is extremely memory intensive for each iteration
-        # data = self._dataset[indices]
-
         # If iteration was stopped, continue where we left off
         start = self.iterations * self.batch_size
 

From 6c9b67b6a52a9b0055bfe503a8a59231e19bfe71 Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Thu, 1 Apr 2021 14:22:48 +0200
Subject: [PATCH 09/26] merge

---
 podium/datasets/iterator.py | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/podium/datasets/iterator.py b/podium/datasets/iterator.py
index b6a83fe0..ecb662b4 100644
--- a/podium/datasets/iterator.py
+++ b/podium/datasets/iterator.py
@@ -493,9 +493,14 @@ def __init__(self, dataset: DatasetBase = None, shuffle=True, add_padding=True):
             iterator. If set to ``False``, numericalized Fields will be
             returned as python lists of ``matrix_class`` instances.
         """
+
+        batch_size = 0
+        if dataset is not None:
+            batch_size=len(dataset)
+
         super().__init__(
             dataset=dataset,
-            batch_size=len(dataset),
+            batch_size=batch_size,
             shuffle=shuffle,
             disable_batch_matrix=not add_padding,
         )

From 2ed6b95809d895e09cc9a5e29b1023e90984ffce Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Thu, 1 Apr 2021 16:40:26 +0200
Subject: [PATCH 10/26] Finalize pytorch rnn example

---
 docs/source/examples/pytorch_rnn_example.rst | 271 ++++++++++++++++++-
 docs/source/quickstart.rst                   |   1 +
 docs/source/walkthrough.rst                  |   2 +
 podium/datasets/iterator.py                  |   2 +-
 4 files changed, 274 insertions(+), 2 deletions(-)

diff --git a/docs/source/examples/pytorch_rnn_example.rst b/docs/source/examples/pytorch_rnn_example.rst
index fa7cc074..991ee4f0 100644
--- a/docs/source/examples/pytorch_rnn_example.rst
+++ b/docs/source/examples/pytorch_rnn_example.rst
@@ -1,4 +1,273 @@
 Pytorch RNN classifier
 =======================
 
-In this example, we will cover a simple RNN-based classifier model implemented in Pytorch. We will
\ No newline at end of file
+In this example, we will cover a simple RNN-based classifier model implemented in Pytorch. We will use the IMDB dataset loaded from 🤗/datasets, preprocess it with Fields and train the model briefly.
+While having a GPU is not necessary it is recommended as otherwise training the model, even for a single epoch, will take a while.
+
+Loading a dataset from 🤗/datasets
+-----------------------------------
+
+As we have covered in :ref:`hf-loading`, we have implemented wrappers around 🤗 dataset classes to enable working with the plethora of datasets implemented therein. We will now briefly go through (1) loading a dataset from 🤗/datasets and (2) wrapping it in Podium classes.
+
+.. code-block:: python
+
+  >>> import datasets
+  >>> imdb = datasets.load_dataset('imdb')
+  >>> print(imdb)
+  DatasetDict({
+      train: Dataset({
+          features: ['text', 'label'],
+          num_rows: 25000
+      })
+      test: Dataset({
+          features: ['text', 'label'],
+          num_rows: 25000
+      })
+      unsupervised: Dataset({
+          features: ['text', 'label'],
+          num_rows: 50000
+      })
+  })
+  >>> from pprint import pprint
+  >>> pprint(imdb['train'].features)
+  {'label': ClassLabel(num_classes=2, names=['neg', 'pos'], names_file=None, id=None),
+   'text': Value(dtype='string', id=None)}
+
+By calling ``load_dataset`` the dataset was downloaded and cached on disk through the ``datasets`` library. The dataset has two splits we are interested in (``train`` and ``test``).
+The main thing we need to pay attention to are the ``features`` of the dataset, in this case ``text`` and ``label``. These features, or data columns, need to be mapped to (and processed by) Podium Fields.
+
+For convenience, we have implemented automatic ``Field`` type inference from 🤗 dataset features -- however it is far from perfect as we have to make many assumptions on the way. We will now wrap the IMDB dataset in Podium and show the automatically inferred Fields.
+
+.. code-block:: python
+
+  >>> from podium.datasets.hf import HFDatasetConverter as HF
+  >>> splits = HF.from_dataset_dict(imdb)
+  >>> imdb_train, imdb_test = splits['train'], splits['test']
+  >>> imdb_train.finalize_fields() # Construct the vocab
+  >>> print(*imdb_train.fields, sep="\n")
+  Field({
+      name: 'text',
+      keep_raw: False,
+      is_target: False,
+      vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 280619})
+  })
+  LabelField({
+      name: 'label',
+      keep_raw: False,
+      is_target: True
+  })
+
+Both of the Fields were constructed well, but there are a couple of drawbacks for this concrete dataset. Firstly, the size of the vocabulary is very large (``280619``) -- we would like to trim this down to a reasonable number as we won't be using subword tokenization in this example. 
+
+.. code-block:: python
+
+  >>> print(imdb_train[0])
+  Example({
+      text: (None, ['Bromwell', 'High', 'is', 'a', 'cartoon', 'comedy.', 'It', 'ran', 'at', 'the', 'same', 'time', 'as', 'some', 'other', 'programs', 'about', 'school', 'life,', 'such', 'as', '"Teachers".', 'My', '35', 'years', 'in', 'the', 'teaching', 'profession', 'lead', 'me', 'to', 'believe', 'that', 'Bromwell', "High's", 'satire', 'is', 'much', 'closer', 'to', 'reality', 'than', 'is', '"Teachers".', 'The', 'scramble', 'to', 'survive', 'financially,', 'the', 'insightful', 'students', 'who', 'can', 'see', 'right', 'through', 'their', 'pathetic', "teachers'", 'pomp,', 'the', 'pettiness', 'of', 'the', 'whole', 'situation,', 'all', 'remind', 'me', 'of', 'the', 'schools', 'I', 'knew', 'and', 'their', 'students.', 'When', 'I', 'saw', 'the', 'episode', 'in', 'which', 'a', 'student', 'repeatedly', 'tried', 'to', 'burn', 'down', 'the', 'school,', 'I', 'immediately', 'recalled', '.........', 'at', '..........', 'High.', 'A', 'classic', 'line:', 'INSPECTOR:', "I'm", 'here', 'to', 'sack', 'one', 'of', 'your', 'teachers.', 'STUDENT:', 'Welcome', 'to', 'Bromwell', 'High.', 'I', 'expect', 'that', 'many', 'adults', 'of', 'my', 'age', 'think', 'that', 'Bromwell', 'High', 'is', 'far', 'fetched.', 'What', 'a', 'pity', 'that', 'it', "isn't!"]),
+      label: (None, 1)
+  })
+
+When inspecting a concrete instance, there are a few more things to note. Firstly, IMDB instances can be quite long (on average around 200 tokens per instance), secondly, the text wasn't tokenized properly near sentence boundaries (due to using the default ``str.split`` tokenizer) and lastly, the text has varying casing.
+We will instead define our own Fields for the corresponding features, add posttokenization hooks which will transform the data, and use those Fields to replace the automatically inferred ones:
+
+.. code-block:: python
+
+  >>> from podium import Field, LabelField, Vocab
+  >>> 
+  >>> # Lowercasing as a post-tokenization hook
+  >>> def lowercase(raw, tokenized):
+  ...   return raw, [token.lower() for token in tokenized]
+  >>>
+  >>> # Truncating as a post-tokenization hook
+  >>> def truncate(raw, tokenized, max_length=200):
+  ...     return raw, tokenized[:max_length]
+  >>>
+  >>> vocab = Vocab(max_size=10000)
+  >>> text = Field(name="text", 
+  ...              numericalizer=vocab,
+  ...              include_lengths=True,
+  ...              tokenizer="spacy-en_core_web_sm",
+  ...              posttokenize_hooks=[truncate, lowercase])
+  >>> 
+  >>> # The labels are already mapped to indices in /datasets so we will
+  >>> # pass them through
+  >>> label = LabelField(name="label", numericalizer=lambda x: x)
+  >>> fields = {
+  ...     'text': text,
+  ...     'label': label
+  ... }
+  >>> 
+  >>> # Use the given Fields to load the dataset again
+  >>> splits = HF.from_dataset_dict(imdb, fields=fields)
+  >>> imdb_train, imdb_test = splits['train'], splits['test']
+  >>> imdb_train.finalize_fields()
+  >>> print(imdb_train)
+  HFDatasetConverter({
+      dataset_name: imdb,
+      size: 25000,
+      fields: [
+              Field({
+                  name: 'text',
+                  keep_raw: False,
+                  is_target: False,
+                  vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 10000})
+              }),
+              LabelField({
+                  name: 'label',
+                  keep_raw: False,
+                  is_target: True
+              })
+      
+      ]
+  })
+  >>> print(imdb_train[0])
+  Example({
+      text: (None, ['bromwell', 'high', 'is', 'a', 'cartoon', 'comedy', '.', 'it', 'ran', 'at', 'the', 'same', 'time', 'as', 'some', 'other', 'programs', 'about', 'school', 'life', ',', 'such', 'as', '"', 'teachers', '"', '.', 'my', '35', 'years', 'in', 'the', 'teaching', 'profession', 'lead', 'me', 'to', 'believe', 'that', 'bromwell', 'high', "'s", 'satire', 'is', 'much', 'closer', 'to', 'reality', 'than', 'is', '"', 'teachers', '"', '.', 'the', 'scramble', 'to', 'survive', 'financially', ',', 'the', 'insightful', 'students', 'who', 'can', 'see', 'right', 'through', 'their', 'pathetic', 'teachers', "'", 'pomp', ',', 'the', 'pettiness', 'of', 'the', 'whole', 'situation', ',', 'all', 'remind', 'me', 'of', 'the', 'schools', 'i', 'knew', 'and', 'their', 'students', '.', 'when', 'i', 'saw', 'the', 'episode', 'in', 'which', 'a', 'student', 'repeatedly', 'tried', 'to', 'burn', 'down', 'the', 'school', ',', 'i', 'immediately', 'recalled', '.........', 'at', '..........', 'high', '.', 'a', 'classic', 'line', ':', 'inspector', ':', 'i', "'m", 'here', 'to', 'sack', 'one', 'of', 'your', 'teachers', '.', 'student', ':', 'welcome', 'to', 'bromwell', 'high', '.', 'i', 'expect', 'that', 'many', 'adults', 'of', 'my', 'age', 'think', 'that', 'bromwell', 'high', 'is', 'far', 'fetched', '.', 'what', 'a', 'pity', 'that', 'it', 'is', "n't", '!']),
+      label: (None, 1)
+  })
+
+Here, we can see the effect of our hooks and using the spacy tokenizer. Now our dataset will be a bit cleaner to work with. Some data cleaning would still be desired, such as removing tokens which only contain punctuation, but we leave this exercise to the reader :)
+
+Loading pretrained embeddings
+-----------------------------
+In most use-cases, we want to use pre-trained word embeddings along with our neural model. With Podium, this process is very simple. If your field uses a vocabulary, it has already built an inventory of tokens for your dataset.
+
+For example, we will use the `GloVe <https://nlp.stanford.edu/projects/glove/>`__ vectors. You can read more about loading pretrained vectors in :ref:`pretrained`, but the procedure to load these vectors has two steps: (1) initialize the vector class, which sets all the required paths and (2) obtain the vectors for a pre-defined list of words by calling ``load_vocab``.
+
+.. code-block:: python
+
+  >>> from podium.vectorizers import GloVe
+  >>> vocab = fields['text'].vocab
+  >>> glove = GloVe()
+  >>> embeddings = glove.load_vocab(vocab)
+  >>> print(f"For vocabulary of size: {len(vocab)} loaded embedding matrix of shape: {embeddings.shape}")
+  For vocabulary of size: 10000 loaded embedding matrix of shape: (10000, 300)
+  >>> # We can obtain vectors for a single word (given the word is loaded) like this:
+  >>> word = "sport"
+  >>> print(f"Vector for {word}: {glove.token_to_vector(word)}")
+  Vector for sport: [ 0.34566    0.15934    0.48444   -0.13693    0.18737    0.2678
+ -0.39159    0.4931    -0.76111   -1.4586     0.41475    0.55837
+  ...
+  0.13802    0.36619    0.19734    0.35701   -0.42228   -0.25242
+ -0.050651  -0.041129   0.15092    0.22084    0.52252   -0.27224  ]
+
+Defining a simple neural model in Pytorch
+------------------------------------------
+
+In this section, we will implement a very simple neural classification model -- a 2-layer BiGRU with a single hidden layer classifier on top of its last hidden state. Many improvements to the model can be made, but this is not our current focus.
+
+.. code-block:: python
+
+  >>> import torch
+  >>> import torch.nn as nn
+  >>> import torch.nn.functional as F
+  >>> 
+  >>> from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence
+  >>> 
+  >>> class RNNClassifier(nn.Module):
+  ...     def __init__(self, embedding, embed_dim=300, hidden_dim=300, num_labels=2):
+  ...       super(NLIModel, self).__init__()
+  ...       self.embedding = embedding
+  ...       self.encoder = nn.GRU(
+  ...             input_size=embed_dim,
+  ...             hidden_size=hidden_dim,
+  ...             num_layers=2,
+  ...             bidirectional=True,
+  ...             dropout=0.3
+  ...       )
+  ...       self.decoder = nn.Sequential(
+  ...             nn.Linear(2*hidden_dim, hidden_dim),
+  ...             nn.Tanh(),
+  ...             nn.Linear(hidden_dim, num_labels)
+  ...       )
+  ...
+  ...     def forward(self, x, lengths):
+  ...         e = self.embedding(x)
+  ...         h_pack = pack_padded_sequence(e, 
+  ...                                       lengths,
+  ...                                       enforce_sorted=False,
+  ...                                       batch_first=True)
+  ...
+  ...         _, h = self.encoder(h_pack) # [2L x B x H]
+  ...
+  ...         # Concat last state of left and right directions
+  ...         h = torch.cat([h[-1], h[-2]], dim=-1) # [B x 2H]
+  ...         return self.decoder(h)
+
+There. We will now define the prerequisites for pytorch model training, where we will use a GPU for speed, however running the model for one epoch will is possible albeit time-consuing even without a GPU.
+
+.. code-block:: python
+
+  >>> embed_dim = 300
+  >>> padding_index = text.vocab.get_padding_index()
+  >>> embedding_matrix = nn.Embedding(len(text.vocab), embed_dim,
+  ...                                 padding_idx=padding_index)
+  >>> # Copy the pretrained GloVe word embeddings
+  >>> embedding_matrix.weight.data.copy_(torch.from_numpy(embeddings))
+  >>>
+  >>> device = torch.device("cuda:0")
+  >>> model = RNNClassifier(embedding_matrix)
+  >>> model = model.to(device)
+  >>> criterion = nn.CrossEntropyLoss()
+  >>> optimizer = torch.optim.Adam(model.parameters())
+
+Now that we have the model setup code ready, we will first define helper method to measure accuracy of our model after each epoch:
+
+.. code-block:: python
+
+  >>> import numpy as np
+  >>> def update_stats(accuracy, confusion_matrix, logits, y):
+  ...     _, max_ind = torch.max(logits, 1)
+  ...     equal = torch.eq(max_ind, y)
+  ...     correct = int(torch.sum(equal))
+  ... 
+  ...     for j, i in zip(max_ind, y):
+  ...         confusion_matrix[int(i),int(j)]+=1
+  ...     return accuracy + correct, confusion_matrix
+
+and now the training loop for the model:
+
+.. code-block:: python
+
+  >>> import tqdm
+  >>> def train(model, data, optimizer, criterion, num_labels):
+  ...     model.train()
+  ...     accuracy, confusion_matrix = 0, np.zeros((num_labels, num_labels), dtype=int)
+  ...     for batch_num, batch in tqdm.tqdm(enumerate(data), total=len(data)):
+  ...         x, lens = batch.text
+  ...         y = batch.label
+  ...         logits = model(x, lens)
+  ...         accuracy, confusion_matrix = update_stats(accuracy, confusion_matrix, logits, y)
+  ...         loss = criterion(logits, y.squeeze())
+  ...         loss.backward()
+  ...         optimizer.step()
+  ...     print("[Accuracy]: {}/{} : {:.3f}%".format(
+  ...           accuracy, len(data)*data.batch_size, accuracy / len(data) / data.batch_size * 100))
+  ...     return accuracy, confusion_matrix
+
+and now, we are done with our model code. Let's turn back to Podium and see how we can set up batching for our training loop to start ticking.
+
+Minibatching data in Podium
+--------------------------------
+
+We have covered batching data in :ref:`minibatching` and advanced batching through bucketing in :ref:`bucketing`. We will use the plain Iterator and leave bucketing for you to change to see how much the model speeds up when minimizing padding. One change we would like to do when iterating over data is to obtain the data matrices as torch tensors on the ``device`` we defined previously. We will now demonstrate how to do this by setting the ``matrix_class`` argument of the :class:`podium.datasets.Iterator`\:
+
+.. code-block:: python
+
+  >>> from podium import Iterator
+  >>> # Closure for converting data to given device
+  >>> def gpu_tensor(data):
+  ...     return torch.tensor(data).to(device)
+  >>> # Initialize our iterator
+  >>> train_iter = Iterator(imdb_train, batch_size=32, matrix_class=gpu_tensor)
+  >>>
+  >>> epochs = 5
+  >>> for epoch in range(epochs):
+  >>>     train(model, train_iter, optimizer, criterion, num_labels=2)
+  [Accuracy]: 20050/25024 : 80.123%
+  [Accuracy]: 22683/25024 : 90.645%
+  [Accuracy]: 23709/25024 : 94.745%
+  [Accuracy]: 24323/25024 : 97.199%
+  [Accuracy]: 24595/25024 : 98.286%
+
+And we are done! In our case, the model takes about one minute per epoch on a GPU, but this can be sped up by using bucketing, which we recommend you try out yourself.
diff --git a/docs/source/quickstart.rst b/docs/source/quickstart.rst
index 715f3452..2e4847de 100644
--- a/docs/source/quickstart.rst
+++ b/docs/source/quickstart.rst
@@ -189,6 +189,7 @@ Since our example dataset is small, we can set ``add_padding=True``, which cause
   When obtaining larger datasets as a single batch, we recommend leaving ``add_padding=False`` (default) or your entire dataset will be padded to the length of the longest instance, causing memory issues.
   When set to ``False``, the output of each Field will be a list instead of a matrix.
 
+.. _minibatching:
 
 Minibatching data
 -----------------------
diff --git a/docs/source/walkthrough.rst b/docs/source/walkthrough.rst
index 4124f7e1..d91ae5de 100644
--- a/docs/source/walkthrough.rst
+++ b/docs/source/walkthrough.rst
@@ -486,6 +486,8 @@ Since datasets can contain multiple input Fields, it is not trivial to determine
 
 And here we can see, that even for our small, two-instance batch, the elements in the batch are now properly sorted according to length.
 
+.. _pretrained:
+
 Loading pretrained word vectors
 -------------------------------
 
diff --git a/podium/datasets/iterator.py b/podium/datasets/iterator.py
index ecb662b4..97e23c30 100644
--- a/podium/datasets/iterator.py
+++ b/podium/datasets/iterator.py
@@ -496,7 +496,7 @@ def __init__(self, dataset: DatasetBase = None, shuffle=True, add_padding=True):
 
         batch_size = 0
         if dataset is not None:
-            batch_size=len(dataset)
+            batch_size = len(dataset)
 
         super().__init__(
             dataset=dataset,

From 0183c38a34a96b4fbe9d1125fe2d88bd4b4b85eb Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Thu, 1 Apr 2021 16:56:43 +0200
Subject: [PATCH 11/26] Finalize examples

---
 docs/source/_static/js/custom.js              |   2 +
 docs/source/notebooks/advanced.ipynb          |   6 +-
 .../notebooks/pytorch_rnn_example.ipynb       | 500 ++++++++++++++++++
 docs/source/notebooks/quickstart.ipynb        |   7 +
 docs/source/notebooks/walkthrough.ipynb       |   9 +-
 .../{examples => }/pytorch_rnn_example.rst    |   0
 .../scripts/convert_doc_to_notebooks.py       |   4 +
 docs/source/{examples => }/tfidf_example.rst  |   2 +-
 8 files changed, 525 insertions(+), 5 deletions(-)
 create mode 100644 docs/source/notebooks/pytorch_rnn_example.ipynb
 rename docs/source/{examples => }/pytorch_rnn_example.rst (100%)
 rename docs/source/{examples => }/tfidf_example.rst (99%)

diff --git a/docs/source/_static/js/custom.js b/docs/source/_static/js/custom.js
index 749084d9..26839340 100644
--- a/docs/source/_static/js/custom.js
+++ b/docs/source/_static/js/custom.js
@@ -16,6 +16,8 @@ const hasNotebook = [
     "advanced",
     "preprocessing",
     "walkthrough",
+    "tfidf_example",
+    "pytorch_rnn_example"
 ]
 
 function addIcon() {
diff --git a/docs/source/notebooks/advanced.ipynb b/docs/source/notebooks/advanced.ipynb
index c9f7d874..639d6af0 100644
--- a/docs/source/notebooks/advanced.ipynb
+++ b/docs/source/notebooks/advanced.ipynb
@@ -957,7 +957,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "As we can see, the difference between using a regular Iterator and a BucketIterator is massive. Not only do we reduce the amount of padding, we have reduced the total amount of tokens processed by about 50%. The SST dataset, however, is a relatively small dataset so this experiment might be a bit biased. Let's take a look at the same statistics for the `IMDB` dataset. After changing the highligted data loading line in the first snippet to:"
+    "As we can see, the difference between using a regular Iterator and a BucketIterator is massive. Not only do we reduce the amount of padding, we have reduced the total amount of tokens processed by about 50%. The SST dataset, however, is a relatively small dataset so this experiment might be a bit biased. Let's take a look at the same statistics for the `IMDB` dataset. After changing the data loading line in the first snippet to:"
    ]
   },
   {
@@ -981,8 +981,8 @@
    "metadata": {},
    "source": [
     "```bash\n",
-    "For Iterator, padding = 13569936 out of 19414616 = 69.89546432440385%\n",
-    "For BucketIterator, padding = 259800 out of 6104480 = 4.255890755641758%\n",
+    "For Iterator, padding = 13569936 out of 19414616 = 69.89%\n",
+    "For BucketIterator, padding = 259800 out of 6104480 = 4.25%\n",
     "```"
    ]
   },
diff --git a/docs/source/notebooks/pytorch_rnn_example.ipynb b/docs/source/notebooks/pytorch_rnn_example.ipynb
new file mode 100644
index 00000000..a29d6256
--- /dev/null
+++ b/docs/source/notebooks/pytorch_rnn_example.ipynb
@@ -0,0 +1,500 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Podium installation\n",
+    "! pip install podium-nlp\n",
+    "# To install from source instead of the last release, comment the command above and uncomment the following one.\n",
+    "# ! pip install git+https://github.com/takelab/podium"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Pytorch RNN classifier"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "In this example, we will cover a simple RNN-based classifier model implemented in Pytorch. We will use the IMDB dataset loaded from 🤗/datasets, preprocess it with Fields and train the model briefly.\n",
+    "While having a GPU is not necessary it is recommended as otherwise training the model, even for a single epoch, will take a while."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Loading a dataset from 🤗/datasets"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "As we have covered in [Loading 🤗 datasets](http://takelab.fer.hr/podium/walkthrough.html#hf-loading), we have implemented wrappers around 🤗 dataset classes to enable working with the plethora of datasets implemented therein. We will now briefly go through (1) loading a dataset from 🤗/datasets and (2) wrapping it in Podium classes."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "DatasetDict({\n",
+       "    train: Dataset({\n",
+       "        features: ['text', 'label'],\n",
+       "        num_rows: 25000\n",
+       "    })\n",
+       "    test: Dataset({\n",
+       "        features: ['text', 'label'],\n",
+       "        num_rows: 25000\n",
+       "    })\n",
+       "    unsupervised: Dataset({\n",
+       "        features: ['text', 'label'],\n",
+       "        num_rows: 50000\n",
+       "    })\n",
+       "})\n",
+       "{'label': ClassLabel(num_classes=2, names=['neg', 'pos'], names_file=None, id=None),\n",
+       " 'text': Value(dtype='string', id=None)}"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "import datasets\n",
+    "imdb = datasets.load_dataset('imdb')\n",
+    "print(imdb)\n",
+    "from pprint import pprint\n",
+    "pprint(imdb['train'].features)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "By calling `load_dataset` the dataset was downloaded and cached on disk through the `datasets` library. The dataset has two splits we are interested in (`train` and `test`).\n",
+    "The main thing we need to pay attention to are the `features` of the dataset, in this case `text` and `label`. These features, or data columns, need to be mapped to (and processed by) Podium Fields.\n",
+    "\n",
+    "For convenience, we have implemented automatic `Field` type inference from 🤗 dataset features -- however it is far from perfect as we have to make many assumptions on the way. We will now wrap the IMDB dataset in Podium and show the automatically inferred Fields."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "Field({\n",
+       "    name: 'text',\n",
+       "    keep_raw: False,\n",
+       "    is_target: False,\n",
+       "    vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 280619})\n",
+       "})\n",
+       "LabelField({\n",
+       "    name: 'label',\n",
+       "    keep_raw: False,\n",
+       "    is_target: True\n",
+       "})"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from podium.datasets.hf import HFDatasetConverter as HF\n",
+    "splits = HF.from_dataset_dict(imdb)\n",
+    "imdb_train, imdb_test = splits['train'], splits['test']\n",
+    "imdb_train.finalize_fields() # Construct the vocab\n",
+    "print(*imdb_train.fields, sep=\"\\n\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Both of the Fields were constructed well, but there are a couple of drawbacks for this concrete dataset. Firstly, the size of the vocabulary is very large (`280619`) -- we would like to trim this down to a reasonable number as we won't be using subword tokenization in this example. "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "Example({\n",
+       "    text: (None, ['Bromwell', 'High', 'is', 'a', 'cartoon', 'comedy.', 'It', 'ran', 'at', 'the', 'same', 'time', 'as', 'some', 'other', 'programs', 'about', 'school', 'life,', 'such', 'as', '\"Teachers\".', 'My', '35', 'years', 'in', 'the', 'teaching', 'profession', 'lead', 'me', 'to', 'believe', 'that', 'Bromwell', \"High's\", 'satire', 'is', 'much', 'closer', 'to', 'reality', 'than', 'is', '\"Teachers\".', 'The', 'scramble', 'to', 'survive', 'financially,', 'the', 'insightful', 'students', 'who', 'can', 'see', 'right', 'through', 'their', 'pathetic', \"teachers'\", 'pomp,', 'the', 'pettiness', 'of', 'the', 'whole', 'situation,', 'all', 'remind', 'me', 'of', 'the', 'schools', 'I', 'knew', 'and', 'their', 'students.', 'When', 'I', 'saw', 'the', 'episode', 'in', 'which', 'a', 'student', 'repeatedly', 'tried', 'to', 'burn', 'down', 'the', 'school,', 'I', 'immediately', 'recalled', '.........', 'at', '..........', 'High.', 'A', 'classic', 'line:', 'INSPECTOR:', \"I'm\", 'here', 'to', 'sack', 'one', 'of', 'your', 'teachers.', 'STUDENT:', 'Welcome', 'to', 'Bromwell', 'High.', 'I', 'expect', 'that', 'many', 'adults', 'of', 'my', 'age', 'think', 'that', 'Bromwell', 'High', 'is', 'far', 'fetched.', 'What', 'a', 'pity', 'that', 'it', \"isn't!\"]),\n",
+       "    label: (None, 1)\n",
+       "})"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "print(imdb_train[0])"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "When inspecting a concrete instance, there are a few more things to note. Firstly, IMDB instances can be quite long (on average around 200 tokens per instance), secondly, the text wasn't tokenized properly near sentence boundaries (due to using the default `str.split` tokenizer) and lastly, the text has varying casing.\n",
+    "We will instead define our own Fields for the corresponding features, add posttokenization hooks which will transform the data, and use those Fields to replace the automatically inferred ones:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "HFDatasetConverter({\n",
+       "    dataset_name: imdb,\n",
+       "    size: 25000,\n",
+       "    fields: [\n",
+       "            Field({\n",
+       "                name: 'text',\n",
+       "                keep_raw: False,\n",
+       "                is_target: False,\n",
+       "                vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 10000})\n",
+       "            }),\n",
+       "            LabelField({\n",
+       "                name: 'label',\n",
+       "                keep_raw: False,\n",
+       "                is_target: True\n",
+       "            })\n",
+       "    \n",
+       "    ]\n",
+       "})\n",
+       "Example({\n",
+       "    text: (None, ['bromwell', 'high', 'is', 'a', 'cartoon', 'comedy', '.', 'it', 'ran', 'at', 'the', 'same', 'time', 'as', 'some', 'other', 'programs', 'about', 'school', 'life', ',', 'such', 'as', '\"', 'teachers', '\"', '.', 'my', '35', 'years', 'in', 'the', 'teaching', 'profession', 'lead', 'me', 'to', 'believe', 'that', 'bromwell', 'high', \"'s\", 'satire', 'is', 'much', 'closer', 'to', 'reality', 'than', 'is', '\"', 'teachers', '\"', '.', 'the', 'scramble', 'to', 'survive', 'financially', ',', 'the', 'insightful', 'students', 'who', 'can', 'see', 'right', 'through', 'their', 'pathetic', 'teachers', \"'\", 'pomp', ',', 'the', 'pettiness', 'of', 'the', 'whole', 'situation', ',', 'all', 'remind', 'me', 'of', 'the', 'schools', 'i', 'knew', 'and', 'their', 'students', '.', 'when', 'i', 'saw', 'the', 'episode', 'in', 'which', 'a', 'student', 'repeatedly', 'tried', 'to', 'burn', 'down', 'the', 'school', ',', 'i', 'immediately', 'recalled', '.........', 'at', '..........', 'high', '.', 'a', 'classic', 'line', ':', 'inspector', ':', 'i', \"'m\", 'here', 'to', 'sack', 'one', 'of', 'your', 'teachers', '.', 'student', ':', 'welcome', 'to', 'bromwell', 'high', '.', 'i', 'expect', 'that', 'many', 'adults', 'of', 'my', 'age', 'think', 'that', 'bromwell', 'high', 'is', 'far', 'fetched', '.', 'what', 'a', 'pity', 'that', 'it', 'is', \"n't\", '!']),\n",
+       "    label: (None, 1)\n",
+       "})"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from podium import Field, LabelField, Vocab\n",
+    "\n",
+    "# Lowercasing as a post-tokenization hook\n",
+    "def lowercase(raw, tokenized):\n",
+    "  return raw, [token.lower() for token in tokenized]\n",
+    "\n",
+    "# Truncating as a post-tokenization hook\n",
+    "def truncate(raw, tokenized, max_length=200):\n",
+    "    return raw, tokenized[:max_length]\n",
+    "\n",
+    "vocab = Vocab(max_size=10000)\n",
+    "text = Field(name=\"text\", \n",
+    "             numericalizer=vocab,\n",
+    "             include_lengths=True,\n",
+    "             tokenizer=\"spacy-en_core_web_sm\",\n",
+    "             posttokenize_hooks=[truncate, lowercase])\n",
+    "\n",
+    "# The labels are already mapped to indices in /datasets so we will\n",
+    "# pass them through\n",
+    "label = LabelField(name=\"label\", numericalizer=lambda x: x)\n",
+    "fields = {\n",
+    "    'text': text,\n",
+    "    'label': label\n",
+    "}\n",
+    "\n",
+    "# Use the given Fields to load the dataset again\n",
+    "splits = HF.from_dataset_dict(imdb, fields=fields)\n",
+    "imdb_train, imdb_test = splits['train'], splits['test']\n",
+    "imdb_train.finalize_fields()\n",
+    "print(imdb_train)\n",
+    "print(imdb_train[0])"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Here, we can see the effect of our hooks and using the spacy tokenizer. Now our dataset will be a bit cleaner to work with. Some data cleaning would still be desired, such as removing tokens which only contain punctuation, but we leave this exercise to the reader :)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Loading pretrained embeddings"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "In most use-cases, we want to use pre-trained word embeddings along with our neural model. With Podium, this process is very simple. If your field uses a vocabulary, it has already built an inventory of tokens for your dataset.\n",
+    "\n",
+    "For example, we will use the [GloVe](https://nlp.stanford.edu/projects/glove/) vectors. You can read more about loading pretrained vectors in [Loading pretrained word vectors](http://takelab.fer.hr/podium/walkthrough.html#pretrained), but the procedure to load these vectors has two steps: (1) initialize the vector class, which sets all the required paths and (2) obtain the vectors for a pre-defined list of words by calling `load_vocab`."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "For vocabulary of size: 10000 loaded embedding matrix of shape: (10000, 300)\n",
+       "Vector for sport: [ 0.34566    0.15934    0.48444   -0.13693    0.18737    0.2678"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from podium.vectorizers import GloVe\n",
+    "vocab = fields['text'].vocab\n",
+    "glove = GloVe()\n",
+    "embeddings = glove.load_vocab(vocab)\n",
+    "print(f\"For vocabulary of size: {len(vocab)} loaded embedding matrix of shape: {embeddings.shape}\")\n",
+    "# We can obtain vectors for a single word (given the word is loaded) like this:\n",
+    "word = \"sport\"\n",
+    "print(f\"Vector for {word}: {glove.token_to_vector(word)}\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    " -0.39159    0.4931    -0.76111   -1.4586     0.41475    0.55837\n",
+    "  ...\n",
+    "  0.13802    0.36619    0.19734    0.35701   -0.42228   -0.25242\n",
+    " -0.050651  -0.041129   0.15092    0.22084    0.52252   -0.27224  ]"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Defining a simple neural model in Pytorch"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "In this section, we will implement a very simple neural classification model -- a 2-layer BiGRU with a single hidden layer classifier on top of its last hidden state. Many improvements to the model can be made, but this is not our current focus."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "import torch.nn as nn\n",
+    "import torch.nn.functional as F\n",
+    "\n",
+    "from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence\n",
+    "\n",
+    "class RNNClassifier(nn.Module):\n",
+    "    def __init__(self, embedding, embed_dim=300, hidden_dim=300, num_labels=2):\n",
+    "      super(NLIModel, self).__init__()\n",
+    "      self.embedding = embedding\n",
+    "      self.encoder = nn.GRU(\n",
+    "            input_size=embed_dim,\n",
+    "            hidden_size=hidden_dim,\n",
+    "            num_layers=2,\n",
+    "            bidirectional=True,\n",
+    "            dropout=0.3\n",
+    "      )\n",
+    "      self.decoder = nn.Sequential(\n",
+    "            nn.Linear(2*hidden_dim, hidden_dim),\n",
+    "            nn.Tanh(),\n",
+    "            nn.Linear(hidden_dim, num_labels)\n",
+    "      )\n",
+    "\n",
+    "    def forward(self, x, lengths):\n",
+    "        e = self.embedding(x)\n",
+    "        h_pack = pack_padded_sequence(e, \n",
+    "                                      lengths,\n",
+    "                                      enforce_sorted=False,\n",
+    "                                      batch_first=True)\n",
+    "\n",
+    "        _, h = self.encoder(h_pack) # [2L x B x H]\n",
+    "\n",
+    "        # Concat last state of left and right directions\n",
+    "        h = torch.cat([h[-1], h[-2]], dim=-1) # [B x 2H]\n",
+    "        return self.decoder(h)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "There. We will now define the prerequisites for pytorch model training, where we will use a GPU for speed, however running the model for one epoch will is possible albeit time-consuing even without a GPU."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "embed_dim = 300\n",
+    "padding_index = text.vocab.get_padding_index()\n",
+    "embedding_matrix = nn.Embedding(len(text.vocab), embed_dim,\n",
+    "                                padding_idx=padding_index)\n",
+    "# Copy the pretrained GloVe word embeddings\n",
+    "embedding_matrix.weight.data.copy_(torch.from_numpy(embeddings))\n",
+    "\n",
+    "device = torch.device(\"cuda:0\")\n",
+    "model = RNNClassifier(embedding_matrix)\n",
+    "model = model.to(device)\n",
+    "criterion = nn.CrossEntropyLoss()\n",
+    "optimizer = torch.optim.Adam(model.parameters())"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Now that we have the model setup code ready, we will first define helper method to measure accuracy of our model after each epoch:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import numpy as np\n",
+    "def update_stats(accuracy, confusion_matrix, logits, y):\n",
+    "    _, max_ind = torch.max(logits, 1)\n",
+    "    equal = torch.eq(max_ind, y)\n",
+    "    correct = int(torch.sum(equal))\n",
+    "\n",
+    "    for j, i in zip(max_ind, y):\n",
+    "        confusion_matrix[int(i),int(j)]+=1\n",
+    "    return accuracy + correct, confusion_matrix"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "and now the training loop for the model:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import tqdm\n",
+    "def train(model, data, optimizer, criterion, num_labels):\n",
+    "    model.train()\n",
+    "    accuracy, confusion_matrix = 0, np.zeros((num_labels, num_labels), dtype=int)\n",
+    "    for batch_num, batch in tqdm.tqdm(enumerate(data), total=len(data)):\n",
+    "        x, lens = batch.text\n",
+    "        y = batch.label\n",
+    "        logits = model(x, lens)\n",
+    "        accuracy, confusion_matrix = update_stats(accuracy, confusion_matrix, logits, y)\n",
+    "        loss = criterion(logits, y.squeeze())\n",
+    "        loss.backward()\n",
+    "        optimizer.step()\n",
+    "    print(\"[Accuracy]: {}/{} : {:.3f}%\".format(\n",
+    "          accuracy, len(data)*data.batch_size, accuracy / len(data) / data.batch_size * 100))\n",
+    "    return accuracy, confusion_matrix"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "and now, we are done with our model code. Let's turn back to Podium and see how we can set up batching for our training loop to start ticking."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Minibatching data in Podium"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "We have covered batching data in [Minibatching data](http://takelab.fer.hr/podium/quickstart.html#minibatching) and advanced batching through bucketing in [Bucketing instances when iterating](http://takelab.fer.hr/podium/advanced.html#bucketing). We will use the plain Iterator and leave bucketing for you to change to see how much the model speeds up when minimizing padding. One change we would like to do when iterating over data is to obtain the data matrices as torch tensors on the `device` we defined previously. We will now demonstrate how to do this by setting the `matrix_class` argument of the `Iterator`:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "[Accuracy]: 20050/25024 : 80.123%\n",
+       "[Accuracy]: 22683/25024 : 90.645%\n",
+       "[Accuracy]: 23709/25024 : 94.745%\n",
+       "[Accuracy]: 24323/25024 : 97.199%\n",
+       "[Accuracy]: 24595/25024 : 98.286%"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from podium import Iterator\n",
+    "# Closure for converting data to given device\n",
+    "def gpu_tensor(data):\n",
+    "    return torch.tensor(data).to(device)\n",
+    "# Initialize our iterator\n",
+    "train_iter = Iterator(imdb_train, batch_size=32, matrix_class=gpu_tensor)\n",
+    "\n",
+    "epochs = 5\n",
+    "for epoch in range(epochs):\n",
+    "    train(model, train_iter, optimizer, criterion, num_labels=2)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "And we are done! In our case, the model takes about one minute per epoch on a GPU, but this can be sped up by using bucketing, which we recommend you try out yourself."
+   ]
+  }
+ ],
+ "metadata": {},
+ "nbformat": 4,
+ "nbformat_minor": 4
+}
diff --git a/docs/source/notebooks/quickstart.ipynb b/docs/source/notebooks/quickstart.ipynb
index 7d23db64..b42e799d 100644
--- a/docs/source/notebooks/quickstart.ipynb
+++ b/docs/source/notebooks/quickstart.ipynb
@@ -432,6 +432,13 @@
     "> When set to `False`, the output of each Field will be a list instead of a matrix."
    ]
   },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "<a id='minibatching'></a>"
+   ]
+  },
   {
    "cell_type": "markdown",
    "metadata": {},
diff --git a/docs/source/notebooks/walkthrough.ipynb b/docs/source/notebooks/walkthrough.ipynb
index f78d4d45..b8a1dc3a 100644
--- a/docs/source/notebooks/walkthrough.ipynb
+++ b/docs/source/notebooks/walkthrough.ipynb
@@ -220,7 +220,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "> **NOTE:** Conversion from features to Fields is **automatically inferred** by default. This is a process which can be error prone, many assumptions have to be made. Nevertheless, it will work for basic use-cases.\n",
+    "> **NOTE:** Conversion from 🤗 dataset features to Fields is **automatically inferred** by default. This is a process which can be error prone. Nevertheless, it will work for basic use-cases.\n",
     "> In general, we recommend you set the `fields` argument of `from_dataset_dict`."
    ]
   },
@@ -961,6 +961,13 @@
     "And here we can see, that even for our small, two-instance batch, the elements in the batch are now properly sorted according to length."
    ]
   },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "<a id='pretrained'></a>"
+   ]
+  },
   {
    "cell_type": "markdown",
    "metadata": {},
diff --git a/docs/source/examples/pytorch_rnn_example.rst b/docs/source/pytorch_rnn_example.rst
similarity index 100%
rename from docs/source/examples/pytorch_rnn_example.rst
rename to docs/source/pytorch_rnn_example.rst
diff --git a/docs/source/scripts/convert_doc_to_notebooks.py b/docs/source/scripts/convert_doc_to_notebooks.py
index 1feb0731..897a823f 100644
--- a/docs/source/scripts/convert_doc_to_notebooks.py
+++ b/docs/source/scripts/convert_doc_to_notebooks.py
@@ -20,6 +20,8 @@
     "advanced.rst",
     "preprocessing.rst",
     "walkthrough.rst",
+    "pytorch_rnn_example.rst",
+    "tfidf_example.rst",
 ]
 
 _re_label = re.compile(r"\.\.\s+_([^:]*):")
@@ -556,6 +558,8 @@ def convert_all_tutorials(path_to_docs=None, path_to_dest=None):
     for file in TUTORIAL_FILES:
         notebook_name = os.path.splitext(file)[0] + ".ipynb"
         doc_file = os.path.join(path_to_docs, file)
+        print(doc_file)
+        print(os.path.exists(doc_file))
         notebook_file = os.path.join(path_to_dest, notebook_name)
         convert_rst_file_to_notebook(doc_file, notebook_file, origin_folder=path_to_docs, dest_folder=path_to_dest, additional_deps=ADDITIONAL_DEPS.get(file))
 
diff --git a/docs/source/examples/tfidf_example.rst b/docs/source/tfidf_example.rst
similarity index 99%
rename from docs/source/examples/tfidf_example.rst
rename to docs/source/tfidf_example.rst
index bc62b3a4..770a8440 100644
--- a/docs/source/examples/tfidf_example.rst
+++ b/docs/source/tfidf_example.rst
@@ -21,7 +21,7 @@ First, we will implement a minimalistic example without much additional preproce
 We have now loaded our dataset, finalized its Fields and obtained it as a batch of input and target data. What we need to do next is define the TF-IDF vectorization for each instance in the dataset. This is done by using our :class:`podium.vectorizers.TfIdfVectorizer`, which adapts the ``scikit-learn`` vectorizer to the Podium input data.
 
 .. code-block:: python
-  
+
   >>> from podium.vectorizers import TfIdfVectorizer
   >>> tfidf_vectorizer = TFIdfVectorizer()
   >>> tfidf_vectorizer.fit(train, field=train.field('text'))

From 0c04847b8f06e56f41bad4b016719938303e1514 Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Thu, 1 Apr 2021 17:01:17 +0200
Subject: [PATCH 12/26] Move examples back to examples folder

---
 docs/source/{ => examples}/pytorch_rnn_example.rst | 0
 docs/source/{ => examples}/tfidf_example.rst       | 0
 2 files changed, 0 insertions(+), 0 deletions(-)
 rename docs/source/{ => examples}/pytorch_rnn_example.rst (100%)
 rename docs/source/{ => examples}/tfidf_example.rst (100%)

diff --git a/docs/source/pytorch_rnn_example.rst b/docs/source/examples/pytorch_rnn_example.rst
similarity index 100%
rename from docs/source/pytorch_rnn_example.rst
rename to docs/source/examples/pytorch_rnn_example.rst
diff --git a/docs/source/tfidf_example.rst b/docs/source/examples/tfidf_example.rst
similarity index 100%
rename from docs/source/tfidf_example.rst
rename to docs/source/examples/tfidf_example.rst

From 9f77e5a3c4ad6e2a01be5373965908bdb8178179 Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Thu, 1 Apr 2021 17:08:17 +0200
Subject: [PATCH 13/26] Add examples directory to notebooks

---
 .../examples/pytorch_rnn_example.ipynb        | 500 ++++++++++++++++++
 .../scripts/convert_doc_to_notebooks.py       |   4 +-
 2 files changed, 502 insertions(+), 2 deletions(-)
 create mode 100644 docs/source/notebooks/examples/pytorch_rnn_example.ipynb

diff --git a/docs/source/notebooks/examples/pytorch_rnn_example.ipynb b/docs/source/notebooks/examples/pytorch_rnn_example.ipynb
new file mode 100644
index 00000000..a29d6256
--- /dev/null
+++ b/docs/source/notebooks/examples/pytorch_rnn_example.ipynb
@@ -0,0 +1,500 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Podium installation\n",
+    "! pip install podium-nlp\n",
+    "# To install from source instead of the last release, comment the command above and uncomment the following one.\n",
+    "# ! pip install git+https://github.com/takelab/podium"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Pytorch RNN classifier"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "In this example, we will cover a simple RNN-based classifier model implemented in Pytorch. We will use the IMDB dataset loaded from 🤗/datasets, preprocess it with Fields and train the model briefly.\n",
+    "While having a GPU is not necessary it is recommended as otherwise training the model, even for a single epoch, will take a while."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Loading a dataset from 🤗/datasets"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "As we have covered in [Loading 🤗 datasets](http://takelab.fer.hr/podium/walkthrough.html#hf-loading), we have implemented wrappers around 🤗 dataset classes to enable working with the plethora of datasets implemented therein. We will now briefly go through (1) loading a dataset from 🤗/datasets and (2) wrapping it in Podium classes."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "DatasetDict({\n",
+       "    train: Dataset({\n",
+       "        features: ['text', 'label'],\n",
+       "        num_rows: 25000\n",
+       "    })\n",
+       "    test: Dataset({\n",
+       "        features: ['text', 'label'],\n",
+       "        num_rows: 25000\n",
+       "    })\n",
+       "    unsupervised: Dataset({\n",
+       "        features: ['text', 'label'],\n",
+       "        num_rows: 50000\n",
+       "    })\n",
+       "})\n",
+       "{'label': ClassLabel(num_classes=2, names=['neg', 'pos'], names_file=None, id=None),\n",
+       " 'text': Value(dtype='string', id=None)}"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "import datasets\n",
+    "imdb = datasets.load_dataset('imdb')\n",
+    "print(imdb)\n",
+    "from pprint import pprint\n",
+    "pprint(imdb['train'].features)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "By calling `load_dataset` the dataset was downloaded and cached on disk through the `datasets` library. The dataset has two splits we are interested in (`train` and `test`).\n",
+    "The main thing we need to pay attention to are the `features` of the dataset, in this case `text` and `label`. These features, or data columns, need to be mapped to (and processed by) Podium Fields.\n",
+    "\n",
+    "For convenience, we have implemented automatic `Field` type inference from 🤗 dataset features -- however it is far from perfect as we have to make many assumptions on the way. We will now wrap the IMDB dataset in Podium and show the automatically inferred Fields."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "Field({\n",
+       "    name: 'text',\n",
+       "    keep_raw: False,\n",
+       "    is_target: False,\n",
+       "    vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 280619})\n",
+       "})\n",
+       "LabelField({\n",
+       "    name: 'label',\n",
+       "    keep_raw: False,\n",
+       "    is_target: True\n",
+       "})"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from podium.datasets.hf import HFDatasetConverter as HF\n",
+    "splits = HF.from_dataset_dict(imdb)\n",
+    "imdb_train, imdb_test = splits['train'], splits['test']\n",
+    "imdb_train.finalize_fields() # Construct the vocab\n",
+    "print(*imdb_train.fields, sep=\"\\n\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Both of the Fields were constructed well, but there are a couple of drawbacks for this concrete dataset. Firstly, the size of the vocabulary is very large (`280619`) -- we would like to trim this down to a reasonable number as we won't be using subword tokenization in this example. "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "Example({\n",
+       "    text: (None, ['Bromwell', 'High', 'is', 'a', 'cartoon', 'comedy.', 'It', 'ran', 'at', 'the', 'same', 'time', 'as', 'some', 'other', 'programs', 'about', 'school', 'life,', 'such', 'as', '\"Teachers\".', 'My', '35', 'years', 'in', 'the', 'teaching', 'profession', 'lead', 'me', 'to', 'believe', 'that', 'Bromwell', \"High's\", 'satire', 'is', 'much', 'closer', 'to', 'reality', 'than', 'is', '\"Teachers\".', 'The', 'scramble', 'to', 'survive', 'financially,', 'the', 'insightful', 'students', 'who', 'can', 'see', 'right', 'through', 'their', 'pathetic', \"teachers'\", 'pomp,', 'the', 'pettiness', 'of', 'the', 'whole', 'situation,', 'all', 'remind', 'me', 'of', 'the', 'schools', 'I', 'knew', 'and', 'their', 'students.', 'When', 'I', 'saw', 'the', 'episode', 'in', 'which', 'a', 'student', 'repeatedly', 'tried', 'to', 'burn', 'down', 'the', 'school,', 'I', 'immediately', 'recalled', '.........', 'at', '..........', 'High.', 'A', 'classic', 'line:', 'INSPECTOR:', \"I'm\", 'here', 'to', 'sack', 'one', 'of', 'your', 'teachers.', 'STUDENT:', 'Welcome', 'to', 'Bromwell', 'High.', 'I', 'expect', 'that', 'many', 'adults', 'of', 'my', 'age', 'think', 'that', 'Bromwell', 'High', 'is', 'far', 'fetched.', 'What', 'a', 'pity', 'that', 'it', \"isn't!\"]),\n",
+       "    label: (None, 1)\n",
+       "})"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "print(imdb_train[0])"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "When inspecting a concrete instance, there are a few more things to note. Firstly, IMDB instances can be quite long (on average around 200 tokens per instance), secondly, the text wasn't tokenized properly near sentence boundaries (due to using the default `str.split` tokenizer) and lastly, the text has varying casing.\n",
+    "We will instead define our own Fields for the corresponding features, add posttokenization hooks which will transform the data, and use those Fields to replace the automatically inferred ones:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "HFDatasetConverter({\n",
+       "    dataset_name: imdb,\n",
+       "    size: 25000,\n",
+       "    fields: [\n",
+       "            Field({\n",
+       "                name: 'text',\n",
+       "                keep_raw: False,\n",
+       "                is_target: False,\n",
+       "                vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 10000})\n",
+       "            }),\n",
+       "            LabelField({\n",
+       "                name: 'label',\n",
+       "                keep_raw: False,\n",
+       "                is_target: True\n",
+       "            })\n",
+       "    \n",
+       "    ]\n",
+       "})\n",
+       "Example({\n",
+       "    text: (None, ['bromwell', 'high', 'is', 'a', 'cartoon', 'comedy', '.', 'it', 'ran', 'at', 'the', 'same', 'time', 'as', 'some', 'other', 'programs', 'about', 'school', 'life', ',', 'such', 'as', '\"', 'teachers', '\"', '.', 'my', '35', 'years', 'in', 'the', 'teaching', 'profession', 'lead', 'me', 'to', 'believe', 'that', 'bromwell', 'high', \"'s\", 'satire', 'is', 'much', 'closer', 'to', 'reality', 'than', 'is', '\"', 'teachers', '\"', '.', 'the', 'scramble', 'to', 'survive', 'financially', ',', 'the', 'insightful', 'students', 'who', 'can', 'see', 'right', 'through', 'their', 'pathetic', 'teachers', \"'\", 'pomp', ',', 'the', 'pettiness', 'of', 'the', 'whole', 'situation', ',', 'all', 'remind', 'me', 'of', 'the', 'schools', 'i', 'knew', 'and', 'their', 'students', '.', 'when', 'i', 'saw', 'the', 'episode', 'in', 'which', 'a', 'student', 'repeatedly', 'tried', 'to', 'burn', 'down', 'the', 'school', ',', 'i', 'immediately', 'recalled', '.........', 'at', '..........', 'high', '.', 'a', 'classic', 'line', ':', 'inspector', ':', 'i', \"'m\", 'here', 'to', 'sack', 'one', 'of', 'your', 'teachers', '.', 'student', ':', 'welcome', 'to', 'bromwell', 'high', '.', 'i', 'expect', 'that', 'many', 'adults', 'of', 'my', 'age', 'think', 'that', 'bromwell', 'high', 'is', 'far', 'fetched', '.', 'what', 'a', 'pity', 'that', 'it', 'is', \"n't\", '!']),\n",
+       "    label: (None, 1)\n",
+       "})"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from podium import Field, LabelField, Vocab\n",
+    "\n",
+    "# Lowercasing as a post-tokenization hook\n",
+    "def lowercase(raw, tokenized):\n",
+    "  return raw, [token.lower() for token in tokenized]\n",
+    "\n",
+    "# Truncating as a post-tokenization hook\n",
+    "def truncate(raw, tokenized, max_length=200):\n",
+    "    return raw, tokenized[:max_length]\n",
+    "\n",
+    "vocab = Vocab(max_size=10000)\n",
+    "text = Field(name=\"text\", \n",
+    "             numericalizer=vocab,\n",
+    "             include_lengths=True,\n",
+    "             tokenizer=\"spacy-en_core_web_sm\",\n",
+    "             posttokenize_hooks=[truncate, lowercase])\n",
+    "\n",
+    "# The labels are already mapped to indices in /datasets so we will\n",
+    "# pass them through\n",
+    "label = LabelField(name=\"label\", numericalizer=lambda x: x)\n",
+    "fields = {\n",
+    "    'text': text,\n",
+    "    'label': label\n",
+    "}\n",
+    "\n",
+    "# Use the given Fields to load the dataset again\n",
+    "splits = HF.from_dataset_dict(imdb, fields=fields)\n",
+    "imdb_train, imdb_test = splits['train'], splits['test']\n",
+    "imdb_train.finalize_fields()\n",
+    "print(imdb_train)\n",
+    "print(imdb_train[0])"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Here, we can see the effect of our hooks and using the spacy tokenizer. Now our dataset will be a bit cleaner to work with. Some data cleaning would still be desired, such as removing tokens which only contain punctuation, but we leave this exercise to the reader :)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Loading pretrained embeddings"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "In most use-cases, we want to use pre-trained word embeddings along with our neural model. With Podium, this process is very simple. If your field uses a vocabulary, it has already built an inventory of tokens for your dataset.\n",
+    "\n",
+    "For example, we will use the [GloVe](https://nlp.stanford.edu/projects/glove/) vectors. You can read more about loading pretrained vectors in [Loading pretrained word vectors](http://takelab.fer.hr/podium/walkthrough.html#pretrained), but the procedure to load these vectors has two steps: (1) initialize the vector class, which sets all the required paths and (2) obtain the vectors for a pre-defined list of words by calling `load_vocab`."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "For vocabulary of size: 10000 loaded embedding matrix of shape: (10000, 300)\n",
+       "Vector for sport: [ 0.34566    0.15934    0.48444   -0.13693    0.18737    0.2678"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from podium.vectorizers import GloVe\n",
+    "vocab = fields['text'].vocab\n",
+    "glove = GloVe()\n",
+    "embeddings = glove.load_vocab(vocab)\n",
+    "print(f\"For vocabulary of size: {len(vocab)} loaded embedding matrix of shape: {embeddings.shape}\")\n",
+    "# We can obtain vectors for a single word (given the word is loaded) like this:\n",
+    "word = \"sport\"\n",
+    "print(f\"Vector for {word}: {glove.token_to_vector(word)}\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    " -0.39159    0.4931    -0.76111   -1.4586     0.41475    0.55837\n",
+    "  ...\n",
+    "  0.13802    0.36619    0.19734    0.35701   -0.42228   -0.25242\n",
+    " -0.050651  -0.041129   0.15092    0.22084    0.52252   -0.27224  ]"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Defining a simple neural model in Pytorch"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "In this section, we will implement a very simple neural classification model -- a 2-layer BiGRU with a single hidden layer classifier on top of its last hidden state. Many improvements to the model can be made, but this is not our current focus."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import torch\n",
+    "import torch.nn as nn\n",
+    "import torch.nn.functional as F\n",
+    "\n",
+    "from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence\n",
+    "\n",
+    "class RNNClassifier(nn.Module):\n",
+    "    def __init__(self, embedding, embed_dim=300, hidden_dim=300, num_labels=2):\n",
+    "      super(NLIModel, self).__init__()\n",
+    "      self.embedding = embedding\n",
+    "      self.encoder = nn.GRU(\n",
+    "            input_size=embed_dim,\n",
+    "            hidden_size=hidden_dim,\n",
+    "            num_layers=2,\n",
+    "            bidirectional=True,\n",
+    "            dropout=0.3\n",
+    "      )\n",
+    "      self.decoder = nn.Sequential(\n",
+    "            nn.Linear(2*hidden_dim, hidden_dim),\n",
+    "            nn.Tanh(),\n",
+    "            nn.Linear(hidden_dim, num_labels)\n",
+    "      )\n",
+    "\n",
+    "    def forward(self, x, lengths):\n",
+    "        e = self.embedding(x)\n",
+    "        h_pack = pack_padded_sequence(e, \n",
+    "                                      lengths,\n",
+    "                                      enforce_sorted=False,\n",
+    "                                      batch_first=True)\n",
+    "\n",
+    "        _, h = self.encoder(h_pack) # [2L x B x H]\n",
+    "\n",
+    "        # Concat last state of left and right directions\n",
+    "        h = torch.cat([h[-1], h[-2]], dim=-1) # [B x 2H]\n",
+    "        return self.decoder(h)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "There. We will now define the prerequisites for pytorch model training, where we will use a GPU for speed, however running the model for one epoch will is possible albeit time-consuing even without a GPU."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "embed_dim = 300\n",
+    "padding_index = text.vocab.get_padding_index()\n",
+    "embedding_matrix = nn.Embedding(len(text.vocab), embed_dim,\n",
+    "                                padding_idx=padding_index)\n",
+    "# Copy the pretrained GloVe word embeddings\n",
+    "embedding_matrix.weight.data.copy_(torch.from_numpy(embeddings))\n",
+    "\n",
+    "device = torch.device(\"cuda:0\")\n",
+    "model = RNNClassifier(embedding_matrix)\n",
+    "model = model.to(device)\n",
+    "criterion = nn.CrossEntropyLoss()\n",
+    "optimizer = torch.optim.Adam(model.parameters())"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Now that we have the model setup code ready, we will first define helper method to measure accuracy of our model after each epoch:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import numpy as np\n",
+    "def update_stats(accuracy, confusion_matrix, logits, y):\n",
+    "    _, max_ind = torch.max(logits, 1)\n",
+    "    equal = torch.eq(max_ind, y)\n",
+    "    correct = int(torch.sum(equal))\n",
+    "\n",
+    "    for j, i in zip(max_ind, y):\n",
+    "        confusion_matrix[int(i),int(j)]+=1\n",
+    "    return accuracy + correct, confusion_matrix"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "and now the training loop for the model:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import tqdm\n",
+    "def train(model, data, optimizer, criterion, num_labels):\n",
+    "    model.train()\n",
+    "    accuracy, confusion_matrix = 0, np.zeros((num_labels, num_labels), dtype=int)\n",
+    "    for batch_num, batch in tqdm.tqdm(enumerate(data), total=len(data)):\n",
+    "        x, lens = batch.text\n",
+    "        y = batch.label\n",
+    "        logits = model(x, lens)\n",
+    "        accuracy, confusion_matrix = update_stats(accuracy, confusion_matrix, logits, y)\n",
+    "        loss = criterion(logits, y.squeeze())\n",
+    "        loss.backward()\n",
+    "        optimizer.step()\n",
+    "    print(\"[Accuracy]: {}/{} : {:.3f}%\".format(\n",
+    "          accuracy, len(data)*data.batch_size, accuracy / len(data) / data.batch_size * 100))\n",
+    "    return accuracy, confusion_matrix"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "and now, we are done with our model code. Let's turn back to Podium and see how we can set up batching for our training loop to start ticking."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Minibatching data in Podium"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "We have covered batching data in [Minibatching data](http://takelab.fer.hr/podium/quickstart.html#minibatching) and advanced batching through bucketing in [Bucketing instances when iterating](http://takelab.fer.hr/podium/advanced.html#bucketing). We will use the plain Iterator and leave bucketing for you to change to see how much the model speeds up when minimizing padding. One change we would like to do when iterating over data is to obtain the data matrices as torch tensors on the `device` we defined previously. We will now demonstrate how to do this by setting the `matrix_class` argument of the `Iterator`:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "[Accuracy]: 20050/25024 : 80.123%\n",
+       "[Accuracy]: 22683/25024 : 90.645%\n",
+       "[Accuracy]: 23709/25024 : 94.745%\n",
+       "[Accuracy]: 24323/25024 : 97.199%\n",
+       "[Accuracy]: 24595/25024 : 98.286%"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from podium import Iterator\n",
+    "# Closure for converting data to given device\n",
+    "def gpu_tensor(data):\n",
+    "    return torch.tensor(data).to(device)\n",
+    "# Initialize our iterator\n",
+    "train_iter = Iterator(imdb_train, batch_size=32, matrix_class=gpu_tensor)\n",
+    "\n",
+    "epochs = 5\n",
+    "for epoch in range(epochs):\n",
+    "    train(model, train_iter, optimizer, criterion, num_labels=2)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "And we are done! In our case, the model takes about one minute per epoch on a GPU, but this can be sped up by using bucketing, which we recommend you try out yourself."
+   ]
+  }
+ ],
+ "metadata": {},
+ "nbformat": 4,
+ "nbformat_minor": 4
+}
diff --git a/docs/source/scripts/convert_doc_to_notebooks.py b/docs/source/scripts/convert_doc_to_notebooks.py
index 897a823f..abbf914f 100644
--- a/docs/source/scripts/convert_doc_to_notebooks.py
+++ b/docs/source/scripts/convert_doc_to_notebooks.py
@@ -20,8 +20,8 @@
     "advanced.rst",
     "preprocessing.rst",
     "walkthrough.rst",
-    "pytorch_rnn_example.rst",
-    "tfidf_example.rst",
+    "examples/pytorch_rnn_example.rst",
+    "examples/tfidf_example.rst",
 ]
 
 _re_label = re.compile(r"\.\.\s+_([^:]*):")

From 73ab6020f8ba5cc88018500f22d3af44cc0ad5c5 Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Thu, 1 Apr 2021 17:09:37 +0200
Subject: [PATCH 14/26] Remove debug print

---
 docs/source/scripts/convert_doc_to_notebooks.py | 2 --
 1 file changed, 2 deletions(-)

diff --git a/docs/source/scripts/convert_doc_to_notebooks.py b/docs/source/scripts/convert_doc_to_notebooks.py
index abbf914f..46288931 100644
--- a/docs/source/scripts/convert_doc_to_notebooks.py
+++ b/docs/source/scripts/convert_doc_to_notebooks.py
@@ -558,8 +558,6 @@ def convert_all_tutorials(path_to_docs=None, path_to_dest=None):
     for file in TUTORIAL_FILES:
         notebook_name = os.path.splitext(file)[0] + ".ipynb"
         doc_file = os.path.join(path_to_docs, file)
-        print(doc_file)
-        print(os.path.exists(doc_file))
         notebook_file = os.path.join(path_to_dest, notebook_name)
         convert_rst_file_to_notebook(doc_file, notebook_file, origin_folder=path_to_docs, dest_folder=path_to_dest, additional_deps=ADDITIONAL_DEPS.get(file))
 

From 7b78c81069bd69d9b5afddde76610d575bbd5347 Mon Sep 17 00:00:00 2001
From: mariosasko <mariosasko777@gmail.com>
Date: Thu, 1 Apr 2021 18:47:31 +0200
Subject: [PATCH 15/26] Fix tfidf example, improve notebooks

---
 docs/source/examples/pytorch_rnn_example.rst  |  35 +-
 docs/source/examples/tfidf_example.rst        |   2 +-
 docs/source/index.rst                         |   4 +-
 docs/source/notebooks/advanced.ipynb          |  10 +-
 .../notebooks/pytorch_rnn_example.ipynb       |  40 ++-
 docs/source/notebooks/quickstart.ipynb        |   2 +-
 docs/source/notebooks/tfidf_example.ipynb     | 330 ++++++++++++++++++
 docs/source/notebooks/walkthrough.ipynb       |  16 +-
 .../scripts/convert_doc_to_notebooks.py       |  15 +-
 9 files changed, 394 insertions(+), 60 deletions(-)
 create mode 100644 docs/source/notebooks/tfidf_example.ipynb

diff --git a/docs/source/examples/pytorch_rnn_example.rst b/docs/source/examples/pytorch_rnn_example.rst
index 991ee4f0..0700e425 100644
--- a/docs/source/examples/pytorch_rnn_example.rst
+++ b/docs/source/examples/pytorch_rnn_example.rst
@@ -11,8 +11,8 @@ As we have covered in :ref:`hf-loading`, we have implemented wrappers around 
 
 .. code-block:: python
 
-  >>> import datasets
-  >>> imdb = datasets.load_dataset('imdb')
+  >>> from datasets import load_dataset
+  >>> imdb = load_dataset('imdb')
   >>> print(imdb)
   DatasetDict({
       train: Dataset({
@@ -106,18 +106,17 @@ We will instead define our own Fields for the corresponding features, add postto
       dataset_name: imdb,
       size: 25000,
       fields: [
-              Field({
-                  name: 'text',
-                  keep_raw: False,
-                  is_target: False,
-                  vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 10000})
-              }),
-              LabelField({
-                  name: 'label',
-                  keep_raw: False,
-                  is_target: True
-              })
-      
+          Field({
+              name: 'text',
+              keep_raw: False,
+              is_target: False,
+              vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 10000})
+          }),
+          LabelField({
+              name: 'label',
+              keep_raw: False,
+              is_target: True
+          })
       ]
   })
   >>> print(imdb_train[0])
@@ -194,7 +193,7 @@ In this section, we will implement a very simple neural classification model --
   ...         h = torch.cat([h[-1], h[-2]], dim=-1) # [B x 2H]
   ...         return self.decoder(h)
 
-There. We will now define the prerequisites for pytorch model training, where we will use a GPU for speed, however running the model for one epoch will is possible albeit time-consuing even without a GPU.
+We will now define the prerequisites for pytorch model training.
 
 .. code-block:: python
 
@@ -205,7 +204,7 @@ There. We will now define the prerequisites for pytorch model training, where we
   >>> # Copy the pretrained GloVe word embeddings
   >>> embedding_matrix.weight.data.copy_(torch.from_numpy(embeddings))
   >>>
-  >>> device = torch.device("cuda:0")
+  >>> device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
   >>> model = RNNClassifier(embedding_matrix)
   >>> model = model.to(device)
   >>> criterion = nn.CrossEntropyLoss()
@@ -256,10 +255,10 @@ We have covered batching data in :ref:`minibatching` and advanced batching throu
 
   >>> from podium import Iterator
   >>> # Closure for converting data to given device
-  >>> def gpu_tensor(data):
+  >>> def device_tensor(data):
   ...     return torch.tensor(data).to(device)
   >>> # Initialize our iterator
-  >>> train_iter = Iterator(imdb_train, batch_size=32, matrix_class=gpu_tensor)
+  >>> train_iter = Iterator(imdb_train, batch_size=32, matrix_class=device_tensor)
   >>>
   >>> epochs = 5
   >>> for epoch in range(epochs):
diff --git a/docs/source/examples/tfidf_example.rst b/docs/source/examples/tfidf_example.rst
index 770a8440..2ba93864 100644
--- a/docs/source/examples/tfidf_example.rst
+++ b/docs/source/examples/tfidf_example.rst
@@ -73,7 +73,7 @@ We will do two things: (1) implement a pre-tokenization hook to lowercase our da
 And then implement flexible ngram extraction where the ``n`` is an interval using ``nltk``\s ``ngrams`` function:
 
 .. code-block:: python
-  
+
   >>> from ntlk import ngrams
   >>> class NGramHook:
   ...   # Transforms a sequence of unigrams into a sequence of
diff --git a/docs/source/index.rst b/docs/source/index.rst
index 22b7b571..36119ebf 100644
--- a/docs/source/index.rst
+++ b/docs/source/index.rst
@@ -15,7 +15,7 @@ Contents
 ---------------------------------
 The documentation is organized in five parts:
 
-- **Quickstart**: an quick preview of the library,
+- **Quickstart**: a quick preview of the library,
 - **Walkthrough**: a description of how the basics work,
 - **In-depth overview**: advanced usage options,
 - **Examples**: full stand-alone examples of NLP models using Podium,
@@ -39,7 +39,7 @@ The documentation is organized in five parts:
 
 .. toctree::
     :maxdepth: 2
-    :caption: Full examples
+    :caption: Examples
 
     examples/tfidf_example.rst
     examples/pytorch_rnn_example.rst
diff --git a/docs/source/notebooks/advanced.ipynb b/docs/source/notebooks/advanced.ipynb
index 639d6af0..d582d1a6 100644
--- a/docs/source/notebooks/advanced.ipynb
+++ b/docs/source/notebooks/advanced.ipynb
@@ -134,7 +134,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "<a id='interact_fields'></a>"
+    "<a name='interact_fields' id='interact_fields'></a>"
    ]
   },
   {
@@ -337,7 +337,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "<a id='specials'></a>"
+    "<a name='specials' id='specials'></a>"
    ]
   },
   {
@@ -477,7 +477,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "<a id='custom_numericalization'></a>"
+    "<a name='custom_numericalization' id='custom_numericalization'></a>"
    ]
   },
   {
@@ -798,7 +798,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "<a id='dataset_concat'></a>"
+    "<a name='dataset_concat' id='dataset_concat'></a>"
    ]
   },
   {
@@ -867,7 +867,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "<a id='bucketing'></a>"
+    "<a name='bucketing' id='bucketing'></a>"
    ]
   },
   {
diff --git a/docs/source/notebooks/pytorch_rnn_example.ipynb b/docs/source/notebooks/pytorch_rnn_example.ipynb
index a29d6256..076f953f 100644
--- a/docs/source/notebooks/pytorch_rnn_example.ipynb
+++ b/docs/source/notebooks/pytorch_rnn_example.ipynb
@@ -9,7 +9,10 @@
     "# Podium installation\n",
     "! pip install podium-nlp\n",
     "# To install from source instead of the last release, comment the command above and uncomment the following one.\n",
-    "# ! pip install git+https://github.com/takelab/podium"
+    "# ! pip install git+https://github.com/takelab/podium\n",
+    "\n",
+    "# Additional dependencies required to run this notebook:\n",
+    "! pip install torch"
    ]
   },
   {
@@ -73,8 +76,8 @@
     }
    ],
    "source": [
-    "import datasets\n",
-    "imdb = datasets.load_dataset('imdb')\n",
+    "from datasets import load_dataset\n",
+    "imdb = load_dataset('imdb')\n",
     "print(imdb)\n",
     "from pprint import pprint\n",
     "pprint(imdb['train'].features)"
@@ -174,18 +177,17 @@
        "    dataset_name: imdb,\n",
        "    size: 25000,\n",
        "    fields: [\n",
-       "            Field({\n",
-       "                name: 'text',\n",
-       "                keep_raw: False,\n",
-       "                is_target: False,\n",
-       "                vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 10000})\n",
-       "            }),\n",
-       "            LabelField({\n",
-       "                name: 'label',\n",
-       "                keep_raw: False,\n",
-       "                is_target: True\n",
-       "            })\n",
-       "    \n",
+       "        Field({\n",
+       "            name: 'text',\n",
+       "            keep_raw: False,\n",
+       "            is_target: False,\n",
+       "            vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 10000})\n",
+       "        }),\n",
+       "        LabelField({\n",
+       "            name: 'label',\n",
+       "            keep_raw: False,\n",
+       "            is_target: True\n",
+       "        })\n",
        "    ]\n",
        "})\n",
        "Example({\n",
@@ -355,7 +357,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "There. We will now define the prerequisites for pytorch model training, where we will use a GPU for speed, however running the model for one epoch will is possible albeit time-consuing even without a GPU."
+    "We will now define the prerequisites for pytorch model training."
    ]
   },
   {
@@ -371,7 +373,7 @@
     "# Copy the pretrained GloVe word embeddings\n",
     "embedding_matrix.weight.data.copy_(torch.from_numpy(embeddings))\n",
     "\n",
-    "device = torch.device(\"cuda:0\")\n",
+    "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
     "model = RNNClassifier(embedding_matrix)\n",
     "model = model.to(device)\n",
     "criterion = nn.CrossEntropyLoss()\n",
@@ -476,10 +478,10 @@
    "source": [
     "from podium import Iterator\n",
     "# Closure for converting data to given device\n",
-    "def gpu_tensor(data):\n",
+    "def device_tensor(data):\n",
     "    return torch.tensor(data).to(device)\n",
     "# Initialize our iterator\n",
-    "train_iter = Iterator(imdb_train, batch_size=32, matrix_class=gpu_tensor)\n",
+    "train_iter = Iterator(imdb_train, batch_size=32, matrix_class=device_tensor)\n",
     "\n",
     "epochs = 5\n",
     "for epoch in range(epochs):\n",
diff --git a/docs/source/notebooks/quickstart.ipynb b/docs/source/notebooks/quickstart.ipynb
index b42e799d..63d8382c 100644
--- a/docs/source/notebooks/quickstart.ipynb
+++ b/docs/source/notebooks/quickstart.ipynb
@@ -436,7 +436,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "<a id='minibatching'></a>"
+    "<a name='minibatching' id='minibatching'></a>"
    ]
   },
   {
diff --git a/docs/source/notebooks/tfidf_example.ipynb b/docs/source/notebooks/tfidf_example.ipynb
new file mode 100644
index 00000000..bb10f8e2
--- /dev/null
+++ b/docs/source/notebooks/tfidf_example.ipynb
@@ -0,0 +1,330 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Podium installation\n",
+    "! pip install podium-nlp\n",
+    "# To install from source instead of the last release, comment the command above and uncomment the following one.\n",
+    "# ! pip install git+https://github.com/takelab/podium"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# TFIDF + scikit-learn SVM"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "In this example, we will cover a once popular family of models -- support vector machines (SVMs) with TF-IDF representations. As a simple example, we will analyse binary classification on the Stanford sentiment treebank (SST) dataset.\n",
+    "\n",
+    "First, we will implement a minimalistic example without much additional preprocessing. Since we're using TFIDF representation of our dataset, it is smart to limit the size of the vocabulary as each word needs to be present in the instance TFIDF representation. Let's load the SST dataset and convert it into a single batch:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from podium import Vocab, Field, LabelField\n",
+    "from podium.datasets import SST\n",
+    "vocab = Vocab(max_size=5000, specials=())\n",
+    "text = Field(name='text', numericalizer=vocab, disable_batch_matrix=True)\n",
+    "label = LabelField(name='label')\n",
+    "fields = {'text': text, 'label': label}\n",
+    "train, dev, test = SST.get_dataset_splits(fields=fields)\n",
+    "train.finalize_fields()\n",
+    "x, y = train.batch(add_padding=True)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "We have now loaded our dataset, finalized its Fields and obtained it as a batch of input and target data. What we need to do next is define the TF-IDF vectorization for each instance in the dataset. This is done by using our `TfIdfVectorizer`, which adapts the `scikit-learn` vectorizer to the Podium input data."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "<class 'scipy.sparse.csr.csr_matrix'> (6920, 5000)"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from podium.vectorizers import TfIdfVectorizer\n",
+    "tfidf_vectorizer = TFIdfVectorizer()\n",
+    "tfidf_vectorizer.fit(train, field=train.field('text'))\n",
+    "tfidf_x = tfidf_vectorizer.transform(x.text)\n",
+    "print(type(tfidf_batch), tfidf_batch.shape)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "We have transformed the train dataset to a sparse matrix containing TF-IDF values for each word in the vocabulary in each instance. What is left to do now is to train our classification model:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "Accuracy on train set: 0.9597"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from sklearn.svm import LinearSVC\n",
+    "from sklearn.metrics import accuracy_score\n",
+    "# Train the SVM on the training set\n",
+    "svm = LinearSVC()\n",
+    "svm.fit(tfidf_batch, y.label.ravel())\n",
+    "# Obtain accuracy on the train set\n",
+    "y_hat = svm.predict(tfidf_batch)\n",
+    "acc = accuracy_score(y_hat, y.label.ravel())\n",
+    "print(f\"Accuracy on the train set: {acc:.4f}\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "And for a more accurate performance evaluation of our model we turn to the test set:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "Accuracy on the test set: 0.7946"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "test_x, test_y = test.batch()\n",
+    "tfidf_test_batch = tfidf_vectorizer.transform(test_x.text)\n",
+    "y_test_hat = svm.predict(tfidf_test_batch)\n",
+    "acc = accuracy_score(y_test_hat, test_y.label.ravel())\n",
+    "print(f\"Accuracy on the test set: {acc:.4f}\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Our basic unigram TF-IDF linear SVM performs pretty well on the SST dataset, reaching accuracy of almost `0.8`. While this example encapsulates the basics of using Podium with `scikit-learn`s `SVM`, we will delve a bit deeper and consider some additional preprocessing."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Using ngram features"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "We have only considered basic unigram features for our model, and this is somewhat prohibitive. Apart from that, we have not implemented any preprocessing for our dataset, and our dataset is cased -- which might be detrimental for the performance of our model since we agressively trim the vocabulary size.\n",
+    "\n",
+    "We will do two things: (1) implement a pre-tokenization hook to lowercase our data, which in our case is fine as we are using the case-insensitive `str.split` as a tokenizer, and (2) implement ngram extraction as a post-tokenization hook. For a more detailed overview of hooks and how to use them, check [Customizing the preprocessing pipeline with Fields](http://takelab.fer.hr/podium/walkthrough.html#fields). We will first implement our lowercase hook:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "def lowercase(raw):\n",
+    "  \"\"\"Lowercases the input string\"\"\"\n",
+    "  return raw.lower()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "And then implement flexible ngram extraction where the `n` is an interval using `nltk`s `ngrams` function:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from ntlk import ngrams\n",
+    "class NGramHook:\n",
+    "  # Transforms a sequence of unigrams into a sequence of\n",
+    "  # [min_n, max_n]-grams\n",
+    "  def __init__(self, min_n, max_n):\n",
+    "    self.min_n = min_n\n",
+    "    self.max_n = max_n\n",
+    "  def __call__(self, raw, tokenized):\n",
+    "    tokenized_ngrams = []\n",
+    "    for n in range(self.min_n, self.max_n+1):\n",
+    "       tokenized_ngrams.extend(ngrams(tokenized, n))\n",
+    "    return raw, tokenized_ngrams"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "We will now incorporate these two hooks into our text input Field:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "                 disable_batch_matrix=True,\n",
+       "                 pretokenization_hooks=[lowercase],\n",
+       "                 posttokenization_hooks=[ngram_hook]\n",
+       "                 )\n",
+       "[('at',), ('from',), ('one',), ('have',), ('I',), ('like',), ('his',), ('in', 'the'), ('all',), (\"'\",)]"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "# Use [1-3]grams, inclusive\n",
+    "ngram_hook = NGramHook(1,3)\n",
+    "vocab = Vocab(max_size=5000, specials=())\n",
+    "text = Field(name='text', numericalizer=vocab, \n",
+    "label = LabelField(name='label')\n",
+    "fields = {'text': text, 'label': label}\n",
+    "train, dev, test = SST.get_dataset_splits(fields=fields)\n",
+    "train.finalize_fields()\n",
+    "print(text.vocab.itos[40:50])"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "We can see that our new Vocab now contains tuples as its tokens -- as long as an item in a sequence is hashable, we can represent it as part of a Vocab! We can see that one 2-gram `('in', 'the')` has made its way into the 50 most frequent tokens.\n",
+    "\n",
+    "As before, we need to train the TFIDF vectorizer and apply it to our data (which now includes 1-, 2- and 3-grams):"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "<class 'scipy.sparse.csr.csr_matrix'> (6920, 5000)"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "dataset_batch = train.batch(add_padding=True)\n",
+    "tfidf_vectorizer = TfIdfVectorizer()\n",
+    "tfidf_vectorizer.fit(train, field=train.field('text'))\n",
+    "tfidf_batch = tfidf_vectorizer.transform(dataset_batch.text)\n",
+    "print(type(tfidf_batch), tfidf_batch.shape)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "We can now train our SVM classification model and evaluate it on the train and test set:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "Accuracy on the train set: 0.9575\n",
+       "Accuracy on the test set: 0.7743"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "svm = LinearSVC()\n",
+    "text, label = dataset_batch\n",
+    "svm.fit(tfidf_batch, label.ravel())\n",
+    "# Compute accuracy on the train set\n",
+    "y_hat = svm.predict(tfidf_batch)\n",
+    "acc = accuracy_score(y_hat, label.ravel())\n",
+    "print(f\"Accuracy on the train set: {acc:.4f}\")\n",
+    "\n",
+    "# Compute accuracy on the test set\n",
+    "test_text, test_label = test.batch(add_padding=True)\n",
+    "tfidf_test_batch = tfidf_vectorizer.transform(test_text)\n",
+    "y_test_hat = svm.predict(tfidf_test_batch)\n",
+    "acc = accuracy_score(y_test_hat, test_label.ravel())\n",
+    "print(f\"Accuracy on the test set: {acc:.4f}\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Sadly, our new model didn't perform better than our initial one on the train set, but there are many avenues we can try further, such as tuning the hyperparameters of the LinearSVC model on the development set or filtering out stop words and punctuation. We encourage you to open this example in Colab and try some things yourself!"
+   ]
+  }
+ ],
+ "metadata": {},
+ "nbformat": 4,
+ "nbformat_minor": 4
+}
diff --git a/docs/source/notebooks/walkthrough.ipynb b/docs/source/notebooks/walkthrough.ipynb
index b8a1dc3a..e6134804 100644
--- a/docs/source/notebooks/walkthrough.ipynb
+++ b/docs/source/notebooks/walkthrough.ipynb
@@ -49,7 +49,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "<a id='builtin-loading'></a>"
+    "<a name='builtin-loading' id='builtin-loading'></a>"
    ]
   },
   {
@@ -121,7 +121,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "<a id='hf-loading'></a>"
+    "<a name='hf-loading' id='hf-loading'></a>"
    ]
   },
   {
@@ -266,7 +266,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "<a id='custom-loading'></a>"
+    "<a name='custom-loading' id='custom-loading'></a>"
    ]
   },
   {
@@ -432,7 +432,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "<a id='vocab'></a>"
+    "<a name='vocab' id='vocab'></a>"
    ]
   },
   {
@@ -494,7 +494,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "<a id='finalizing_vocab'></a>"
+    "<a name='finalizing_vocab' id='finalizing_vocab'></a>"
    ]
   },
   {
@@ -682,7 +682,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "<a id='fields'></a>"
+    "<a name='fields' id='fields'></a>"
    ]
   },
   {
@@ -817,7 +817,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "<a id='iterating'></a>"
+    "<a name='iterating' id='iterating'></a>"
    ]
   },
   {
@@ -965,7 +965,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "<a id='pretrained'></a>"
+    "<a name='pretrained' id='pretrained'></a>"
    ]
   },
   {
diff --git a/docs/source/scripts/convert_doc_to_notebooks.py b/docs/source/scripts/convert_doc_to_notebooks.py
index c8189d59..8f8a8f33 100644
--- a/docs/source/scripts/convert_doc_to_notebooks.py
+++ b/docs/source/scripts/convert_doc_to_notebooks.py
@@ -20,8 +20,8 @@
     "advanced.rst",
     "preprocessing.rst",
     "walkthrough.rst",
-    "pytorch_rnn_example.rst",
-    "tfidf_example.rst",
+    "examples/pytorch_rnn_example.rst",
+    "examples/tfidf_example.rst",
 ]
 
 _re_label = re.compile(r"\.\.\s+_([^:]*):")
@@ -451,7 +451,12 @@ def rm_first_line(text):
         ! pip install datasets spacy
         ! python -m spacy download en_core_web_sm
         """
-    )
+    ),
+    "examples/pytorch_rnn_example.rst": textwrap.dedent(
+        """\
+        ! pip install torch
+        """
+    ),
 }
 ADDITIONAL_DEPS = {k: "# Additional dependencies required to run this notebook:\n" + v for k, v in ADDITIONAL_DEPS.items()}
 
@@ -558,9 +563,7 @@ def convert_all_tutorials(path_to_docs=None, path_to_dest=None):
     for file in TUTORIAL_FILES:
         notebook_name = os.path.splitext(file)[0] + ".ipynb"
         doc_file = os.path.join(path_to_docs, file)
-        print(doc_file)
-        print(os.path.exists(doc_file))
-        notebook_file = os.path.join(path_to_dest, notebook_name)
+        notebook_file = os.path.join(path_to_dest, Path(notebook_name).name)
         convert_rst_file_to_notebook(doc_file, notebook_file, origin_folder=path_to_docs, dest_folder=path_to_dest, additional_deps=ADDITIONAL_DEPS.get(file))
 
 

From 0ce4f200cfbed6fc1d3f58a00d4ef2aca8e58ce1 Mon Sep 17 00:00:00 2001
From: mariosasko <mariosasko777@gmail.com>
Date: Thu, 1 Apr 2021 19:19:58 +0200
Subject: [PATCH 16/26] Move examples notebooks to notebooks/examples

---
 docs/source/_static/js/custom.js              |  13 +-
 .../examples/pytorch_rnn_example.ipynb        |  40 +-
 .../{ => examples}/tfidf_example.ipynb        |   0
 .../notebooks/pytorch_rnn_example.ipynb       | 502 ------------------
 .../scripts/convert_doc_to_notebooks.py       |   3 +-
 5 files changed, 29 insertions(+), 529 deletions(-)
 rename docs/source/notebooks/{ => examples}/tfidf_example.ipynb (100%)
 delete mode 100644 docs/source/notebooks/pytorch_rnn_example.ipynb

diff --git a/docs/source/_static/js/custom.js b/docs/source/_static/js/custom.js
index 26839340..e503da3b 100644
--- a/docs/source/_static/js/custom.js
+++ b/docs/source/_static/js/custom.js
@@ -16,8 +16,8 @@ const hasNotebook = [
     "advanced",
     "preprocessing",
     "walkthrough",
-    "tfidf_example",
-    "pytorch_rnn_example"
+    "examples/tfidf_example",
+    "examples/pytorch_rnn_example"
 ]
 
 function addIcon() {
@@ -51,13 +51,12 @@ function addGithubButton() {
 }
 
 function addColabLink() {
-    if (location.toString().indexOf("package_reference") !== -1) {
-        return; 
+    if (location.protocol === "file:") {
+        const pageName = location.pathname.split("/html/")[1].split(".")[0]
+    } else {
+        const pageName = location.pathname.split("/podium/")[1].split(".")[0]
     }
 
-    const parts = location.toString().split('/');
-    const pageName = parts[parts.length - 1].split(".")[0];
-
     if (hasNotebook.includes(pageName)) {
         const colabLink = `<a href="https://colab.research.google.com/github/TakeLab/podium/blob/master/docs/source/notebooks/${pageName}.ipynb">
             <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>
diff --git a/docs/source/notebooks/examples/pytorch_rnn_example.ipynb b/docs/source/notebooks/examples/pytorch_rnn_example.ipynb
index a29d6256..076f953f 100644
--- a/docs/source/notebooks/examples/pytorch_rnn_example.ipynb
+++ b/docs/source/notebooks/examples/pytorch_rnn_example.ipynb
@@ -9,7 +9,10 @@
     "# Podium installation\n",
     "! pip install podium-nlp\n",
     "# To install from source instead of the last release, comment the command above and uncomment the following one.\n",
-    "# ! pip install git+https://github.com/takelab/podium"
+    "# ! pip install git+https://github.com/takelab/podium\n",
+    "\n",
+    "# Additional dependencies required to run this notebook:\n",
+    "! pip install torch"
    ]
   },
   {
@@ -73,8 +76,8 @@
     }
    ],
    "source": [
-    "import datasets\n",
-    "imdb = datasets.load_dataset('imdb')\n",
+    "from datasets import load_dataset\n",
+    "imdb = load_dataset('imdb')\n",
     "print(imdb)\n",
     "from pprint import pprint\n",
     "pprint(imdb['train'].features)"
@@ -174,18 +177,17 @@
        "    dataset_name: imdb,\n",
        "    size: 25000,\n",
        "    fields: [\n",
-       "            Field({\n",
-       "                name: 'text',\n",
-       "                keep_raw: False,\n",
-       "                is_target: False,\n",
-       "                vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 10000})\n",
-       "            }),\n",
-       "            LabelField({\n",
-       "                name: 'label',\n",
-       "                keep_raw: False,\n",
-       "                is_target: True\n",
-       "            })\n",
-       "    \n",
+       "        Field({\n",
+       "            name: 'text',\n",
+       "            keep_raw: False,\n",
+       "            is_target: False,\n",
+       "            vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 10000})\n",
+       "        }),\n",
+       "        LabelField({\n",
+       "            name: 'label',\n",
+       "            keep_raw: False,\n",
+       "            is_target: True\n",
+       "        })\n",
        "    ]\n",
        "})\n",
        "Example({\n",
@@ -355,7 +357,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "There. We will now define the prerequisites for pytorch model training, where we will use a GPU for speed, however running the model for one epoch will is possible albeit time-consuing even without a GPU."
+    "We will now define the prerequisites for pytorch model training."
    ]
   },
   {
@@ -371,7 +373,7 @@
     "# Copy the pretrained GloVe word embeddings\n",
     "embedding_matrix.weight.data.copy_(torch.from_numpy(embeddings))\n",
     "\n",
-    "device = torch.device(\"cuda:0\")\n",
+    "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
     "model = RNNClassifier(embedding_matrix)\n",
     "model = model.to(device)\n",
     "criterion = nn.CrossEntropyLoss()\n",
@@ -476,10 +478,10 @@
    "source": [
     "from podium import Iterator\n",
     "# Closure for converting data to given device\n",
-    "def gpu_tensor(data):\n",
+    "def device_tensor(data):\n",
     "    return torch.tensor(data).to(device)\n",
     "# Initialize our iterator\n",
-    "train_iter = Iterator(imdb_train, batch_size=32, matrix_class=gpu_tensor)\n",
+    "train_iter = Iterator(imdb_train, batch_size=32, matrix_class=device_tensor)\n",
     "\n",
     "epochs = 5\n",
     "for epoch in range(epochs):\n",
diff --git a/docs/source/notebooks/tfidf_example.ipynb b/docs/source/notebooks/examples/tfidf_example.ipynb
similarity index 100%
rename from docs/source/notebooks/tfidf_example.ipynb
rename to docs/source/notebooks/examples/tfidf_example.ipynb
diff --git a/docs/source/notebooks/pytorch_rnn_example.ipynb b/docs/source/notebooks/pytorch_rnn_example.ipynb
deleted file mode 100644
index 076f953f..00000000
--- a/docs/source/notebooks/pytorch_rnn_example.ipynb
+++ /dev/null
@@ -1,502 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# Podium installation\n",
-    "! pip install podium-nlp\n",
-    "# To install from source instead of the last release, comment the command above and uncomment the following one.\n",
-    "# ! pip install git+https://github.com/takelab/podium\n",
-    "\n",
-    "# Additional dependencies required to run this notebook:\n",
-    "! pip install torch"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Pytorch RNN classifier"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "In this example, we will cover a simple RNN-based classifier model implemented in Pytorch. We will use the IMDB dataset loaded from 🤗/datasets, preprocess it with Fields and train the model briefly.\n",
-    "While having a GPU is not necessary it is recommended as otherwise training the model, even for a single epoch, will take a while."
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Loading a dataset from 🤗/datasets"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "As we have covered in [Loading 🤗 datasets](http://takelab.fer.hr/podium/walkthrough.html#hf-loading), we have implemented wrappers around 🤗 dataset classes to enable working with the plethora of datasets implemented therein. We will now briefly go through (1) loading a dataset from 🤗/datasets and (2) wrapping it in Podium classes."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "text/plain": [
-       "DatasetDict({\n",
-       "    train: Dataset({\n",
-       "        features: ['text', 'label'],\n",
-       "        num_rows: 25000\n",
-       "    })\n",
-       "    test: Dataset({\n",
-       "        features: ['text', 'label'],\n",
-       "        num_rows: 25000\n",
-       "    })\n",
-       "    unsupervised: Dataset({\n",
-       "        features: ['text', 'label'],\n",
-       "        num_rows: 50000\n",
-       "    })\n",
-       "})\n",
-       "{'label': ClassLabel(num_classes=2, names=['neg', 'pos'], names_file=None, id=None),\n",
-       " 'text': Value(dtype='string', id=None)}"
-      ]
-     },
-     "execution_count": null,
-     "metadata": {},
-     "output_type": "execute_result"
-    }
-   ],
-   "source": [
-    "from datasets import load_dataset\n",
-    "imdb = load_dataset('imdb')\n",
-    "print(imdb)\n",
-    "from pprint import pprint\n",
-    "pprint(imdb['train'].features)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "By calling `load_dataset` the dataset was downloaded and cached on disk through the `datasets` library. The dataset has two splits we are interested in (`train` and `test`).\n",
-    "The main thing we need to pay attention to are the `features` of the dataset, in this case `text` and `label`. These features, or data columns, need to be mapped to (and processed by) Podium Fields.\n",
-    "\n",
-    "For convenience, we have implemented automatic `Field` type inference from 🤗 dataset features -- however it is far from perfect as we have to make many assumptions on the way. We will now wrap the IMDB dataset in Podium and show the automatically inferred Fields."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "text/plain": [
-       "Field({\n",
-       "    name: 'text',\n",
-       "    keep_raw: False,\n",
-       "    is_target: False,\n",
-       "    vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 280619})\n",
-       "})\n",
-       "LabelField({\n",
-       "    name: 'label',\n",
-       "    keep_raw: False,\n",
-       "    is_target: True\n",
-       "})"
-      ]
-     },
-     "execution_count": null,
-     "metadata": {},
-     "output_type": "execute_result"
-    }
-   ],
-   "source": [
-    "from podium.datasets.hf import HFDatasetConverter as HF\n",
-    "splits = HF.from_dataset_dict(imdb)\n",
-    "imdb_train, imdb_test = splits['train'], splits['test']\n",
-    "imdb_train.finalize_fields() # Construct the vocab\n",
-    "print(*imdb_train.fields, sep=\"\\n\")"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "Both of the Fields were constructed well, but there are a couple of drawbacks for this concrete dataset. Firstly, the size of the vocabulary is very large (`280619`) -- we would like to trim this down to a reasonable number as we won't be using subword tokenization in this example. "
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "text/plain": [
-       "Example({\n",
-       "    text: (None, ['Bromwell', 'High', 'is', 'a', 'cartoon', 'comedy.', 'It', 'ran', 'at', 'the', 'same', 'time', 'as', 'some', 'other', 'programs', 'about', 'school', 'life,', 'such', 'as', '\"Teachers\".', 'My', '35', 'years', 'in', 'the', 'teaching', 'profession', 'lead', 'me', 'to', 'believe', 'that', 'Bromwell', \"High's\", 'satire', 'is', 'much', 'closer', 'to', 'reality', 'than', 'is', '\"Teachers\".', 'The', 'scramble', 'to', 'survive', 'financially,', 'the', 'insightful', 'students', 'who', 'can', 'see', 'right', 'through', 'their', 'pathetic', \"teachers'\", 'pomp,', 'the', 'pettiness', 'of', 'the', 'whole', 'situation,', 'all', 'remind', 'me', 'of', 'the', 'schools', 'I', 'knew', 'and', 'their', 'students.', 'When', 'I', 'saw', 'the', 'episode', 'in', 'which', 'a', 'student', 'repeatedly', 'tried', 'to', 'burn', 'down', 'the', 'school,', 'I', 'immediately', 'recalled', '.........', 'at', '..........', 'High.', 'A', 'classic', 'line:', 'INSPECTOR:', \"I'm\", 'here', 'to', 'sack', 'one', 'of', 'your', 'teachers.', 'STUDENT:', 'Welcome', 'to', 'Bromwell', 'High.', 'I', 'expect', 'that', 'many', 'adults', 'of', 'my', 'age', 'think', 'that', 'Bromwell', 'High', 'is', 'far', 'fetched.', 'What', 'a', 'pity', 'that', 'it', \"isn't!\"]),\n",
-       "    label: (None, 1)\n",
-       "})"
-      ]
-     },
-     "execution_count": null,
-     "metadata": {},
-     "output_type": "execute_result"
-    }
-   ],
-   "source": [
-    "print(imdb_train[0])"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "When inspecting a concrete instance, there are a few more things to note. Firstly, IMDB instances can be quite long (on average around 200 tokens per instance), secondly, the text wasn't tokenized properly near sentence boundaries (due to using the default `str.split` tokenizer) and lastly, the text has varying casing.\n",
-    "We will instead define our own Fields for the corresponding features, add posttokenization hooks which will transform the data, and use those Fields to replace the automatically inferred ones:"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "text/plain": [
-       "HFDatasetConverter({\n",
-       "    dataset_name: imdb,\n",
-       "    size: 25000,\n",
-       "    fields: [\n",
-       "        Field({\n",
-       "            name: 'text',\n",
-       "            keep_raw: False,\n",
-       "            is_target: False,\n",
-       "            vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 10000})\n",
-       "        }),\n",
-       "        LabelField({\n",
-       "            name: 'label',\n",
-       "            keep_raw: False,\n",
-       "            is_target: True\n",
-       "        })\n",
-       "    ]\n",
-       "})\n",
-       "Example({\n",
-       "    text: (None, ['bromwell', 'high', 'is', 'a', 'cartoon', 'comedy', '.', 'it', 'ran', 'at', 'the', 'same', 'time', 'as', 'some', 'other', 'programs', 'about', 'school', 'life', ',', 'such', 'as', '\"', 'teachers', '\"', '.', 'my', '35', 'years', 'in', 'the', 'teaching', 'profession', 'lead', 'me', 'to', 'believe', 'that', 'bromwell', 'high', \"'s\", 'satire', 'is', 'much', 'closer', 'to', 'reality', 'than', 'is', '\"', 'teachers', '\"', '.', 'the', 'scramble', 'to', 'survive', 'financially', ',', 'the', 'insightful', 'students', 'who', 'can', 'see', 'right', 'through', 'their', 'pathetic', 'teachers', \"'\", 'pomp', ',', 'the', 'pettiness', 'of', 'the', 'whole', 'situation', ',', 'all', 'remind', 'me', 'of', 'the', 'schools', 'i', 'knew', 'and', 'their', 'students', '.', 'when', 'i', 'saw', 'the', 'episode', 'in', 'which', 'a', 'student', 'repeatedly', 'tried', 'to', 'burn', 'down', 'the', 'school', ',', 'i', 'immediately', 'recalled', '.........', 'at', '..........', 'high', '.', 'a', 'classic', 'line', ':', 'inspector', ':', 'i', \"'m\", 'here', 'to', 'sack', 'one', 'of', 'your', 'teachers', '.', 'student', ':', 'welcome', 'to', 'bromwell', 'high', '.', 'i', 'expect', 'that', 'many', 'adults', 'of', 'my', 'age', 'think', 'that', 'bromwell', 'high', 'is', 'far', 'fetched', '.', 'what', 'a', 'pity', 'that', 'it', 'is', \"n't\", '!']),\n",
-       "    label: (None, 1)\n",
-       "})"
-      ]
-     },
-     "execution_count": null,
-     "metadata": {},
-     "output_type": "execute_result"
-    }
-   ],
-   "source": [
-    "from podium import Field, LabelField, Vocab\n",
-    "\n",
-    "# Lowercasing as a post-tokenization hook\n",
-    "def lowercase(raw, tokenized):\n",
-    "  return raw, [token.lower() for token in tokenized]\n",
-    "\n",
-    "# Truncating as a post-tokenization hook\n",
-    "def truncate(raw, tokenized, max_length=200):\n",
-    "    return raw, tokenized[:max_length]\n",
-    "\n",
-    "vocab = Vocab(max_size=10000)\n",
-    "text = Field(name=\"text\", \n",
-    "             numericalizer=vocab,\n",
-    "             include_lengths=True,\n",
-    "             tokenizer=\"spacy-en_core_web_sm\",\n",
-    "             posttokenize_hooks=[truncate, lowercase])\n",
-    "\n",
-    "# The labels are already mapped to indices in /datasets so we will\n",
-    "# pass them through\n",
-    "label = LabelField(name=\"label\", numericalizer=lambda x: x)\n",
-    "fields = {\n",
-    "    'text': text,\n",
-    "    'label': label\n",
-    "}\n",
-    "\n",
-    "# Use the given Fields to load the dataset again\n",
-    "splits = HF.from_dataset_dict(imdb, fields=fields)\n",
-    "imdb_train, imdb_test = splits['train'], splits['test']\n",
-    "imdb_train.finalize_fields()\n",
-    "print(imdb_train)\n",
-    "print(imdb_train[0])"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "Here, we can see the effect of our hooks and using the spacy tokenizer. Now our dataset will be a bit cleaner to work with. Some data cleaning would still be desired, such as removing tokens which only contain punctuation, but we leave this exercise to the reader :)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Loading pretrained embeddings"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "In most use-cases, we want to use pre-trained word embeddings along with our neural model. With Podium, this process is very simple. If your field uses a vocabulary, it has already built an inventory of tokens for your dataset.\n",
-    "\n",
-    "For example, we will use the [GloVe](https://nlp.stanford.edu/projects/glove/) vectors. You can read more about loading pretrained vectors in [Loading pretrained word vectors](http://takelab.fer.hr/podium/walkthrough.html#pretrained), but the procedure to load these vectors has two steps: (1) initialize the vector class, which sets all the required paths and (2) obtain the vectors for a pre-defined list of words by calling `load_vocab`."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "text/plain": [
-       "For vocabulary of size: 10000 loaded embedding matrix of shape: (10000, 300)\n",
-       "Vector for sport: [ 0.34566    0.15934    0.48444   -0.13693    0.18737    0.2678"
-      ]
-     },
-     "execution_count": null,
-     "metadata": {},
-     "output_type": "execute_result"
-    }
-   ],
-   "source": [
-    "from podium.vectorizers import GloVe\n",
-    "vocab = fields['text'].vocab\n",
-    "glove = GloVe()\n",
-    "embeddings = glove.load_vocab(vocab)\n",
-    "print(f\"For vocabulary of size: {len(vocab)} loaded embedding matrix of shape: {embeddings.shape}\")\n",
-    "# We can obtain vectors for a single word (given the word is loaded) like this:\n",
-    "word = \"sport\"\n",
-    "print(f\"Vector for {word}: {glove.token_to_vector(word)}\")"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    " -0.39159    0.4931    -0.76111   -1.4586     0.41475    0.55837\n",
-    "  ...\n",
-    "  0.13802    0.36619    0.19734    0.35701   -0.42228   -0.25242\n",
-    " -0.050651  -0.041129   0.15092    0.22084    0.52252   -0.27224  ]"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Defining a simple neural model in Pytorch"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "In this section, we will implement a very simple neural classification model -- a 2-layer BiGRU with a single hidden layer classifier on top of its last hidden state. Many improvements to the model can be made, but this is not our current focus."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "import torch\n",
-    "import torch.nn as nn\n",
-    "import torch.nn.functional as F\n",
-    "\n",
-    "from torch.nn.utils.rnn import pack_padded_sequence, pad_packed_sequence\n",
-    "\n",
-    "class RNNClassifier(nn.Module):\n",
-    "    def __init__(self, embedding, embed_dim=300, hidden_dim=300, num_labels=2):\n",
-    "      super(NLIModel, self).__init__()\n",
-    "      self.embedding = embedding\n",
-    "      self.encoder = nn.GRU(\n",
-    "            input_size=embed_dim,\n",
-    "            hidden_size=hidden_dim,\n",
-    "            num_layers=2,\n",
-    "            bidirectional=True,\n",
-    "            dropout=0.3\n",
-    "      )\n",
-    "      self.decoder = nn.Sequential(\n",
-    "            nn.Linear(2*hidden_dim, hidden_dim),\n",
-    "            nn.Tanh(),\n",
-    "            nn.Linear(hidden_dim, num_labels)\n",
-    "      )\n",
-    "\n",
-    "    def forward(self, x, lengths):\n",
-    "        e = self.embedding(x)\n",
-    "        h_pack = pack_padded_sequence(e, \n",
-    "                                      lengths,\n",
-    "                                      enforce_sorted=False,\n",
-    "                                      batch_first=True)\n",
-    "\n",
-    "        _, h = self.encoder(h_pack) # [2L x B x H]\n",
-    "\n",
-    "        # Concat last state of left and right directions\n",
-    "        h = torch.cat([h[-1], h[-2]], dim=-1) # [B x 2H]\n",
-    "        return self.decoder(h)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "We will now define the prerequisites for pytorch model training."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "embed_dim = 300\n",
-    "padding_index = text.vocab.get_padding_index()\n",
-    "embedding_matrix = nn.Embedding(len(text.vocab), embed_dim,\n",
-    "                                padding_idx=padding_index)\n",
-    "# Copy the pretrained GloVe word embeddings\n",
-    "embedding_matrix.weight.data.copy_(torch.from_numpy(embeddings))\n",
-    "\n",
-    "device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\n",
-    "model = RNNClassifier(embedding_matrix)\n",
-    "model = model.to(device)\n",
-    "criterion = nn.CrossEntropyLoss()\n",
-    "optimizer = torch.optim.Adam(model.parameters())"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "Now that we have the model setup code ready, we will first define helper method to measure accuracy of our model after each epoch:"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "import numpy as np\n",
-    "def update_stats(accuracy, confusion_matrix, logits, y):\n",
-    "    _, max_ind = torch.max(logits, 1)\n",
-    "    equal = torch.eq(max_ind, y)\n",
-    "    correct = int(torch.sum(equal))\n",
-    "\n",
-    "    for j, i in zip(max_ind, y):\n",
-    "        confusion_matrix[int(i),int(j)]+=1\n",
-    "    return accuracy + correct, confusion_matrix"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "and now the training loop for the model:"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "import tqdm\n",
-    "def train(model, data, optimizer, criterion, num_labels):\n",
-    "    model.train()\n",
-    "    accuracy, confusion_matrix = 0, np.zeros((num_labels, num_labels), dtype=int)\n",
-    "    for batch_num, batch in tqdm.tqdm(enumerate(data), total=len(data)):\n",
-    "        x, lens = batch.text\n",
-    "        y = batch.label\n",
-    "        logits = model(x, lens)\n",
-    "        accuracy, confusion_matrix = update_stats(accuracy, confusion_matrix, logits, y)\n",
-    "        loss = criterion(logits, y.squeeze())\n",
-    "        loss.backward()\n",
-    "        optimizer.step()\n",
-    "    print(\"[Accuracy]: {}/{} : {:.3f}%\".format(\n",
-    "          accuracy, len(data)*data.batch_size, accuracy / len(data) / data.batch_size * 100))\n",
-    "    return accuracy, confusion_matrix"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "and now, we are done with our model code. Let's turn back to Podium and see how we can set up batching for our training loop to start ticking."
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Minibatching data in Podium"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "We have covered batching data in [Minibatching data](http://takelab.fer.hr/podium/quickstart.html#minibatching) and advanced batching through bucketing in [Bucketing instances when iterating](http://takelab.fer.hr/podium/advanced.html#bucketing). We will use the plain Iterator and leave bucketing for you to change to see how much the model speeds up when minimizing padding. One change we would like to do when iterating over data is to obtain the data matrices as torch tensors on the `device` we defined previously. We will now demonstrate how to do this by setting the `matrix_class` argument of the `Iterator`:"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "text/plain": [
-       "[Accuracy]: 20050/25024 : 80.123%\n",
-       "[Accuracy]: 22683/25024 : 90.645%\n",
-       "[Accuracy]: 23709/25024 : 94.745%\n",
-       "[Accuracy]: 24323/25024 : 97.199%\n",
-       "[Accuracy]: 24595/25024 : 98.286%"
-      ]
-     },
-     "execution_count": null,
-     "metadata": {},
-     "output_type": "execute_result"
-    }
-   ],
-   "source": [
-    "from podium import Iterator\n",
-    "# Closure for converting data to given device\n",
-    "def device_tensor(data):\n",
-    "    return torch.tensor(data).to(device)\n",
-    "# Initialize our iterator\n",
-    "train_iter = Iterator(imdb_train, batch_size=32, matrix_class=device_tensor)\n",
-    "\n",
-    "epochs = 5\n",
-    "for epoch in range(epochs):\n",
-    "    train(model, train_iter, optimizer, criterion, num_labels=2)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "And we are done! In our case, the model takes about one minute per epoch on a GPU, but this can be sped up by using bucketing, which we recommend you try out yourself."
-   ]
-  }
- ],
- "metadata": {},
- "nbformat": 4,
- "nbformat_minor": 4
-}
diff --git a/docs/source/scripts/convert_doc_to_notebooks.py b/docs/source/scripts/convert_doc_to_notebooks.py
index 8f8a8f33..bfca3189 100644
--- a/docs/source/scripts/convert_doc_to_notebooks.py
+++ b/docs/source/scripts/convert_doc_to_notebooks.py
@@ -563,7 +563,8 @@ def convert_all_tutorials(path_to_docs=None, path_to_dest=None):
     for file in TUTORIAL_FILES:
         notebook_name = os.path.splitext(file)[0] + ".ipynb"
         doc_file = os.path.join(path_to_docs, file)
-        notebook_file = os.path.join(path_to_dest, Path(notebook_name).name)
+        notebook_file = os.path.join(path_to_dest, notebook_name)
+        Path(notebook_file).parent.mkdir(exist_ok=True)
         convert_rst_file_to_notebook(doc_file, notebook_file, origin_folder=path_to_docs, dest_folder=path_to_dest, additional_deps=ADDITIONAL_DEPS.get(file))
 
 

From 23b44d1003c76081dd77126eb04712c811f5fe18 Mon Sep 17 00:00:00 2001
From: mariosasko <mariosasko777@gmail.com>
Date: Thu, 1 Apr 2021 19:31:33 +0200
Subject: [PATCH 17/26] Fix JS colab condition

---
 docs/source/_static/js/custom.js | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)

diff --git a/docs/source/_static/js/custom.js b/docs/source/_static/js/custom.js
index e503da3b..49e2a609 100644
--- a/docs/source/_static/js/custom.js
+++ b/docs/source/_static/js/custom.js
@@ -51,11 +51,7 @@ function addGithubButton() {
 }
 
 function addColabLink() {
-    if (location.protocol === "file:") {
-        const pageName = location.pathname.split("/html/")[1].split(".")[0]
-    } else {
-        const pageName = location.pathname.split("/podium/")[1].split(".")[0]
-    }
+    const pageName = location.protocol === "file:" ? location.pathname.split("/html/")[1].split(".")[0] : location.pathname.split("/podium/")[1].split(".")[0]
 
     if (hasNotebook.includes(pageName)) {
         const colabLink = `<a href="https://colab.research.google.com/github/TakeLab/podium/blob/master/docs/source/notebooks/${pageName}.ipynb">

From dcb6b9e2c2654aabf0c568dfcbf3f027c72c7be4 Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Thu, 1 Apr 2021 20:08:56 +0200
Subject: [PATCH 18/26] Comments

---
 docs/source/examples/pytorch_rnn_example.rst | 2 +-
 podium/datasets/iterator.py                  | 4 +---
 2 files changed, 2 insertions(+), 4 deletions(-)

diff --git a/docs/source/examples/pytorch_rnn_example.rst b/docs/source/examples/pytorch_rnn_example.rst
index 0700e425..48f4f505 100644
--- a/docs/source/examples/pytorch_rnn_example.rst
+++ b/docs/source/examples/pytorch_rnn_example.rst
@@ -2,7 +2,7 @@ Pytorch RNN classifier
 =======================
 
 In this example, we will cover a simple RNN-based classifier model implemented in Pytorch. We will use the IMDB dataset loaded from 🤗/datasets, preprocess it with Fields and train the model briefly.
-While having a GPU is not necessary it is recommended as otherwise training the model, even for a single epoch, will take a while.
+While having a GPU is not necessary, it is recommended as otherwise training the model -- even for a single epoch -- will take a while.
 
 Loading a dataset from 🤗/datasets
 -----------------------------------
diff --git a/podium/datasets/iterator.py b/podium/datasets/iterator.py
index 97e23c30..bd8781aa 100644
--- a/podium/datasets/iterator.py
+++ b/podium/datasets/iterator.py
@@ -494,9 +494,7 @@ def __init__(self, dataset: DatasetBase = None, shuffle=True, add_padding=True):
             returned as python lists of ``matrix_class`` instances.
         """
 
-        batch_size = 0
-        if dataset is not None:
-            batch_size = len(dataset)
+        batch_size = len(dataset) if dataset else 0
 
         super().__init__(
             dataset=dataset,

From 63bed39ea3a912ad3f45897d2a8a24c7d762b851 Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Thu, 1 Apr 2021 20:13:17 +0200
Subject: [PATCH 19/26] Delete examples (the camera ready ones are migrated
 into docs)

---
 examples/README.md                            |  26 -
 examples/__init__.py                          |   3 -
 examples/fine_tuning_bert_example.py          | 173 ------
 examples/imdb_example.py                      | 160 ------
 examples/keywords_example.py                  |  79 ---
 examples/model_example.py                     | 102 ----
 ...ace Transformers on the IMDB Dataset.ipynb | 437 ---------------
 examples/notebooks/Podium_example.ipynb       | 501 ------------------
 examples/notebooks/img/imdb_logo_small.png    | Bin 237996 -> 0 bytes
 examples/tfidf_svm_example.py                 |  63 ---
 10 files changed, 1544 deletions(-)
 delete mode 100644 examples/README.md
 delete mode 100644 examples/__init__.py
 delete mode 100644 examples/fine_tuning_bert_example.py
 delete mode 100644 examples/imdb_example.py
 delete mode 100644 examples/keywords_example.py
 delete mode 100644 examples/model_example.py
 delete mode 100644 examples/notebooks/Fine-tuning BERT from HuggingFace Transformers on the IMDB Dataset.ipynb
 delete mode 100644 examples/notebooks/Podium_example.ipynb
 delete mode 100644 examples/notebooks/img/imdb_logo_small.png
 delete mode 100644 examples/tfidf_svm_example.py

diff --git a/examples/README.md b/examples/README.md
deleted file mode 100644
index 757e1d46..00000000
--- a/examples/README.md
+++ /dev/null
@@ -1,26 +0,0 @@
-# Podium usage examples
-
-
-## Basic
-
-[Loading and caching a Dataset](dataset_example.py): the most basic example of loading and storing a built-in podium dataset.
-
-[Full podium walkthrough](notebooks/Podium_example.ipynb): a basic-to-advanced level walkthrough of all podium features, with detailed comments.
-
-## Intermediate
-
-[TFIDF SVM example](tfidf_svm_example.py): use TfIDFVectorizer to train an SVM model for SST sentiment classification.
-
-[Model example](model_example.py): implement a basic scikit-learn MLP model on Pauza HR dataset:
-
-[KEX](keywords_example.py): use a built-in keyword extraction model.
-
-## Advanced
-
-[Experiment example](experiment_example.py): optimize hyperparameters with grid search on Pauza HR dataset.
-
-[IMDB Example](imdb_example.py): Run a Pytorch NN model on the IMDB sentiment classification dataset.
-
-[IMDB Example with BERT](notebooks/Fine-tuning%20BERT%20from%20HuggingFace%20Transformers%20on%20the%20IMDB%20Dataset.ipynb): Fine-tune a BERT model from HuggingFace Transformers on the IMDB sentiment classification dataset
-
-[NER Example](ner_example.py): Run a NER model on the Croatian NER dataset.
\ No newline at end of file
diff --git a/examples/__init__.py b/examples/__init__.py
deleted file mode 100644
index 7b51df24..00000000
--- a/examples/__init__.py
+++ /dev/null
@@ -1,3 +0,0 @@
-"""
-Package contains examples of podium usage.
-"""
diff --git a/examples/fine_tuning_bert_example.py b/examples/fine_tuning_bert_example.py
deleted file mode 100644
index c6b1b1c8..00000000
--- a/examples/fine_tuning_bert_example.py
+++ /dev/null
@@ -1,173 +0,0 @@
-# flake8: noqa
-import pickle
-
-import numpy as np
-import torch
-import torch.nn as nn
-import torch.optim as optim
-from sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score
-from transformers import DistilBertForSequenceClassification, DistilBertTokenizer
-
-from podium import Field, Iterator, LabelField, Vocab
-from podium.datasets import IMDB
-from podium.experimental.models import Experiment
-from podium.experimental.models.impl.pytorch import TorchModel, TorchTrainer
-from podium.experimental.pipeline import Pipeline
-
-
-def create_fields():
-    tokenizer = DistilBertTokenizer.from_pretrained("distilbert-base-uncased")
-
-    def text_to_tokens(string):
-        input_ids = tokenizer(
-            string,
-            max_length=128,
-            padding=False,
-            truncation=True,
-            return_attention_mask=False,
-        )["input_ids"]
-
-        return tokenizer.convert_ids_to_tokens(input_ids)
-
-    text = Field(
-        name="text",
-        tokenizer=text_to_tokens,
-        numericalizer=tokenizer.convert_tokens_to_ids,
-        padding_token=0,
-    )
-
-    label = LabelField(name="label")
-
-    return {"text": text, "label": label}
-
-
-class BertModelWrapper(nn.Module):
-    def __init__(self, **kwargs):
-        super().__init__()
-        self.model = DistilBertForSequenceClassification.from_pretrained(
-            "distilbert-base-uncased", return_dict=True
-        )
-
-    def forward(self, x):
-        attention_mask = (x != 0).long()
-        return_dict = self.model(x, attention_mask)
-        return_dict["pred"] = return_dict["logits"]
-        return return_dict
-
-
-def main():
-    # loading the IMDB dataset
-    fields = create_fields()
-    imdb_train, imdb_test = IMDB.get_dataset_splits(fields)
-
-    # setting up the experiment for fine-tuning the model
-    model_config = {
-        "lr": 1e-5,
-        "clip": float("inf"),  # disable gradient clipping
-        "num_epochs": 5,
-    }
-    model_config["num_classes"] = len(fields["label"].vocab)
-
-    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-
-    iterator = Iterator(batch_size=32)
-    trainer = TorchTrainer(model_config["num_epochs"], device, iterator, imdb_test)
-
-    # we have to swap axes to nullify the effect of swapping axes afterwards
-    # because we work with the batch-first model (we should add this option to Podium!!!)
-    feature_transformer = (
-        lambda feature_batch: feature_batch[0].astype(np.int64).swapaxes(0, 1)
-    )
-    label_transformer = lambda label_batch: label_batch[0].astype(np.int64)
-
-    experiment = Experiment(
-        TorchModel,
-        trainer=trainer,
-        feature_transformer=feature_transformer,
-        label_transform_fn=label_transformer,
-    )
-
-    experiment.fit(
-        imdb_train,
-        model_kwargs={
-            "model_class": BertModelWrapper,
-            "criterion": nn.CrossEntropyLoss(),
-            "optimizer": optim.AdamW,
-            "device": device,
-            **model_config,
-        },
-    )
-
-    # utilities for saving/loading the model
-    def save_model(model, file_path):
-        with open(file_path, "wb") as f:
-            pickle.dump(model, f)
-
-    def load_model(file_path):
-        with open(file_path, "rb") as f:
-            model = pickle.load(f)
-        return model
-
-    fitted_model = experiment.model
-
-    model_file = "bert_model.pt"
-    save_model(fitted_model, model_file)
-    loaded_model = load_model(model_file)
-
-    # here we show how to use the raw model to make predictions,
-    # this is how you can use the model that is already fine-tuned
-    cast_to_torch_transformer = lambda t: torch.from_numpy(t[0].astype(np.int64)).to(
-        device
-    )
-
-    @torch.no_grad()
-    def make_predictions(raw_model, dataset, batch_size=64):
-        raw_model.eval()
-
-        def predict(batch):
-            predictions = raw_model(cast_to_torch_transformer(batch))["pred"]
-            return predictions.cpu().numpy()
-
-        iterator = Iterator(batch_size=batch_size, shuffle=False)
-
-        predictions = []
-        for x_batch, _ in iterator(dataset):
-            batch_prediction = predict(x_batch)
-            predictions.append(batch_prediction)
-
-        return np.concatenate(predictions)
-
-    # model evaluation
-    loaded_model_raw = loaded_model.model
-    predictions = make_predictions(loaded_model_raw, imdb_test)
-    y_pred = predictions.argmax(axis=1)
-
-    _, y_true = imdb_test.batch()
-    y_true = y_true[0].ravel()
-
-    print("accuracy score:", accuracy_score(y_true, y_pred))
-    print("precision score:", precision_score(y_true, y_pred))
-    print("recall score:", recall_score(y_true, y_pred))
-    print("f1 score:", f1_score(y_true, y_pred))
-
-    # we use `Pipeline` to make predictions on raw data
-    pipe = Pipeline(
-        fields=list(fields.values()),
-        example_format="list",
-        feature_transformer=cast_to_torch_transformer,
-        model=loaded_model,
-    )
-
-    instances = [["This movie is horrible"], ["This movie is great!"]]
-
-    for instance in instances:
-        predictions = pipe.predict_raw(instance)
-        print(
-            f"instance: {instance}, predicted label: "
-            f'{fields["label"].vocab.itos[predictions.argmax()]}, '
-            f"predictions: {predictions}"
-        )
-
-
-if __name__ == "__main__":
-    main()
diff --git a/examples/imdb_example.py b/examples/imdb_example.py
deleted file mode 100644
index 35fd506d..00000000
--- a/examples/imdb_example.py
+++ /dev/null
@@ -1,160 +0,0 @@
-# flake8: noqa
-import pickle
-
-import torch
-import torch.nn as nn
-
-from podium import Field, Iterator, LabelField, Vocab
-from podium.datasets import IMDB
-from podium.experimental.models import Experiment
-from podium.experimental.models.impl.pytorch import AttentionRNN, TorchModel, TorchTrainer
-from podium.experimental.pipeline import Pipeline
-from podium.vectorizers import GloVe
-
-
-def lowercase(raw, tokenized):
-    """
-    Applies lowercasing as a post-tokenization hook.
-
-    Parameters
-    ----------
-    Raw : str
-        the untokenized input data
-    Tokenized: list(str)
-        list of tokens.
-    Returns
-    -------
-    Raw: str
-        unmodified input
-    Tokenized: list(str)
-        lowercased tokenized data
-    """
-    return raw, [token.lower() for token in tokenized]
-
-
-def max_length(raw, data, length=200):
-    """
-    Applies lowercasing as a post-tokenization hook.
-
-    Parameters
-    ----------
-    Raw : str
-        the untokenized input data
-    Tokenized: list(str)
-        list of tokens.
-    Length: int
-        maximum length for each instance
-    Returns
-    -------
-    Raw: str
-        unmodified input
-    Tokenized: list(str)
-        tokenized data truncated to `length`
-    """
-    return raw, data[:length]
-
-
-def create_fields():
-    # Define the vocabulary
-    max_vocab_size = 10000
-    min_frequency = 5
-    vocab = Vocab(max_size=max_vocab_size, min_freq=min_frequency)
-
-    text = Field(name="text", numericalizer=vocab, tokenizer="spacy-en", keep_raw=False)
-    # Add preprpocessing hooks to model
-    # 1. Lowercase
-    text.add_posttokenize_hook(lowercase)
-    # 2. Truncate to length
-    text.add_posttokenize_hook(max_length)
-
-    label = LabelField(name="label", numericalizer=Vocab(specials=()))
-    return {text.name: text, label.name: label}
-
-
-def main():
-    fields = create_fields()
-    imdb_train, imdb_test = IMDB.get_dataset_splits(fields)
-
-    # Construct vectoziter based on vocab
-    vocab = fields["text"].vocab
-    embeddings = GloVe().load_vocab(vocab)
-    print(
-        f"For vocabulary of size: {len(vocab)} loaded embedding matrix of shape: {embeddings.shape}"
-    )
-
-    # First, we will define the hyperparameters for our model.
-    # These are only used when a concrete model is trained, and can be changed between calls.
-    model_config = {
-        "rnn_type": "LSTM",
-        "embed_dim": 300,
-        "hidden_dim": 150,
-        "nlayers": 1,
-        "lr": 1e-3,
-        "clip": 5,
-        "epochs": 1,
-        "batch_size": 32,
-        "dropout": 0.0,
-        "bidirectional": True,
-    }
-
-    # Task-specific metadata
-    label_vocab = fields["label"].vocab
-    model_config["num_classes"] = len(label_vocab)
-    model_config["vocab_size"] = len(vocab)
-    model_config["pretrained_embedding"] = embeddings
-    # Run on CPU since we don't have a GPU on this machine
-    device = torch.device("cuda:0")
-    # Define the model criterion
-    criterion = nn.CrossEntropyLoss()
-
-    data_iterator = Iterator(batch_size=32)
-
-    trainer = TorchTrainer(model_config["epochs"], device, data_iterator, imdb_test)
-    experiment = Experiment(TorchModel, trainer=trainer)
-
-    model = experiment.fit(
-        imdb_train,  # Data on which to fit the model
-        model_kwargs={  # Arguments passed to the model constructor
-            "model_class": AttentionRNN,  # The wrapped concrete model
-            "criterion": criterion,  # The loss for the concrete model
-            "optimizer": torch.optim.Adam,  # Optimizer _class_
-            "device": device,  # The device to store the data on
-            **model_config,  # Delegated to the concrete model
-        },
-    )
-
-    # Check serialization for _model_ only (should be for experiment as well)
-    fitted_model = experiment.model
-
-    model_save_file = "model.pt"
-    with open(model_save_file, "wb") as dump_file:
-        pickle.dump(fitted_model, dump_file)
-
-    with open(model_save_file, "rb") as load_file:
-        loaded_model = pickle.load(load_file)
-
-    ft = experiment.feature_transformer
-    cast_to_torch_transformer = lambda t: torch.from_numpy(
-        ft.transform(t).swapaxes(0, 1)
-    ).to(device)
-
-    pipe = Pipeline(
-        fields=list(fields.values()),
-        example_format="list",
-        feature_transformer=cast_to_torch_transformer,
-        model=fitted_model,
-    )
-
-    instances = [["This movie is horrible"], ["This movie is great!"]]
-
-    for instance in instances:
-        prediction = pipe.predict_raw(instance)
-        print(
-            f"For instance: {instance}, the prediction is: "
-            f"{fields['label'].vocab.itos[prediction.argmax()]},"
-            f" with logits: {prediction}"
-        )
-
-
-if __name__ == "__main__":
-    main()
diff --git a/examples/keywords_example.py b/examples/keywords_example.py
deleted file mode 100644
index d8b70338..00000000
--- a/examples/keywords_example.py
+++ /dev/null
@@ -1,79 +0,0 @@
-"""
-Example shows how to add keyword extraction features to a dataset.
-"""
-from podium import Field
-from podium.datasets import Dataset, ExampleFactory
-from podium.preproc.yake import YAKE
-
-
-class DummyDataset(Dataset):
-    """
-    Dummmy dataset.
-    """
-
-    TEXT_FIELD_NAME = "text"
-
-    def __init__(self, texts, fields):
-        """
-        Dataset constructor.
-
-        Parameters
-        ----------
-        texts : list of str
-            list of document represented as strings
-        """
-        example_factory = ExampleFactory(fields)
-        examples = [
-            example_factory.from_dict({DummyDataset.TEXT_FIELD_NAME: text})
-            for text in texts
-        ]
-        super(DummyDataset, self).__init__(**{"examples": examples, "fields": fields})
-
-
-def keyword_extraction_main():
-    """
-    Function creates a dummmy keyword extraction dataset in Croatian language
-    and extracts the keywords.
-
-    The created dataset demonstrates how to map the input text to two fields:
-    tokens (tokenized using str.split) and keywords (extracted using YAKE).
-    """
-
-    sample_texts = [
-        """Karijera u turizmu Pjevačica Renata Končić Minea već dva tjedna radi kao
-                       prodajni predstavnik u odjelu korporativnog poslovanja turističke
-                       agencije Adriatica.net, no zbog toga se neće, tvrdi 27-godišnja
-                       Zagrepčanka, odreći glazbe. Minea se prije deset godina, kad je
-                       počinjala pjevačku karijeru i imala veliki hit 'Vrapci i komarci',
-                       ispisala iz ekonomske škole, a poslije je maturirala na dopisnoj
-                       birotehničkoj školi. (Marijana Marinović/Matko Stanković)""",
-        """MISS UNIVERSE TESTIRANA NA AIDS JOHANNESBURG - Aktualna Miss Universe
-                       podvrgnula se u utorak testu na AIDS u jednoj bolnici u
-                       Johannesburgu i izrazila nadu da će njezina popularnost uvjeriti
-                       druge ljude da učine isto. Brineta plavih očiju Natalie Gtebova,
-                       23-godišnja Kanađanka rođena u Rusiji, izjavila je da želi
-                       iskoristiti svoju titulu za podizanje svjesnosti i borbe protiv
-                       stigme koja okružuju tu bolest. »Mislim da će činjenica da sam se
-                       javno testirala govoriti vrlo mnogo. To će ohrabriti puno mladih
-                       žena da učine isto«, rekla je ona. Južnoafrička Republika bilježi
-                       najveći broj zaraženih HlV-om - više od pet milijuna ljudi. (H)""",
-    ]
-
-    tokens = Field(
-        name="tokens",
-        tokenizer="split",
-        keep_raw=True,
-    )
-    kws = Field(
-        name="keywords",
-        tokenizer=YAKE("hr"),
-        keep_raw=True,
-    )
-    fields = {DummyDataset.TEXT_FIELD_NAME: (tokens, kws)}
-    dummy_dataset = DummyDataset(texts=sample_texts, fields=fields)
-    keywords = [ex["keywords"][1] for ex in dummy_dataset]
-    print(*keywords, sep="\n")
-
-
-if __name__ == "__main__":
-    keyword_extraction_main()
diff --git a/examples/model_example.py b/examples/model_example.py
deleted file mode 100644
index f1a34b68..00000000
--- a/examples/model_example.py
+++ /dev/null
@@ -1,102 +0,0 @@
-"""
-Example how to use model on simple PauzaHR dataset.
-"""
-
-from podium import Field, Iterator, LabelField, Vocab
-from podium.datasets import PauzaHRDataset
-from podium.experimental.models import FeatureTransformer
-from podium.experimental.models.impl.fc_model import ScikitMLPClassifier
-from podium.experimental.models.impl.simple_trainers import SimpleTrainer
-from podium.storage import LargeResource
-
-
-def numericalize_pauza_rating(rating):
-    """
-    Function numericalizes pauza_hr dataset rating field.
-    """
-    label = round(float(rating) * 2) - 1
-    return label
-
-
-def label_extraction_fun(y_batch):
-    """
-    Label transform function that returns a 1-d array of rating labels.
-    """
-    return y_batch.Rating.ravel()
-
-
-def feature_extraction_fn(x_batch):
-    """
-    Feature transform function that returns an matrix containing word indexes.
-
-    Serves only as a simple demonstration.
-    """
-    x_tensor = x_batch.Text
-    return x_tensor
-
-
-def basic_pauza_hr_fields():
-    """
-    Function returns pauza-hr fields used for classification.
-    """
-    rating = LabelField(
-        name="Rating",
-        numericalizer=Vocab(specials=()),
-        pretokenize_hooks=[numericalize_pauza_rating],
-    )
-    text = Field(
-        name="Text",
-        numericalizer=Vocab(),
-        tokenizer="split",
-        keep_raw=False,
-        fixed_length=100,
-    )
-    return {"Text": text, "Rating": rating}
-
-
-def pauza_mlp_example():
-    """
-    Adjustable example that demonstrates how to use pauzahr dataset with scikit
-    MLP classifier using podium.
-    """
-
-    # Set the base repository directory
-    # This directory will be used by podium to cache all LargeResources
-    # like datasets and vectorizers loaded trough the LargeResource API
-    LargeResource.BASE_RESOURCE_DIR = "downloaded_datasets"
-
-    fields = basic_pauza_hr_fields()
-    train_set, test_set = PauzaHRDataset.get_train_test_dataset(fields=fields)
-
-    train_iter = Iterator(batch_size=100)
-
-    model = ScikitMLPClassifier(
-        classes=[i for i in range(len(fields["Rating"].vocab.itos))],
-        verbose=True,
-        hidden_layer_sizes=(50, 20),
-        solver="adam",
-    )
-
-    # Define a FeatureTranformer used to extract and transform feature matrices
-    # from feature batches
-    feature_transformer = FeatureTransformer(feature_extraction_fn)
-
-    trainer = SimpleTrainer()
-    trainer.train(
-        model,
-        train_set,
-        iterator=train_iter,
-        feature_transformer=feature_transformer,
-        label_transform_fun=label_extraction_fun,
-        **{trainer.MAX_EPOCH_KEY: 10},
-    )
-
-    test_batch_x, test_batch_y = next(iter(train_iter(train_set)))
-    x_test = feature_transformer.transform(test_batch_x)
-    y_test = label_extraction_fun(test_batch_y)
-    prediction = model.predict(X=x_test)
-    print("Expected:\t", y_test, "\n", "Given:\t\t", prediction[model.PREDICTION_KEY])
-
-
-if __name__ == "__main__":
-    pauza_mlp_example()
diff --git a/examples/notebooks/Fine-tuning BERT from HuggingFace Transformers on the IMDB Dataset.ipynb b/examples/notebooks/Fine-tuning BERT from HuggingFace Transformers on the IMDB Dataset.ipynb
deleted file mode 100644
index 24c9641b..00000000
--- a/examples/notebooks/Fine-tuning BERT from HuggingFace Transformers on the IMDB Dataset.ipynb	
+++ /dev/null
@@ -1,437 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Fine-tuning BERT from HuggingFace Transformers on the IMDB Dataset\n",
-    "> __Note__: In this tutorial we use a smaller version of BERT, called __DistilBert__, that is easier and faster to train."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 1,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "import pickle\n",
-    "\n",
-    "import numpy as np\n",
-    "\n",
-    "import torch\n",
-    "import torch.nn as nn\n",
-    "import torch.optim as optim\n",
-    "\n",
-    "from sklearn.metrics import accuracy_score, f1_score, precision_score, recall_score\n",
-    "\n",
-    "from transformers import DistilBertTokenizer, DistilBertForSequenceClassification\n",
-    "\n",
-    "from podium.datasets import IMDB, Iterator\n",
-    "from podium.models import Experiment\n",
-    "from podium.models.impl.pytorch import TorchTrainer, TorchModel\n",
-    "from podium.pipeline import Pipeline\n",
-    "from podium.storage import Field, LabelField, Vocab"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "### Defining the fields\n",
-    "\n",
-    "* text - applies `BertTokenizer` to the instance data. We don't store `attention_mask` to reduce the memory footprint of the dataset, instead we create it ourselves on the fly.\n",
-    "* label - stores binary labels that represent the sentiment of an instance."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 2,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def create_fields():\n",
-    "    tokenizer = DistilBertTokenizer.from_pretrained('distilbert-base-uncased')\n",
-    "    \n",
-    "    def text_to_tokens(string):\n",
-    "        input_ids = tokenizer(string,\n",
-    "                              max_length=128,\n",
-    "                              padding=False,\n",
-    "                              truncation=True,\n",
-    "                              return_attention_mask=False\n",
-    "                             )['input_ids']\n",
-    "        \n",
-    "        return tokenizer.convert_ids_to_tokens(input_ids)\n",
-    "        \n",
-    "    text = Field(name='text',\n",
-    "                 tokenizer=text_to_tokens,\n",
-    "                 custom_numericalize=tokenizer.convert_tokens_to_ids,\n",
-    "                 padding_token=0)\n",
-    "    \n",
-    "    label = LabelField(name='label', vocab=Vocab(specials=()))\n",
-    "    \n",
-    "    return {\n",
-    "        'text': text,\n",
-    "        'label': label\n",
-    "    } "
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "### Wrapping the model"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "In this tutorial we will be using `DistilBertForSequenceClassification`. This model has additional layers on top of the base model, `DistilBertModel`, to perform classification. Moreover, `DistilBertForSequenceClassification` is a standard PyTorch `Module` so it can be easily wrapped in another `Module` that has a proper interface to Podium - the model has to return a dictionary that has a key `pred` that points to the model predictions. As mentioned earlier, we are creating the attention mask ourselves so this is a good place to do it. "
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 3,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "class BertModelWrapper(nn.Module):\n",
-    "    \n",
-    "    def __init__(self, **kwargs):\n",
-    "        super().__init__()\n",
-    "        self.model = DistilBertForSequenceClassification.from_pretrained('distilbert-base-uncased', \n",
-    "                                                                         return_dict=True)\n",
-    "        \n",
-    "    def forward(self, x):\n",
-    "        attention_mask = (x != 0).long()\n",
-    "        return_dict = self.model(x, attention_mask)\n",
-    "        return_dict['pred'] = return_dict['logits']\n",
-    "        return return_dict"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "### Loading the IMDB dataset"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 5,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "100%|██████████| 84.1M/84.1M [00:08<00:00, 9.56MB/s]\n"
-     ]
-    }
-   ],
-   "source": [
-    "fields = create_fields()\n",
-    "imdb_train, imdb_test = IMDB.get_dataset_splits(fields)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "#### Check the loaded data \n",
-    "\n",
-    "> __Note__: `None` values in the output are for caching purposes."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 6,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "(None, ['[CLS]', 'dominic', '##k', '(', 'nicky', ')', 'luciano', 'wears', 'a', \"'\", 'hulk', \"'\", 't', '-', 'shirt', 'and', 'tr', '##udge', '##s', 'off', 'everyday', 'to', 'perform', 'his', 'duties', 'as', 'a', 'garbage', 'man', '.', 'he', 'uses', 'his', 'physical', 'power', 'in', 'picking', 'up', 'other', \"'\", 's', 'trash', 'and', 'hauling', 'it', 'to', 'the', 'town', 'dump', '.', 'he', 'reads', 'comic', '-', 'book', 'hero', 'stories', 'and', 'loves', 'wrestlers', 'and', 'wrestling', ',', 'going', 'to', 'wrestlemania', 'with', 'his', 'twin', 'brother', 'eugene', 'on', 'their', 'birthday', 'is', 'a', 'yearly', 'tradition', '.', 'he', 'talks', 'kindly', 'with', 'the', 'many', 'people', 'he', 'comes', 'in', 'contact', 'with', 'during', 'his', 'day', '.', 'he', 'reads', 'comic', 'books', ',', 'which', 'he', 'finds', 'in', 'the', 'trash', ',', 'with', 'a', 'young', 'boy', 'who', 'he', 'often', 'passes', 'by', 'while', 'on', 'the', 'garbage', 'route', '.', 'unfortunately', ',', 'dominic', '##k', 'has', '[SEP]'])\n",
-      "('positive', None)\n"
-     ]
-    }
-   ],
-   "source": [
-    "print(imdb_train[0].text)\n",
-    "print(imdb_train[0].label)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "### Setting up the Podium Experiment\n",
-    "\n",
-    "To fine-tune the model, we define an `Experiment`."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 7,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "model_config = {\n",
-    "    'lr': 1e-5,\n",
-    "    'clip': float('inf'), # disable gradient clipping\n",
-    "    'num_epochs': 5,\n",
-    "}\n",
-    "model_config['num_classes'] = len(fields['label'].vocab)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 8,
-   "metadata": {
-    "scrolled": false
-   },
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "[Batch]: 781 in 0.10555 seconds, loss=0.43808\n",
-      "Total time for train epoch: 205.91309213638306\n",
-      "[Valid]: 781 in 0.00576 seconds, loss=0.38564\n",
-      "Total time for valid epoch: 63.414066314697266\n",
-      "[Batch]: 781 in 0.10557 seconds, loss=0.39357\n",
-      "Total time for train epoch: 199.31143307685852\n",
-      "[Valid]: 781 in 0.00574 seconds, loss=0.03775\n",
-      "Total time for valid epoch: 60.488966941833496\n",
-      "[Batch]: 781 in 0.08967 seconds, loss=0.17306\n",
-      "Total time for train epoch: 196.89669013023376\n",
-      "[Valid]: 781 in 0.00577 seconds, loss=0.15879\n",
-      "Total time for valid epoch: 60.47404861450195\n",
-      "[Batch]: 781 in 0.10608 seconds, loss=0.04704\n",
-      "Total time for train epoch: 199.86573767662048\n",
-      "[Valid]: 781 in 0.00578 seconds, loss=0.00348\n",
-      "Total time for valid epoch: 60.53778100013733\n",
-      "[Batch]: 781 in 0.10526 seconds, loss=0.00671\n",
-      "Total time for train epoch: 200.03476238250732\n",
-      "[Valid]: 781 in 0.00653 seconds, loss=0.74549\n",
-      "Total time for valid epoch: 60.55412006378174\n"
-     ]
-    }
-   ],
-   "source": [
-    "device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n",
-    "\n",
-    "iterator = Iterator(batch_size=32)\n",
-    "trainer = TorchTrainer(model_config['num_epochs'], device, iterator, imdb_test)\n",
-    "\n",
-    "# here we have to swap axes to nullify the effect of swapping axes afterwards\n",
-    "# because we work with the batch-first model (we should add this option to Podium!!!)\n",
-    "feature_transformer = lambda feature_batch: feature_batch[0].astype(np.int64).swapaxes(0, 1)\n",
-    "label_transformer = lambda label_batch: label_batch[0].astype(np.int64)\n",
-    "\n",
-    "experiment = Experiment(TorchModel, \n",
-    "                        trainer=trainer, \n",
-    "                        feature_transformer=feature_transformer,\n",
-    "                        label_transform_fn=label_transformer)\n",
-    "\n",
-    "experiment.fit(imdb_train,  \n",
-    "               model_kwargs={\n",
-    "                   'model_class': BertModelWrapper,  \n",
-    "                   'criterion': nn.CrossEntropyLoss(),  \n",
-    "                   'optimizer': optim.AdamW,\n",
-    "                   'device': device,\n",
-    "                   **model_config\n",
-    "               })"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "### Saving and loading the fitted model"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "#### Utilities for saving/loading the model "
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 9,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "def save_model(model, file_path):\n",
-    "    with open(file_path, 'wb') as f:\n",
-    "        pickle.dump(model, f)\n",
-    "\n",
-    "def load_model(file_path):\n",
-    "    with open(file_path, 'rb') as f:\n",
-    "        model = pickle.load(f)\n",
-    "    return model"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 10,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "fitted_model = experiment.model\n",
-    "\n",
-    "model_file = 'bert_model.pt'\n",
-    "save_model(fitted_model, model_file)\n",
-    "loaded_model = load_model(model_file)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "### Testing and making predictions on raw data"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "#### Utilities for making predictions with the raw model on the parsed dataset"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 11,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "cast_to_torch_transformer = lambda t: torch.from_numpy(t[0].astype(np.int64)).to(device)\n",
-    "\n",
-    "@torch.no_grad()\n",
-    "def make_predictions(raw_model, dataset, batch_size=32):\n",
-    "    raw_model.eval()\n",
-    "    \n",
-    "    def predict(batch):\n",
-    "        predictions = raw_model(cast_to_torch_transformer(batch))['pred']\n",
-    "        return predictions.cpu().numpy()\n",
-    "\n",
-    "    iterator = Iterator(batch_size=batch_size, \n",
-    "                        shuffle=False)\n",
-    "    \n",
-    "    predictions = []\n",
-    "    for x_batch, _ in iterator(dataset):\n",
-    "        batch_prediction = predict(x_batch)\n",
-    "        predictions.append(batch_prediction)\n",
-    "        \n",
-    "    return np.concatenate(predictions)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "#### Model evaluation"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 14,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "accuracy score: 0.87036\n",
-      "precision score: 0.8693258875149581\n",
-      "recall score: 0.87176\n",
-      "f1 score: 0.8705412422608348\n"
-     ]
-    }
-   ],
-   "source": [
-    "loaded_model_raw = loaded_model.model\n",
-    "predictions = make_predictions(loaded_model_raw, imdb_test)\n",
-    "y_pred = predictions.argmax(axis=1)\n",
-    "\n",
-    "_, y_true = imdb_test.batch()\n",
-    "y_true = y_true[0].ravel()\n",
-    "\n",
-    "print('accuracy score:', accuracy_score(y_true, y_pred))\n",
-    "print('precision score:', precision_score(y_true, y_pred))\n",
-    "print('recall score:', recall_score(y_true, y_pred))\n",
-    "print('f1 score:', f1_score(y_true, y_pred))"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "> __Note__: `loaded_model_raw` is an instance of `BertModelWrapper`."
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "#### Making predictions on raw data "
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 15,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "instance: ['This movie is horrible'], predicted label: negative, predictions: [-3.139483   3.0472898]\n",
-      "instance: ['This movie is great!'], predicted label: positive, predictions: [ 2.5500994 -3.031134 ]\n"
-     ]
-    }
-   ],
-   "source": [
-    "pipe = Pipeline(fields=list(fields.values()), \n",
-    "                example_format='list',\n",
-    "                feature_transformer=cast_to_torch_transformer,\n",
-    "                model=loaded_model)\n",
-    "\n",
-    "instances = [\n",
-    "    ['This movie is horrible'],\n",
-    "    ['This movie is great!']\n",
-    "]\n",
-    "\n",
-    "for instance in instances:\n",
-    "    predictions = pipe.predict_raw(instance)\n",
-    "    print(f'instance: {instance}, predicted label: '\n",
-    "          f'{fields[\"label\"].vocab.itos[predictions.argmax()]}, '\n",
-    "          f'predictions: {predictions}')"
-   ]
-  }
- ],
- "metadata": {
-  "kernelspec": {
-   "display_name": "podium-env",
-   "language": "python",
-   "name": "podium-env"
-  },
-  "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 3
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
-   "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython3",
-   "version": "3.6.10"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 4
-}
diff --git a/examples/notebooks/Podium_example.ipynb b/examples/notebooks/Podium_example.ipynb
deleted file mode 100644
index 7ea8558a..00000000
--- a/examples/notebooks/Podium_example.ipynb
+++ /dev/null
@@ -1,501 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "code",
-   "execution_count": 1,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "import math\n",
-    "import time\n",
-    "import podium\n",
-    "import pickle\n",
-    "\n",
-    "import torch\n",
-    "import torch.nn as nn\n",
-    "import numpy as np"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Dataset loading & preprocessing\n",
-    "\n",
-    "When using `podium` you have three options for data loading:\n",
-    "1. Use one of our built-in datasets\n",
-    "2. Use a flexible data loader from a predefined format (`TabularDataset` loads from `.csv`, `.tsv` files)\n",
-    "3. Write your own data loader for a dataset in a custom format\n",
-    "\n",
-    "### IMDB sentiment classification\n",
-    "\n",
-    "![Imdb logo](img/imdb_logo_small.png)\n",
-    "\n",
-    "For this walkthough, we will use the [IMDB sentiment classification dataset](https://ai.stanford.edu/~amaas/data/sentiment/). This dataset is built-in, so let's check what exactly does that mean.\n",
-    "\n",
-    "- Each built-in dataset has a static method `get_dataset_splits` which downloads and caches the splits for that model and returns them as a tuple (train, valid?, test).\n",
-    "  - Note: the IMDB dataset only has a train and test split\n",
-    "- We will first load the IMDB dataset with default `Fields` (preprocessing pipelines) and check whether we might want to modify something.\n",
-    "- You can inspect the default fields by calling the `get_default_fields` static method of the dataset"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 2,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "from podium.datasets import IMDB\n",
-    "imdb_train, imdb_test = IMDB.get_dataset_splits()"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 3,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Data in a single dataset instance:\n",
-      "==================================================\n",
-      "(None, ['I', 'am', 'and', 'was', 'very', 'entertained', 'by', 'the', 'movie', '.', 'It', 'was', 'my', 'all', 'time', 'favorite', 'movie', 'of', '1976', '.', 'Being', 'raised', 'in', 'the', '70', \"'s\", ',', 'I', 'was', 'so', 'in', 'love', 'with', 'Kris', 'Kristoffersons', 'look', 'and', 'demeanor', ',', 'of', 'course', 'I', 'am', 'no', 'movie', 'critic', ',', 'but', 'for', 'the', 'time', 'era', ',', 'I', 'think', 'it', 'was', 'very', 'good', '.', 'I', 'very', 'much', 'like', 'the', 'combo', 'of', 'Streisand', 'and', 'Kristofferson', '.', 'I', 'thought', 'they', 'worked', 'very', 'well', 'together', '.', 'I', 'have', 'seen', 'the', 'movie', 'many', 'times', 'and', 'still', 'love', 'the', 'two', 'of', 'them', 'as', 'Esther', 'and', 'John', 'Norman', '.', 'I', 'am', 'a', 'very', 'huge', 'fan', 'of', 'Kris', 'and', 'see', 'him', 'in', 'concert', 'when', 'I', 'can', '.', 'What', 'a', 'talented', 'singer', 'song', 'writer', ',', 'not', 'to', 'mention', ',', 'actor', '.', 'I', 'have', 'seen', 'him', 'in', 'many', 'movies', ',', 'but', 'still', 'think', 'back', 'to', 'A', 'star', 'is', 'Born', '.'])\n",
-      "('positive', None)\n",
-      "==================================================\n",
-      "Input text length interval [11, 2789] \n",
-      "Average length 272.42864 +- 202.5628456251304\n"
-     ]
-    }
-   ],
-   "source": [
-    "first_instance = imdb_train[0]\n",
-    "text, label = first_instance.text, first_instance.label\n",
-    "\n",
-    "\n",
-    "# Note that the text is cased\n",
-    "print(\"Data in a single dataset instance:\")\n",
-    "print(\"=\"*50)\n",
-    "print(text)\n",
-    "print(label)\n",
-    "print(\"=\"*50)\n",
-    "\n",
-    "def get_text_statistics(dataset):\n",
-    "    instance_lengths = [len(ex.text[1]) for ex in dataset]\n",
-    "    print(f\"Input text length interval [{min(instance_lengths)}, {max(instance_lengths)}] \\n\" \n",
-    "                             f\"Average length {np.mean(instance_lengths)} +- {np.std(instance_lengths)}\")\n",
-    "get_text_statistics(imdb_train)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "### Using hooks during data preprocessing\n",
-    "The average length of instances in the dataset is large, while the longest instance has 2789 tokens. \n",
-    "Instances of this length are likely to cause memory issues when batched and transferred to GPU, so we would like to limit this. We might also want only lowercase data in our instances.\n",
-    "\n",
-    "TODO: add data processing graph (cf prez)\n",
-    "\n",
-    "We can implement this ourselves easily by adding `hooks` to our model. Hooks are methods with a standardized signature which view and modify the data flowing through the preprocessing pipeline at **two** points.\n",
-    "1. **Pre-tokenization hooks**:\n",
-    "  - pre-tokenization hooks work on raw data (the loaded input string). You might want to lowercase data during pre-tokenization, but keep in mind that most tokenizers (such as `spacy`) are sensitive to casing and might produce bad results. Since we use `spacy` as the `IMDB` tokenizer, this is not a good choice and we might want to delegate lowercasing to post-tokenization.\n",
-    "2. **Post-tokenization hooks**:\n",
-    "  - post-tokenization hooks work on raw **and** tokenized data. Here you might want to limit the length of your instances to a fixed amount or filter out stop-words.\n",
-    "  \n",
-    "Hooks are added to a `Field` and only apply to data being processed by that field. If you want to use the same hook on multiple fields, you can do that but you need to add the same hook to multiple fields."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 4,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "## The signature of post-tokenization hooks has *two* arguments: raw and tokenized data\n",
-    "def lowercase(raw, tokenized):\n",
-    "    \"\"\"Applies lowercasing as a post-tokenization hook\n",
-    "    \n",
-    "    Parameters\n",
-    "    ----------\n",
-    "    Raw : str\n",
-    "        the untokenized input data\n",
-    "    Tokenized: list(str)\n",
-    "        list of tokens.\n",
-    "    Returns\n",
-    "    -------\n",
-    "    Raw: str \n",
-    "        unmodified input\n",
-    "    Tokenized: list(str) \n",
-    "        lowercased tokenized data\n",
-    "    \"\"\"\n",
-    "    return raw, [token.lower() for token in tokenized]\n",
-    "\n",
-    "def max_length(raw, data, length=200):\n",
-    "    \"\"\"Applies lowercasing as a post-tokenization hook\n",
-    "    \n",
-    "    Parameters\n",
-    "    ----------\n",
-    "    Raw : str\n",
-    "        the untokenized input data\n",
-    "    Tokenized: list(str)\n",
-    "        list of tokens.\n",
-    "    Length: int\n",
-    "        maximum length for each instance \n",
-    "    Returns\n",
-    "    -------\n",
-    "    Raw: str \n",
-    "        unmodified input\n",
-    "    Tokenized: list(str) \n",
-    "        tokenized data truncated to `length`\n",
-    "    \"\"\"\n",
-    "    return raw, data[:length]\n",
-    "\n",
-    "from podium.storage import LabelField, Field, Vocab\n",
-    "\n",
-    "def create_fields():\n",
-    "    # Define the vocabulary\n",
-    "    max_vocab_size = 10000\n",
-    "    min_frequency = 5\n",
-    "    vocab = Vocab(max_size=max_vocab_size, min_freq=min_frequency)\n",
-    "\n",
-    "    text = Field(name='text', vocab=vocab, tokenizer='spacy', store_as_raw=False)\n",
-    "    # Add preprpocessing hooks to model\n",
-    "    # 1. Lowercase\n",
-    "    text.add_posttokenize_hook(lowercase)\n",
-    "    # 2. Truncate to length\n",
-    "    text.add_posttokenize_hook(max_length)\n",
-    "\n",
-    "    label = LabelField(name='label', vocab = Vocab(specials=()))\n",
-    "    return {text.name : text, label.name: label}"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "Now let's create our modified fields and load the dataset"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 5,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "fields = create_fields()\n",
-    "imdb_train, imdb_test = IMDB.get_dataset_splits(fields)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 6,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Input text length interval [11, 200] \n",
-      "Average length 170.3254 +- 41.14426515129417\n"
-     ]
-    }
-   ],
-   "source": [
-    "# Check whether the preprocessing worked\n",
-    "get_text_statistics(imdb_train)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Load pretrained embeddings\n",
-    "\n",
-    "In most use-cases, we want to use pre-trained word embeddings. With `podium`, this process is incredibly simple. If your field uses a vocabulary, it has already built an inventory of tokens for your dataset.\n",
-    "\n",
-    "`Podium` offers a number of implemented `vectorizers` and a class ([BasicVectorStorage](https://github.com/mttk/podium/blob/master/podium/storage/vectorizers/vectorizer.py#L218)) which is able to load the standardized word2vec-style format of word embeddings from disk.\n",
-    "\n",
-    "For example, we will use the [GloVe](https://nlp.stanford.edu/projects/glove/) vectors. The procedure to load these vectors has two steps:\n",
-    "1. Initialize the vector class, which sets all the required paths\n",
-    "  - Right now, the vectors are not yet loaded from disk as you usually don't want to load the full file\n",
-    "2. Get the vectors for a pre-defined list of words by calling `load_vocab`\n",
-    "  - The argument can be a `Vocab` object (which is itself an `iterable` of strings), or any sequence of strings\n",
-    "  \n",
-    "The output of the function call is a numpy matrix of word embeddings which you can then pass to your model to initialize the embedding matrix or to be used otherwise."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 7,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "For vocabulary of size: 10000 loaded embedding matrix of shape: (10000, 300)\n",
-      "Vector for sport: [ 0.34566    0.15934    0.48444   -0.13693    0.18737    0.2678\n",
-      " -0.39159    0.4931    -0.76111   -1.4586     0.41475    0.55837\n",
-      "  0.021504   0.28509   -0.30284    0.021432   0.071542   0.53333\n",
-      "  0.18084   -0.40818   -0.37935    0.86781    0.4492     0.67524\n",
-      "  0.052925  -0.42635    0.46103    0.031358   0.5166    -0.081332\n",
-      "  0.35399   -0.53411   -0.22646   -0.091881  -0.5428     0.61143\n",
-      "  0.67188    0.079147  -0.21608    0.2817     0.22489    0.0087532\n",
-      "  0.096962  -0.19739    0.61631    0.19901    0.37232    0.13627\n",
-      "  0.18799    0.35549   -0.74315   -0.08624    0.1883    -0.53688\n",
-      "  0.082768   0.78362   -0.31835   -0.21211   -0.023984   0.0095728\n",
-      "  0.28296    0.30896   -0.38585   -0.55332   -0.66063    0.30262\n",
-      " -0.65959    0.25509   -0.55906   -0.3507    -0.11521    0.32024\n",
-      " -0.25448    0.12392   -0.22941   -0.30697   -0.073869  -0.021054\n",
-      "  0.26082    0.45367    0.32809    0.56492    0.067392  -0.54504\n",
-      " -0.45747    0.29705   -0.007364  -0.15747   -0.19674    0.44562\n",
-      "  0.10822   -0.010847  -0.020063  -0.54242    0.85393   -0.098667\n",
-      "  0.22144   -0.1058     0.08902    0.059909   0.17442    0.34239\n",
-      "  0.071011   0.32127    0.099133   0.60742    0.52472   -0.13034\n",
-      " -0.10823    0.16208    0.30687   -0.079381   0.046322  -0.96637\n",
-      " -0.034772  -0.032899   0.44984   -0.0086    -0.24193    0.067582\n",
-      " -0.13938    0.19137    0.46692    0.58191   -0.20529   -0.29932\n",
-      " -0.19122    0.47751    0.011554   0.016327   0.14571   -0.29702\n",
-      " -0.22549    0.078491  -0.2856     0.027681  -0.050709   0.08904\n",
-      " -0.10603    0.5987     0.17281    0.0021526  0.93064   -0.057695\n",
-      "  0.17412   -0.31447    0.0025599 -0.079155  -0.08038    0.62715\n",
-      "  0.62819    0.25948   -0.25359   -0.38077    0.30185   -0.93266\n",
-      " -0.02277    0.40116   -0.0079393 -0.029146   0.2834    -0.42381\n",
-      " -0.75231   -0.03341    0.18793   -0.80745   -0.1471     0.12451\n",
-      "  0.16207   -0.83302   -0.48239    0.12213   -1.0567     0.60332\n",
-      "  0.25437   -0.12537    0.28882   -0.13624    0.16769    0.82878\n",
-      "  0.33491    0.57352    0.0038668  0.052082   0.87144   -0.40841\n",
-      "  0.0081564  0.0021284 -0.33046   -0.16384   -0.31893   -0.16997\n",
-      " -0.24397    0.24522    0.19511   -0.17615   -0.12781    0.61104\n",
-      "  0.62262   -0.28578    1.032     -0.24093   -0.072303   0.065576\n",
-      " -0.45258   -0.15914    0.27673    0.046784  -0.23509    0.078567\n",
-      "  0.18001    0.023338  -0.7807     0.48395   -0.13503    0.15531\n",
-      "  0.12585    0.14729   -0.3931    -0.44495   -0.0121    -0.49026\n",
-      "  0.33196   -0.63298    0.49953   -0.25245   -0.30707    0.30539\n",
-      " -0.0075252  0.1769     0.20692   -0.59478    0.3011    -0.38093\n",
-      "  0.36627   -0.05798    0.35727    0.65025   -0.23389   -0.056722\n",
-      "  0.35972   -0.15963    0.15001   -0.056637  -0.63519    0.13256\n",
-      " -0.4007    -0.19513   -0.27042    0.33505   -0.15308   -0.058018\n",
-      " -0.025477  -0.4279     0.013337  -0.26202   -0.012633  -0.34509\n",
-      "  0.14835   -0.63398   -0.15411   -0.82738    0.37643    0.11571\n",
-      "  0.39404   -0.36529   -0.3185     0.35226   -0.20441    0.10508\n",
-      " -0.022458   0.32689   -0.22491   -0.14179    0.79285    0.3091\n",
-      " -1.7894    -0.24364    0.33151    0.60022   -0.088763   0.11163\n",
-      " -0.1362     0.32127    0.017934  -0.6352    -0.61595   -0.24006\n",
-      "  0.13802    0.36619    0.19734    0.35701   -0.42228   -0.25242\n",
-      " -0.050651  -0.041129   0.15092    0.22084    0.52252   -0.27224  ]\n"
-     ]
-    }
-   ],
-   "source": [
-    "# Load GloVe embeddings\n",
-    "from podium.storage.vectorizers.impl import GloVe\n",
-    "vocab = fields['text'].vocab\n",
-    "glove = GloVe()\n",
-    "embeddings = glove.load_vocab(vocab)\n",
-    "print(f\"For vocabulary of size: {len(vocab)} loaded embedding matrix of shape: {embeddings.shape}\")\n",
-    "# We can obtain vectors for a single word (given the word is loaded) like this:\n",
-    "word = \"sport\"\n",
-    "print(f\"Vector for {word}: {glove.token_to_vector(word)}\")"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Define & train a model\n",
-    "\n",
-    "Now we need to train a concrete model on our data!\n",
-    "\n",
-    "We will use a pre-defined RNN classifier with self-attention as our model.\n",
-    "The model, which is implemented in pytorch, needs to be wrapped in the `podium.model` interface so other convenience classes can be used. In this case, the classes you need to use are:\n",
-    "\n",
-    "- `podium.model` subclass: \n",
-    "  - Exposes abstract methods required to train or evaluate the model, or predict on raw data\n",
-    "- `podium.trainer` subclass:\n",
-    "  - Handles the data <-> model communication (e.g. batching, early stopping). The user only implements this class but does not explicitly use its methods.\n",
-    "- `podium.experiment` instance:\n",
-    "  - Wraps the model and its parameters to simplify multiple restarts with different choices of hyperparameters (in order to use grid search)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 8,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# First, we will define the hyperparameters for our model. \n",
-    "# These are only used when a concrete model is trained, and can be changed between calls.\n",
-    "model_config = {\n",
-    "    'rnn_type': 'LSTM',\n",
-    "    'embed_dim': 300,\n",
-    "    'hidden_dim': 150,\n",
-    "    'nlayers': 1,\n",
-    "    'lr': 1e-3,\n",
-    "    'clip': 5,\n",
-    "    'epochs': 1,\n",
-    "    'batch_size': 32,\n",
-    "    'dropout': 0.,\n",
-    "    'bidirectional': True,\n",
-    "    'gpu': -1\n",
-    "}\n",
-    "\n",
-    "# Task-specific metadata\n",
-    "label_vocab = fields['label'].vocab\n",
-    "model_config['num_classes'] = len(label_vocab)\n",
-    "model_config['vocab_size'] = len(vocab)\n",
-    "model_config['pretrained_embedding'] = embeddings\n",
-    "# Run on CPU since we don't have a GPU on this machine\n",
-    "device = torch.device('cpu:0')\n",
-    "# Define the model criterion\n",
-    "criterion = nn.CrossEntropyLoss()"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 9,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Total parameter size: 3543002\n",
-      "[Batch]: 781 in 0.42820 seconds, loss=0.52127\n",
-      "Total time for train epoch: 439.7152373790741\n",
-      "[Valid]: 781 in 0.03686 seconds, loss=0.75586\n",
-      "Total time for valid epoch: 89.66827535629272\n"
-     ]
-    }
-   ],
-   "source": [
-    "from podium.datasets import Iterator\n",
-    "from podium.experimental.models import Experiment\n",
-    "\n",
-    "from podium.experimental.models.impl.pytorch import TorchTrainer, TorchModel, AttentionRNN\n",
-    "\n",
-    "data_iterator = Iterator(batch_size=32)\n",
-    "\n",
-    "trainer = TorchTrainer(model_config['epochs'], device, data_iterator, imdb_test)\n",
-    "experiment = Experiment(TorchModel, trainer=trainer)\n",
-    "\n",
-    "model = experiment.fit(\n",
-    "    imdb_train, # Data on which to fit the model\n",
-    "    model_kwargs={ # Arguments passed to the model constructor\n",
-    "        'model_class': AttentionRNN, # The wrapped concrete model\n",
-    "        'criterion': criterion, # The loss for the concrete model\n",
-    "        'optimizer': torch.optim.Adam, # Optimizer _class_\n",
-    "        'device': device, # The device to store the data on\n",
-    "        **model_config # Delegated to the concrete model\n",
-    "    },\n",
-    ")"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 10,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Total parameter size: 3543002\n"
-     ]
-    }
-   ],
-   "source": [
-    "# Check serialization for _model_ only (should be for experiment as well)\n",
-    "import pickle\n",
-    "fitted_model = experiment.model\n",
-    "\n",
-    "model_save_file = 'model.pt'\n",
-    "with open(model_save_file, 'wb') as dump_file:\n",
-    "    pickle.dump(fitted_model, dump_file)\n",
-    "\n",
-    "with open(model_save_file, 'rb') as load_file:\n",
-    "    loaded_model = pickle.load(load_file)"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "## Pipeline: enable your model to process raw data\n",
-    "\n",
-    "So far, we have been dealing with data wrapped in podium `Dataset` instances. This might not be the case in real-world scenarios, where you want to use a trained model to process raw data.\n",
-    "\n",
-    "To simplify this, we provide a `Pipeline` class, designed to streamline raw data processing. Pipeline extends your `Experiment` class with the following functionality:\n",
-    "1. Obtain predictions from raw data\n",
-    "2. Fine-tune your model on raw data\n",
-    "3. Retrain your model on raw data"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": 11,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "For instance: ['This movie is horrible'], the prediction is: negative, with logits: [-3.9965053  4.1739044]\n",
-      "For instance: ['This movie is great!'], the prediction is: positive, with logits: [ 1.8435774 -1.751819 ]\n"
-     ]
-    }
-   ],
-   "source": [
-    "from podium.experimental.pipeline import Pipeline\n",
-    "\n",
-    "ft = experiment.feature_transformer\n",
-    "cast_to_torch_transformer = lambda t: torch.from_numpy(ft.transform(t).swapaxes(0,1)).to(device)\n",
-    "\n",
-    "pipe = Pipeline(\n",
-    "  fields = list(fields.values()),\n",
-    "  example_format = 'list',\n",
-    "  feature_transformer = cast_to_torch_transformer,\n",
-    "  model = fitted_model\n",
-    "  )\n",
-    "\n",
-    "instances = [\n",
-    "        ['This movie is horrible'], \n",
-    "        ['This movie is great!']\n",
-    "]\n",
-    "\n",
-    "for instance in instances:\n",
-    "    prediction = pipe.predict_raw(instance)\n",
-    "    print(f\"For instance: {instance}, the prediction is: {fields['label'].vocab.itos[prediction.argmax()]}, with logits: {prediction}\")\n"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": []
-  }
- ],
- "metadata": {
-  "kernelspec": {
-   "display_name": "Python 3",
-   "language": "python",
-   "name": "python3"
-  },
-  "language_info": {
-   "codemirror_mode": {
-    "name": "ipython",
-    "version": 3
-   },
-   "file_extension": ".py",
-   "mimetype": "text/x-python",
-   "name": "python",
-   "nbconvert_exporter": "python",
-   "pygments_lexer": "ipython3",
-   "version": "3.7.4"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 2
-}
diff --git a/examples/notebooks/img/imdb_logo_small.png b/examples/notebooks/img/imdb_logo_small.png
deleted file mode 100644
index 5416d95b1988a0a64ba7c919df44f8b2a932ab2a..0000000000000000000000000000000000000000
GIT binary patch
literal 0
HcmV?d00001

literal 237996
zcmdqoWo+F{xG3sS!_3SLGc!ZOP0}zkGvf*~Cr#2Yr)ij(nVFfHnbtb}zP<Ok`%3pp
zN810+NHb%P{Eo+xC41(vBfl$3Bf;atgMon|$$XPg`CE_tlVPF%mX(@SIe#lC7crS1
zuz!;;tZBsGeK^N&S}tH<2<ZQ0@VQO@kH3wCu9BLrst)F^9!Ab)V4|iD#%82aR?cP~
z4$d|%q~fX^&~@{9e~-id=Qwd^Gb2|k2Yb>VR(57!EUe6IEX=HIoPg_9FfdejM`v{n
z&#X(G$)p1*v{dg3IvM_TIxkB3`iq8PL4*k}VhC7AD5&(@9SPJJ-Z!^5GfyyS$iUau
z059;E$<YneSDl|pf0m<BP{9k;8y7O{&HR78nx+1+-U1yn3)61VZ)KLua}brD8f9FS
z&Rg*jmCpCO2HhL;Z7B`!aS$oZ_gl3IXUF|*J%yqkJoOYHB2gTEpReWF3lFETzwrn7
zdh!Tuh!<WG>!UnBKd%_lUqu(pKsG3qhNB1re4e`({BP3$A4&}#py110we@h2{rku5
zy^E!g=c>hUT4Pt0kJn&w*X4Qb)v$L6yJz*VQ%G{6lKlH>u5h+(^j@$aQ|4y`e8d4|
z7u78zZQtH)(WQ<6-foNzKkz{*SF!1J#?#uLMY<bv^co1UrfM*aX;}>JDz;sRWoohG
zXDbZsr^aX^`OMLQ;%V;F?YLF^4#^NJa<ja5-ik2f;r~4PFa=Zgiugu#?d`3KkcS2_
zBBh6)a3$&<L6%oAdIi;aq-ojxz8+7#E`;$p2CLK=V#ov`J-BUysChcG1)SV2X$EAk
zSrl-K_SA7gnLzG}7l@1tMIe%5Q@#VvHAobi{8Fzg@s)|3MGOk@f{IDLfb~oFFBgc0
zplEd80u89wZ~WyR<PaoQY1VeHv|K0QIar}%AmBUVS-)K#CwVkfzI5OCY3hF$c6`Y9
z+)l<q2iL|jh_*})NWIheXg<*>vw*xr&uG8twCF7wim*M2g)lv*Ss&Q0OJ#Zi-3?IB
zDp@Ch(1cF|(N+uzqi%O+hQNNuLXYF0Xq*H)^i4o0;~aFEAn?ot7bSlgz=a$FD-{`Q
zCSP0Y)*H`dCGD2(-lq-l_$~YWMgm&tblBxi<F+L}3s#C2V!SQu7_h%w5Nxu&{}buz
zcYr3@H@Y5NL{gT9v1YF)yIv6Kazok;U{H9hJFY01ypwt#1{rEdY403X+r_a0aQ<RY
z5Dzp+itDQqKi}sCp<x<w;<JeKM>?_qms^s>1eU75gO)G;My19<_ejyzK#El|8}~+V
zq^tEDfPs>Qfqr#%dZ1xn<h&7SXpsw}2b1WGLaZ5x8x!?BlJ}tqte3|f0PG{qI!Y1a
z3kid@4L~U9l*p#eD0T#@<GCUzU8FykKEYkm_eE&I*KYfLXJ74;<am}uDdia_a2BR9
zz)~I?6{cs4=2_v)Za4bUBkyQERkAyYINe{drF8OiI{*q7hOr&|avM4`hzAX#<;L#!
zYg6{)C&OK~=e|k;jCyENzzQjcVGx=)spChe<dgVIA?WC|aI~rZ<T#GH+anU3fGtGM
zU<j2Gxq(tB;)UzcU%7p!8<}&PYaAfiJ>egBNcTwdF81YN^NSY6A4GC!a<X3}urutA
ze0fR1BvFG+dmy^^$m`X2#l4<DvA9Oqqs2l9T=4H<q);%U0w>@>%kRY>Zl4C~LZK_l
z5Bz{xT8wXa$Bo)zU}8svG|0-{^UVn`Ln5~@QN3;?Y%lDI!b%XZDMt5Qi9^>^P?B&B
zeLa-6l#qR#1Hz^^)ib3rI3Y&lxP1N#K^(7HKTu2qi+XszVIQ312GPG-_>xHQ@!})n
z5P^r`<R5K1+=x>%B{hSGiZVe!(|E(mL@BP+19JuMdELE+ji}gm&!E?eAtkdQ7km-p
zT=UKj<Vhj43-U;usQXp5>|B7{C)qc;l>Kl)kVe0Q*LUDZF0c4;>`?VYOvS0e;QZ!Q
zFM4zC0S}$Y6H*V(8{dh=I!zlqH(XD@DV*RApzcA$1|~`@WcjNDbJadz(@7JdDQ%H5
z;!kCz*S<vIl``Q%I0tYT-5h@~`<LcPwNAknyYosj91|?t?LEfzheWJJei;(ki$p8v
z(E_?ixhV5(>oUswe0d-0)#u|=_f?royS1p0qUN%Mh%Q|>u@~M47>iJke4T<d`{JOK
zMbB?bdWLS6)S;sP8<ML-Dpdp~3E}f6RDPY!$T48_`k)#8O21FagK_@oChE20<J-2V
zf~fmwlpUQ2rlzyumFV)0F7Q}D|I(^`^}4xgV{?SJf$?|a=6NXh^YT~UkbQrEf?ad_
zT97Wspt`^F{YD}L`c|xF_v!sd=}P&PYdhU|zusETPqkL|Toh7}eeLBg_qxL3;B_8z
zNJ<vg)S$2x?6#|3%IeR<t*v&3Xy(l6X318++8;ai8i5&`%Rhi&xM=e#;*>n+0L1po
z%B{1<y1F|QNwq2A>f1sHRNULQ{4+{=L#i}}^{iXryQX!@cMmnA!%cNFqfaNA=&TPM
z9eMx-yu|H?>Ca=~;-wH!w3(O5bw~Cv<V$zkxUFHp0SLWcSs_7;bQmzIyQo0{Yz2w>
z&&A$K3vUbh5Q`f*d^iVKp3qI@jDh(|X^xTn4i3KAO!h1+H)aBGKR;JKf|?>0R?@1q
zKGbB&qS-PR(&RzmY?v8kt_G1W<)I&YWBysMfJ|Q_FSt8aWl`-O_{$&`Lo85AE<as;
zq-yjFOuzOUg+E5Q<gYXX_w@c8(Y+&N2=g6Ie@sB=3)wN%hZ$>hx473Q`Ufjl#Yd)P
zuiygG)mIb2k(Jz=V2D(Z%lr1!3C}TTX^*)8<bYv>{PkGu&nPbtW)EOy`?jVF+>G(3
z|6tyGq)WrOa+mSV-tCUy{@DKd!H}+6PX~UNT(#Z-x@=<RFO$Oi&?Ki-nD`*MM<MYj
z(q6C%6h1XOxUQzi%0#yT_dsxYSp=~HV?Upz>Y#VacL2&INnSMkR|H(IZ_*Dipy3aX
zpBJDQ<|ONO&u*7nhx7Qu0uxBn3lbi14?|!IXp4p%luJ!ce0PTg{56q+;2>KGf%l=}
z$A1ngKLF{G_lY`q(nKAUvjfBLmy!I{4K;;9$QzGt>>ukX01w*xP22s?sa>TH7@c(O
zoff4GKcHR9^S*p6`{VFP_tg0GJ;1*zd++03*sV);9^^xD2+~W!RC0*c`WS9$d;*NG
z?3K9c22TlqR38Jk#41jgKqVS)UxJ05&2(@-EIyU8dkOt7KPT?G8v(xx$k)1Ap01|`
zKpUWJSMjMEl6R6YPzmr_<&1gyRagUr4IBo0@B&7I>_8T2FWu3C)9WWQ8l!+FY!%?o
zILj^27ULP<)qEhgu8I%y5<~}7avJSqVQAz9$$@(tzX^{5YtQSOUo4)c_Y7`w?*Kx;
zK~NLuIpemwperuB=hw#{7Ppd`Po;UBf)8oGn=BBd@N;RF^_kMy+m-FdSCFuhUVUdT
zfPY@{ej5*@3O4Z$8UuJfrHpGxHUZNopFr380Cf7{F4sA5y=M@<*W1O#ef%LHA8LaN
zgkzBlh|xcVdt418fvkU2k`~U(CBEN|0j*&%-2=IG%&e<6t4`2%bU!Vip}EgNUx79>
z={sG5&mT7T25oVt@zT`O2f}Q~P$GS~Io!Ec!e<G5KUV#ZcHv1p$brno7tDjg*C1u+
zzmFReK*F+`)}{Sc!7@QWBvXOQA%1)};nUr1W#;GO6?|?}Uf8){lbJNp49@f|Ee{V7
ztoEpbo2~FpkFwTmW->UHum3(PZJQ4vKs!!oHXRdb?B{JO%2vCbhF#IaE$)o<2ZjX=
z{yrr8QrUjIKnGR7|DBi~y?3ju6r|bm_RX{eUpyEG-Ybu2DCGSgk%ZgfLtKYq@dUn$
z;X+#iV)Piu;=AI9UsA4Kw!AxrAKF5S$uTRODMqkAyJW;`>TC4{sRm>=%-zR2k$=?^
z3n+Z4IGBIf>ESwY4eKqQwZ{+R#MX-aB5lFgHGk{aSvnxT<HTl2+pTshE?5IC3iDu=
z6WXXHqN5oapY!F8`DLf|3?>7<?^ik&(0{<Ht^}nhV2b?@(aFui>_m7P9tsXu(D(1c
zhrF=MdN(m6UpcEOm@R!XVBlw#KEzR8YDu*5EA34gR_?coSE;J!NpV7ukM&URLRDAF
z({Q$eU(}BSmv1%hA&BAv!=z~4*5Zd^U6*2y{JE`;=wZ)RIddO=YO&V{Zux|e4m0)x
z;K5ctBizqE4K0Y>uo*@=zFSbf;HjsURL(zw6FbBpsCe>9gn>AOquGCQSa5oh7ybgv
z-sX$ka}$Z;;{+;qfJY!ZF^*(_9&l79qoFxe1=Fp_vXN7aI!qDO#vwemEJX(ptdK;g
zfL6{e)b@83toWoc9422~R0Kd3>Wj4lQzMbfdScg|@*)j;?_4+lH|^Zs`$iEdOyN22
zC=Q(Se)*&p?%rmd2Er|{tCI2AGx?85<4c|dYE)sxoqaxttcHT!*@QjPQAaNXS1B0X
zq=0&2ljPLEiu?GA8FKdTRAS^Y9FZ6M4?p3cIt?3s;gJ@QcuGQsC07NWEWdmOqP|67
zJk`{HfzGbirpc?Vd3cBDyne)Kq58H2S@l{v9&D8jU2>8~nb+*y!-?dwgVv=Z#;vTB
z&1kS6M^C=TcF_1K(**n#RWzaZ%n^9Piwa|-m!_F;FQ`ATy?8gDA)cTvQhCW79b(pL
zix3Ky7GGq0V)ocG0fE;|{jjlo+d>GpX)v!1rjOuB1sd0$OdO|-F#95<JtIaokg8eZ
zc;pdGT#(}rHNp$@EoV!;n>=&Xqd_PZ87#bqmJkhtZDYCqz^)A0#9PQtzlemy48j9u
zE}s5?F?~cIG?4QW8M=CPcNu89MMQ8R9z7>eJTMJRyEW*wNh~n_Ef02!^d*RajK>1P
zD`}5+8#)=qsWglPEqciJ^LT{NtGpO}Nd>t>JCKOwyJj>rcH>dFXnl@^AL%SoxM5~F
zC`xyZR1$(G;5m=Zkq^I>k}{I&Of`sfsoYclanTxkVgaO&Afj6){|B*So{qeLRELZK
zY+burb;7eEvf1DCy`HK8n=!asahF<`(0Vq{p^Q|9^85EvM4Z$SEy#q3b}a;{@w@Md
z^1#D`aoUJbRS~W+i3+KXgwQ{Ggr5~mje#ygK3VFSoDJkDv;Bt~_tLpB)u2kYKem`C
zPbBPnY4jz~z469B1VW>sz|!t@fpW+*xDkGh;G;8hN%oG!NBLOZiZMu3-thHdV``JU
zI(>=@BKQ@r4ofFP_}M^@uru&1#(_-jnRz;X=GP^cs6E^Ky;@jn_|d&%1yd<1LjYK=
zw)dGxe4`dowz~ZDoDG2_p1!3jw{FWQ+=C`0{CBdcJEhlL*?<K|DsoyWPRp<?d8LPe
z;?$NQyPG=b`dugv$ERKv+ZJY4^C4k$(4(*#BAM4A4Sskzel;2vLPN(GUatM`4vt2#
zYK%{Oee4U#!1|A7|M*AEH0J&JeD_k$$yqv|Ge(u}NCqm~!j$1X1*oH2X`A(vW$&5!
z6IgxIgO({xbrqvrp@ugA=Z`uWC%7P7KV@+A?lHi}Lu`C^gWT<~F<OF`M2~)|)0f<~
zpdUo1NFVauGbEAAl9bo^vpEMU!2PEdLHqdie=87Ry&KG6xacN_a8UB^y1_R)8!@T3
ztuf%=ZK%Kbf0Ss8*z>`_d?4&>eyDn8U36->5J;yAM_o19CKk|)=G+KOAW)Q)It+l~
zF4w>(1kj*F>n6V-B{UVCz)I!^B=Ib9ZEWBJZ)NN^P@G7vE!vs6r+I%hnwnJ(NR>$t
zd|AkN<fjH$j3%o|3wzkx1Mm>19X3yygA4C<{MHCAB(Q_KiE_8sx(+R`1~P?&H*Oso
zPZ6{~yk{OAkQ6enNV4YAdaq`>?eOP&Z!WSfXI4#)Z*I360G^&M>!<bc*_ZR<>qFWm
z3pe=H_@6^Xw>f_N4jtLVge0|$=eFZMO_B~DJxw{>z}_cMej-;TDaWx(6M=g!n>t=A
z1^!J)OY`H1#!U-^j^rsv91r{_$8|>%oNRI3)mJ<m{^1X(cEJxpp22KyqK`#_oxOB*
z<oDxdJBu~m;@DExr+G5HgNP~V=wLg`L)He_On0S+`*my`T*r*7$jE479jls9Q3MR^
zd9jg?diY;M>L!_$oyOe&!|koQ;`J6&ouTLf`ECmHw+<Gc0oHXJT^3y-uB{q7jaV5f
zE^H(A(H<#P<9#ny+Xh724s0V6{NmYJ#-QP^p1MJ$k~a;oIZLi9fYB<p4CHX%Rbivw
zt4un*-I$O!crcEow}G_FFwhjLj3utKf`x{47!(;{*`XymU<xwxbu%d(li7CH60k(l
z;JY>3`HbS)Qe=K3WS52j0S8aQ8EV2AGRS$0ww6WQ8i!@uO`hguy4nRlo2JjD2f8Pn
zX4`Ise#>B!hUq1Pyf&h`%Io7r2f39%^cBnd-Xj@^O#ZX!7{<dpw6$*!HDhZO&y{R%
z^#@&FX*#1%Iz%@~eRG_B$#rgUEN3GUeYq5eqMQ?INwXBSs`)K)2K)v`vv?JiQg}i;
z?}Q49rU{D`zRG3|j+->Nwx@<BD8R0UB=}Jau35-6JI%kiG7C}&Ue&KdB`dJBLM6+w
zlnLHZ%G-d;lcYPfYv;|Up9cA;OOBVIao~#<{%4_$n1O^m$z_s;V7|$JagQ&u2PSv`
zhGjyUU<E}oOdjCq!z98;4@nwxNsbzOeS(J<k{}AJ6_OwctA%+JsA+;mvH=}NZ!3w1
zD7oP~ws1}&KUzh5GcKhf^J7)y%nrGjbqqeoj7BbW`e!f4+F;Mdf#DTPs82L)`m$)-
zE(Xkl4f07f;!rtm%p~jX3T7FeTll22#O{n{IOiR|t(nq-C6;(i|2;~!tS4M5K}{va
z({ui<ex&yJ{_jQ$)30d}T3Rtsv!AhU(=IFYVF_j!VsGD67!n`mVJVo)0dMwbDG4wX
zJiHs&J+t@9RcVgOIT@^vPAkIYXLz{L9=|O}>M%+PB{mrdCC)R{^cQVoNhQRQaNo4F
zDpvzE)j$hj>R0*6tM1<$32YcDf`#D0yn}^0hJOqr&6R8qxS)pqh@uGoWL=h=M~d%<
zq^ikN;l-%>;m86O5;<?|UZ1Cel+J{g4EH`K<33_LA)#?m$ftK0|32VE(Q)@y;Gh<u
zFoUN-{~%-@rEM^?^8JOV$KoI$g84J<Mr0rj{5bp$^*WCqCMJzYeH1l#i>3n&Zofn$
z=b9m>&x3d9Xe9$-#Hnx=4eoR9o4QlNVh&MUN>T7B-eE7|hRHf9M_E6PxN0uu(=~(}
zhp*X9TQJo=2A_kIK4Hyc_o3;Gc8GfBj2H>UTxVlDbQQLd<V=rXLrZfID_p=A6Rm*M
z#K6?aevKo53UrZzCodsOIK`3U7fQEaW(os*We5kan)uW%-1I|;P0y_8TlK*Q09uuN
zftJXlym+xUvUC#UuP~FIREXg|*;kg6aTb9KPJjCOXasPriCA@t$ZR}ZCk_{r*q8`E
zO)%bwHIlk-rxEaB7e-_dhz(g{*1BhBA!qNa$nsbd1eW!*-@f;sQig5t6Al_t^A6-7
zsdMoA_NF_PI}LOz;NwVY*2Qq0Jr#We%G%`}!x0x(C&2|9uUwiW+=ao=JIn-ke8;BF
zQy%Q97_>$FwHFer9QO4@1{>*Uqt#f0kS@_!7?D&~gM*j(khb1ON(u~Thz5(WxAjzO
z;UG*u@1)7w2EG5@Chj=0fC@UPWCx><^aC{<N-rmI`l_yqZHwM8f~;n2q|cLy%*t_k
zh`q?sVWLw6{-?7otchG-{UEp<S^+UL=N(UpzFgDDjTvbdWwbeTFV3#m6U@dLA~bme
z9#x_APns^(_FaKOzc?q;t#5Yq#%dB|^ZlLj+dt=cb*n3tCk7jSL|tlHi>~-43>+0w
z+Tu;?B$JWkZ4x8HY{cDf8&13BCM<pG=8=4Xi~Vxa%F9%AtkQt-+sph9!=PVt%ey^V
zQ;GD--rT^~KOAM{D#f#&;A`Jep=sXjodosczss^#y3t!Rwk3LL+sM`?-iS&bL0So;
z;hlcj`=+_RKej<(CS|D1*$l_ggCn{j?ehJH<T4~yrx*?iT>(?{iz0a%{SeUA!+E(@
zIIfU!7nj?5Cmy->j89Kg@|gKgp4h9au;VJ4@TePL;`SY@{Vt@GyQZf;=>5ut@RKL5
zi=V=+qnw~vT~ZtiIgtXDQ(qL5vUKF0)fd;ZMA90!=ytvbkC96qr7S1O>AOh!NVA75
z3hj7&aEv)_VPP<H7osBh+n>ZPS+#~iRwnEY0!A}<<iZj>at31$)r0MeJ+SgmP1WKG
zP46uAtco!|L>@+e%Z}tWUbSS?`QoxZMSP(jL$W368>(WBN>6u*KXaoIT2#gc19K?6
z9|nt<59nDva(%zP1o{6iOso$%@*(~4Q(l=>eozyDvZ0Js;K+_fz~osU6Q07m495E7
zkZ18J+{F41%cHjSas4D@IhWFSBT?P9<*?+~Z;ZAq@=!|2r`M=gi2ixocZ}<WQ0O4|
zUEJk~#JWD2yPcm;)RPgVUnBPmF$uXi_Ct3ufP53{p4&J1XHZmdWrhoORdI^CM4Ra9
z8_LdwC3{k`uo`8-=o7fPuf~qn1I3YYXU)jmv~`<`FpLkRpiDPC9FtL7*@kjt>ogWB
zOm?Y`8}z~^dH$a%u;)I$e(_5%n>a{r$?K<7wFM_b-=f|zwf~%rPT~4r&#D^~qA-=p
zdr!EF*UGM@MDo4<_${S$kmG8`z}sZZi=lC9e3rZy*CwALB}a6V*NX=g8DHc_3r~xf
zmb;JH(<+{UJH+xcy-DX*Af=j8(G8In3D(Pc?}rwNSD#LL;NBa-WB}^h1#~7CTDdCn
z2MnQa{cI)6!8$`>^OrqWX*ibi_NWb&+K+G*o9EGF=HU5f!cqg4^4(p+WVx(*crEXA
zu@)>kM=!VHppb%Zbd8*GMK~TpvjC+tYg!rwVv7vlJ}{kvsym}*8ls2Reaz^Dgd_@?
z4t0U>G?kT1s)X1cX3yY95*Aq&nI9TsJN#<N3f{M(Nu+d~*>@Df<aCpnnC;9#MW3b#
zDKMM+q;nquy>kRoO&uz<nULn(qhICRbjH@WVLd4_Y*I#9ixa`Tiu;Q!eunRuL-hg-
z8@fUh{j3mcy~`!syaEb;UgxUkoS`6Ff)fpdL@fWnWZXfesX{f7kiNc01+z~ak|t*j
z%rclt4!UtxT6v>C*zIUdY-77)KK5T7gpr|SUz$Rxk=TvE#G4q2s0<&<p}in3$@S0W
zH<Sf%BBz^HQaW%FFP(cdzt3+OJs&mUWCxy8IEqh<W=ogQyuz-q3bE`EiFu{Ag+i_t
z@}Ht@KAs?Zx394J1T$T+;m$PqX$yo8HsBy#{ucG^V-9gtNiNWtQfdBSlO(EO9Nn2G
zIj!l_u>+22a`m7qnN#YBmyQ#U0v-r+&FOElKtrrCyZ#;u4(Jf47B<{|e!D?TB^ET6
z><7$`Q0dw3O&bK-F&k#OKi{8VDUosYcI!77_a3`x9j9+~4!x%xa_yXaRikYy&e)Jo
z%G4UKjM?C7-GD~Ki{E6VRMnT8TSqi4T05@j5n7}?OOn{sE-<5PBYuz=!3*-iZVdFN
z%%y3{HnT0K;*F-XSWGIP!;yi|z%@63?I`N{z4cF=hp~oUwWPLGQNF|B-h5}Dv03i%
z@4>hFl*h#Zv*eZwZEdxpMQQ4}=n2VZuO#&fVL2@W1#N{S_4-tGEUQP<M5lo81lblK
zgP9+vRpU49N^NbkqQ`VOn=~bf&xj9fH$`HX*WXV(<C}#U4C|ER?u_GXbsu5lb-{(@
zX}y}=B=V+hu-@4nO&M7NNy!@HWieSxK|9EA@Pe1J$RquX&LYJ2WmNU7qqe%^6;W#M
zST;!1aOK!l|1l1Xq@|)%^`D0AB*@D!LO|~wBB|#8xZYgK<6-<JwBNHutk$l|m<}}=
zA~B@!jfu^<hql#A-?A;djaC;aBA(i$F4FMc%?ACm`?N~u1?3Gjl;=y)I6KXKkQ_5c
z5eS8gZv(z-DHMk_aN3zla7w(-A}=z?K??pLkx4#wOT_!_^)(<@2245+>wJzB$Z9+A
zja`@f5;!chPPo028b1-@HH4#$y<mWTs19M73<;$p^#YbfpQTJ?lV50`wi@&#UF^`%
zS1Gk_H`J`{-r{yJGR&tH&c%+rc!)ik)C!n1NEH*<g%_>p4iC+*aX@A*W_Ji-<Uj%{
z*)_>J-nWg^?S|6TqG%?#F7oWUwCEovC(SPDwr7h`DHK0;^n0aFu7PWWe`LcX<SkUF
zV}bu(@k;&j_bsQ93<#CR4mG6GShCqbsK>T%?$})DG+w_?pwp{9%udx-*pV11OzgE!
zSu$?bS2s{QHHhYwg|nV&BCx8!8;8-oJM-4ikV>_lHuJ_4SCNf&4hh3m96K3dgBV`=
z==)?Ky5BlCY0yNTa4VlNf%!+q0R_bwUPV3eo*}a??B!#MAhIrnZi#EXV%Lh!9PcKr
z#F)azH0}Wt%O$JmDvXXRrzaO3>ic<WU&u!H5wjc;@fYixM4aB9`Sk6>WV5+__5uP7
zQw**ZM|MKDRSi=qYFW;*z|6IuEf#hLspt`i^Gxc8CZDv5$?yfS0&XlE^Kv5i8y!N{
z+>grI%X8GJ4y#sr<Jq1WIZKU6)JuohSaT6S?STwQ-R+%Mlb80)Y*oCXi#$DbCa7NF
zep@>(N-|%&B+bC}J#i*@^FC*P+=PR!pTfDDH?#F#Uhvf7(rEURS#&q5$8+iSj2nAR
z*EU~6FJHr<ZL|+YsO4XMvl!qBU3=p4y!eklI$2y7G_rE3YE4MOD^bS6nMNM>hd*m)
zZCB=4+vNwfLc;5xotnPt?BV*zB9QE_U+J5~&}?1b6UVnB#<7@4+wMSCF%fhZUojhy
z`SPUVHwtq;gr$nuN2EgnNQ`dxumExsy!dWTjCy=g);A_~C}n{pj|%$+c(Ove>`t2s
zSuG%bCCPxM7>wcS+Wp3n+|`?6{;lBvws%DT10|t;3`#~%Pp<yP+5SA&q;=$1ed57}
zYgm#>x0#c>LQtjC_ybH6A0i+jVD=@@FPy%YTa$apgQ97rjcQ$NxyEUl>V07q`-L-B
z&y6wY!)hPn8;d0S5`_;i%0*KPfat2DbU}2coL7&A#8%u@gFQ>yN)#3{FVq&D0`=WW
zPC|VyQ9SF3Qz~1T=oeL{n*e$opOG_=ISEbp3cMW|w_Gj~jhMBg$|(kZC7|Vm4r{Y9
zpwZO!P!NPtV$)r#{(u?^S0cg0gHL@S>wxF|AW`FiIEXSLhGVPuqkU=xILrYaXK<(-
z65dQMA2XmARNuXuwOJV1!pXmUp<91%F*(Ji$jM(V#E(*Q_2f1=^-ueUR(-u*{e7p=
zabrMs^{qGQV@AGBSQXxzi@9Z~`H9c)Ok#}d<P4|yGUBHedxJjvL#NS++tduR@*DFi
zr#^Q~Mxjky74KX9k!AVOiBJD*Uv!+~e+$rJe~4mjDA->5>A`8SoYm<%wWJYJe?HbO
z>FB<KbKSS<cBuEyNB}Yyx*A)X!tPb{*VgV;tFLl~$8C34z#11lFAes;9U9<^FaPv-
zj?e!a`_B-8D!6S`9qq!1dKz1cgnBLyX~VN_hnt_fCvA6C<|i5X-5g<m`%;h1Z?|Pm
zP=ivk%1WqL2+F?ht9=z5Fw5Ao*Sv6;m|D(|K^a}bZ`I&eQMVU#ch$GEaW(LsvmeQ>
zj8XsUlRkPil-k1P9g?x0PeOn&%at#x!%9R~Zl~1J6k6TqiM4(%H!q}-v~|&-^ptd*
z?+38t(2#oOYQ2(VY`=&VI#`2o8;top&uHX-F9>rDUBl6fVg8eL)F$~}P%@x)+V#6a
z(rs|<^G<?WOvnt8X!eMo<sm?Cm(ZyUug(1AePpIQmoY)B+5BYsXFf@|?;5U!^FuOj
zl+Fm!gGxwqB8|Nu*SV5yyRiJ_x3^!@1?rOK<x|kmU2wI7Z(PwgTJb?S4mHAzhLD;t
z?*R=1YT6?tLQ)MqaBKuc=lP9bf<Mj`u#kL%lDwlm(4yf9g!>G{a6q9(8%B&zxBlOp
zt2M`&zk3@F`df5Y2ep6Ed(X|=L_xi0d|{v+R4$x6*r`?8If8R;3w=`jhE|76Do$c1
z#|yqfP;|+D8ovs3q7|}@K{yH`=FpAI28~Ra%e{N@K7C)>d_9rMktnwum6G31O|z(L
z?R=V52^Msy6OO(#=wwctWM75P)Hwc?<P;;Tv)9(#du2265n^L|st)DDdgO(>5;Be{
z<UQ8p3r0BG?=z_NiX|@B;rop-`t@doy^PbL#UssjIye1kEeYz-2QHqN^JLTIbol6*
zXvQzv?s@jUxT~<9eI2g4We{&+PzbuyW&#)<k;nL$J&F0itl2JICM2__Cl#-<xeskK
z4CPk5@{?HliP`_>gHJ}$P^Q*1Cb8N_&RlhhyV9!0>JQzo)!_ol`ehNs<CYQhLybJ=
zQ&B{<Wwu8C23jUQ5C)wlxpmN|zkH+)$$eO61(R3CD8^<x58D@TdS>bD8HcChSDtKQ
z)|(2B|KNE`Ay&d5+ASjL{d>>jikIHnXpwls9Qp=3YgVRjnNEC(;q1e;3p%#CCse?l
z>~)_ZN80z_Y3Dj}8^&E9_gahnUgwEIf`6Re+Lm?oh-c~ZIiI<H9L|4(K7rGa9(Inn
zUeb7DMt2Hny*n`t=QW^}bDub~_$$XvU+9|(_qf%w_kml3PKTR=MMkn`LI%Hv8AZ2|
zrSE_fQqPfUkK<@!9{CoSWlLynefX*@jx^1hN)@Dc3<1GlPa;mmz%Lpp)`1hGi6Meg
zw&#w!R64Jp$*;?^rK^>QX|c0WTInYCF55^vW6Ce~+M?TbLTnNH8p!V!v3>`sx%lFQ
zrL99Y5E(2&c}+|^PSnqSp>_}?2YxI$q~|-T6UK8VxFpgvt{c`X^-hM<G5o9b+_?Fj
z@b%*=HjDe#=NCM65U-w-($i5Vvb>k?IWmvr`F|&%g{o!VJSv2KfQM0p2bu0*Awf~W
zsDefTBwFg}mMQ_m<xq=EMe;5Ylujc2v|{f>Q!Znzx`)~|5tBObbwu<BCs#Vc)|nZh
zfJr9&_33e`a>ogA5N%q9=hy{yyl#^gy-*))&)By<5j@-#Vci4sct%D-;TUae0_oT&
zweSgP^w;#E;yX2#5=yB{eXgBIglk8ANP8_ysF##W@nsyehtMQD;K%{mai+x>1t^nI
zPk;W7mS>6PQmff3o6pliu_qr|D(`#0P@l5djv~zGHJXm%Od)aMR{C#*=$QalICD9F
zjBA~hwZm_q-0xHQ-8f^cehYd_K@IGb();VhjXmbhwnp*p-}{7jfCsfYE7N{4t*#aP
zHMKy)pbQqHbwd<xG=oWLgu3Eim2ft_TRNVp2$Ek<?raODnqJd`8-mbe65;-&Ugv0(
zkVhJGUF}{kO6V9EHHcr_Y1#*@Cb~4hj~lf*+~!CSjIFg^tHkI0tRyI0P^iGk7c$?$
z(iw3^m0KBcCe<PFvX22KhI;Q5ZghUB)lKE7=iIe2i6?T4z3u1K4mVrvSXSXLV8g2G
zLOvtwHsb+BD2u-3epJ*u7e;8aL|Inp3QwnL>z-}jXcr(<0Q^aQu;lDkz}1@3{}$CN
zI_MtGA#U2kiX+t9>zF7Hs>kjc=NV>jr9S_lYpRp0SJ)WRq_;J>Ybc30vKv^g$_p<x
z;jD;nwHlLcX1%kbIv_jyn)Wr`l_C0@n0`Cmi_h=FgyTRu*Ri*mie%HJ($*gzW>AN|
z)2wk{#Zt9SfLufUI=iV0@CrE3zJqI2PYgrc+qCf=M{=!ee2f@f4691tqwtrA;#sDT
z^L-T_It1(!cjt$G2^p(H9AHzW=Yz%DvidPAq&jaGdWGMvIo@tW`$>TMv(2jqavxYg
ztMtx(azFubJw9@W9I}@Sv*VGGdXn>BVSg9glGJ_UL_UE@W|N${1BBJpDo1W*TABhR
zSGdzC{(g1HN5BaRT_h^SDai8HmZQClXpQLT%dgJ2%ZR1TxWh4cHWB!=hR}TLU!L%0
z`YKm)%F%P<oEGXM1q)(Xqk>zvYldn<S>a=R8-v>#xy&(aFU}LJbR}X`pcY>`eR@+(
zD`3|coJ~s}C=)mvU~}2xg4Aow!1Z3dzZPGa*stv2k;KP)H&fipTQW5cJrtbio_Q_O
zRmJ&s4|=C7>7HBWJWxMv?!k)w1lsS!677EGdo&oUJ(KDY`M5?eC%y8Bdt;8v@M?mL
z_r8(&&N*ntIc*wC=_S*sr^zg!ggQarNeeNrHsA~mYgwwN&0>_z?PBUcUIIG~Bz}HI
z;~kGkaWIu%*EC&ph~mAF*-Z0-pl9JdEtRbV5zis7i+wZX=@bZY9QIy#eECigRxQzR
z$Z{#eoFWCpE~t45p*}2GEyR9E{C!Mx^#PwtM5muZR`#IZbspAaK>JL}pBg9eUWvH+
zaZY!|Q2Rj`F+9F8orbQ~LDI0atta6ewbPcqM<KnPsa)$2k3P`}D!f^{+p`Igj&JC~
z@9U-9&Dq)g_d46lH`@);-4!KfI&TfjC<mXY(h>Th(~fWg@WP4PymFgtCpj_pt7?k;
z$2BXLyT+cb{x&rRrN_+|8UKCq{_x;+e3Z)c5fCaQfp@*C@agOFZj*aCKUebZ>&53g
zIyHQC`Vg==cq;n*(R3QmXaI?;nEPz*QgC{851h2SKEbp%I(7t$29j;b00=~d1j0(5
zjWGDNKo9BeMj|Kkmq0ScFvscjb^aHd_ULFb{_&8@)zJOtovgKo>d)@%?X?=hxoHgT
zhuxRXEpP@PUx=S>(?)S_sbgC=GJnLm5ec|pSkomEiqYoo4I;LlHX*Vy7=J$nkFAFV
z?6dQS@wfGxW#B1VmMJJDJukWg&ij?0sg}v}t%Ey+Uy$dlE*luEdKu#P4hPcb0&>4^
zNZli%q=rDei<i(<EB^!->)pYIhX9O;%!0;=3oiiu_RtcaWIX#y72xd`E09tc?Xlqe
z7BDfQCOm?^T^%;U!eXu2$f#0bxI51s?Xh`JIfZ+l%u*F31TS|-`o~GK8~hdPPle0r
z#3WkW%@_M0OTYH|N<$S96trv1P@XLO#&h<9w8}R3i8OQUNsJ5My8HHcBb0NmT&GqV
z32Y<v>`UuoZQ9pg%6(j2d{+1$9U4h4PfulXyAY9IGN#}_Y@1$j@LYm?=4rxwS9`p_
zUX(fCVZs{%76IlHynT4VRLxc#1~Y@VfSYc@m<R0^9JP?y<T>4#VC@ryIDG7R>7};Q
zUb3=Vgp&F2b!Obs&kh*+Dd?4UWL}~~hCpIH*LS9p@Nn4SS=49#wcdGGl&~9{W7HN0
z9L3gq_KI!65F3TkZo-=P_fYo?MU3V7m&Y!+>i(*-;i}@<#)k_VBErir^f|c?j?N5>
zd#YxM)g9QkSUThejeD+#+7vEh-2{w2<Xz2F``93%TWTiLEyd<ru0w+CE2K8=W*_S5
z-6S7sW=B@j?g^@7;MUC-|N6BsqlQu+Qz^oaVdP6T$nNDq{_S(1{L_#HXr<b3C;y7s
zaWa`KBQ%-JdaGf3(`cE51>32#8!>K#=U*XZWoq;rpWTS04tD;>MWxo>^U7|_YWN=<
zGKXAQ1cK*1p4{_@R-%M<w&7wo@sVS?jXx96`HmAcl456m2c9okoP)pv!&q#x?TX%}
z{yTo2z`ofY!(JD$ZwzLKt+4VNZm$#skIkNb#6)MC_(QNIWP`6IE28Y4KGbBo$5|aF
z+<LA5avv#jS3JEa6>l3s%Z|r}S)Um*vBXln$H=yV`!D^>LC)H|#A{$NVr2}O5H3&X
zl_9&HO#}wE-!M5}c_yZxdJM?t;+V(tm>&fA4oEAD)bSz^MA!T>H7um9z3sr|(uU@W
zkw7Jwr-_ZTM-yP`-A5N-{MS88`h^N<95{%OU(r8sgI^i?$v`TIJr;yLj!#^IAdWT*
zOnG->iTRrP8CQfnS}SerjHHs*`paxsGKy8okbeL0CPXW|&<C>chTzY{FY=-V!C+XB
z@j4Tw`Od$c?5ov`%ifo4-z!hmE}HFE{WHU6p@%I=#2Z*qTDGRkw^M5%zt+qi1y3eJ
zad|A{;^~YXQZyu<82xONzS1#!uo(4|`|gc1>lNHaRw9ool*ssk&=4^vti^*zKPEa9
z<j6NzG&k%Moy_Z!l|$nM?{nxUA(P81n<<;Z6P-3}dFsxHK$eMG2S@g<=-n`lp`UwI
zY*94q=bmX;2S=4==nO~ID4i2x#Tal3sXogvWi{d2D_CA&HV=y*n>YKP1tat%bhvqD
zEkb9@Ekr4F2OM-MY@RV96kF`hz9tl^{q&dLW@ca^wR%tmw?u*z!KiGoWYMT~e+vj9
zR2zo9;PoEI&`O%%{D@4WmID$b?+~S&f?u))W6&xNE;O;q>NgDHT6Dfuh$XX3&_`4-
zE-=HVC!=cQn>Vz-VaP?uB)zeMYk8>%Gs!s%uWGd7XbADoYoz?cD=#APD~(~@J%RbE
zaLm2Vkc7!0_;b&)3);lem4X!Xtr8Ps#*`6P4i!FgD~_0C<5VFJiA9>^r=@68`s$#E
zO>!DC6e&mN7RG?W<3e5L-2wqlle7IAfcj!Ra$Hqlj^Bu4mNTkby(4O+jk=dwJ!cRq
z5~tybg;}FHq_NtkDuH%uxT5*G=%Rf^Z4L4p4Oj$~lDFp34Ocoo&bTP13T+(X)37d%
z0=ED%N_tO_?4f*%N^dV1DwSgN>ST~Cr(v6^Jk^C9^(t9CD~4XJlB&eru-{$_n!cCc
zY%^UC)k#b_mNyBzWF@nuEzJk=4!iST)AX12B;6YixEJ?|@Cfn_x~P#2!+nt2js1qp
z-FfG0af{hC&dN2sq$?6f5F7XHBu%O=D!Ah5!m)^wNPe0I#e9c&9Xelzs0%mat*8co
z^Sf}DK6sgk121>REBc;X;O4M73sF0I#u`KW#6+U4*?qZZk-P#Hza<>|S)GLnyvC3r
znyz$iw$AHi>_cfRpaG6K9P-F<fUqi{$o|IcK#oFohY{szb`A%!UN?^*-i3ARPdbpf
z@23n0!tWfb-#?~zzd4i6J0h;A>}YDzl0f^tBB^r$)k|=2AnMWyDy7i**;)21kB0E^
z(B5mnoKYKvO%vzPun<@K#X{&SScM#XQ`-3Q-~5{(RlnRrui3(*FF5i=l073!JFUnd
z*IEQ)O(Gkxm~ZT1kox!2{efW6k|J>PQK1ABbIsgC#m{MshKRCWJiQK#-OTQ?xfk$&
z$I?BSF`Ao<U{c%0urY`#)mgF}Phqw1eUHTS;ep}Qn?2}ADV?~&!vojFE&SmKbzEYh
zoJ`;ySg?LpDT6p&&6`6Vm648&o!A!ZWSb+Qa(TdXPPg9)5rzv<9>WlGjGhGjY%3rR
z|2(vTQ1qR4IE<x<Qb31r2Ak~cR)r$o<gC#&8qolHPp;0=9Mu<%$rR92iJVfTYRN7y
z6;G$^Tk78LC6!rndK|mJMXxT2$tJil{-8&j-T%0TKd37&SG*Eej2$uwBV#Kh+n0LV
z!HIVVzIM6m*mKDPLVD7DhO^q7{_>`rFryielesVE%_PYmm|;;ern2D?-F}U2WaAye
z$fTfKAe-Dqr-mPklZW)hD4^~ddwKs6wit6>ySZSgfZ$hYV+ke!9LWz;p@b)l#6*44
zxR6G{stG|Ab0};zmrM~zU$g-~4j4%_SHbVf-V-A}S$w)Z9_X#QtWTxw$?b8UPK0)F
z!!62?0M`h$<x_Y%K7jHUtg+Jf!11^jnW}e_xb5nPn=d(A5naxOG5;5sX0B?b2&O%i
zasL;G)Y}BMV|>!EKZK=gKu386tiOzX>D%OGxOimM+9du&aM|(=&rK{%wJ4!zM96^E
zz|b?p4zQQqHY4(#jBr1LS4EvihPY!6ScG@mK<VV064W1N-ofJA_xJU+>XirX{*O~t
zkaBY?AD@$(tDL5j^k{5G6ag)nZZWbfi;3@1bZ;428WmI7Hs40kh&w+~!{7gN!f*EX
zkJ&h<tN4EF^`(EGz`gC0i&<m4{Gmc#aeezJs7>N(z_FlUUz#}(RX{*Nd!=ew_J=xH
z+NRqfgxN1ym8x?{he4e-lTDdA7k=?1awEAa=IgLJgT)@5p1hyRhH_fuoMjm#F$YgD
z>1|>Ivn1K&{Dk+!;UwPN083ceK*MI#-$~q8r#~j&tVBjsF-phYMliPe5HXjxEulkZ
zEVx%1A~my}cZ7F#GGpyS$+pe$IvG(*`DFvNw$%<&qYi~rwA+<(eDSinj(KMHB?Yof
zoJS2zsff8SX#QuIDL&2H8_qDY_>=8LanUqc-#39WLRO*lW281h_MU3+Dw@<v7?1K$
z^qD7uL)ft90ovfj_ZNSzX6`0s8pEO;t<5&+PON2>GSzU36vV+#<m4PR;-E}4C{W<+
z8@AL_G3&rtCHmqac$|6{K-=q&Fi+kWS5ypgeO2X#0P(yQF<&&I&S`P57D;@P$N?y4
ze_kc{zVYI4#BN3wZWRqF;Nb>y`wMBL>CoCMcq$EHB6%<#UhU(|3VSHlbZ|EJ|F>ki
z9};UVMu)jCkP7YMi1C!k%raV(>HOgw&7-Ov)SoG^s1Ve@0gGbw@U9}CR&7C?h#n3P
z=m~zt?`E~ej4waA{Ok(|r$+1@MOk?k^pr<^@51nE&@@eE;<`Pq?rFlQQD}endWb4T
z&D=3y)~?;TZ>w;Tzt-RX9CwcHAX7}zah2n2{YA-LpQ?k`YUkXF^fG^aGFjntY6s1x
za_OROm}`yhCR1R~ILn5vhvH3sfY>V-y(S`@q=dOX#nz;joF$o5Z#}AZ8okFzKr$1&
z>nY=>+&Qfk=yx?x)utca5c5rZdu*`OCv<R3Xj(b?(Sr88JaU}MX|=K8&$se;#Uh+v
z(O37Bm70Lmx~`0d49ZG##!9-mGcxjBIf9TMeg6i?luOQ8FV4|XR1WKExBHTxNgp)-
zcDYL*mu?)^x*}dQU8T}$ve8OSJml6*B_h=8L>zkiG$fywoH;*eTs$~P%<9lqG&^X(
zhVq!vII+_B2P6|A?;Pbgf5pm}?`xZ@RjgGbd5g5zF8=2UZN%m>7Uuq-+dlX2zV!?n
zpmg+~G}Xu0Em9sQAAD%`b3t$hg?oAY|Dj}|gUShht%3PP8rHj86pCW*l@UaHREALO
z?^Dion}B(F{hb%uSeTG8YdX%NN-GMAKV&!Z&mxm`95dpNi5y<~IKKijG$3AnK{-Nc
z*D@jPk@QQ&aP+&)UjZ)aCvRizUPt89EqwFML_fbO+kXw<;uWbe3qk2=?ZY#x!{$44
z|HWtt`0*x1zhT|5Y3^HY7Oo>5=Ub8NX34()UC%y|;n(}}hZ*evMiuJ7n~x?g{QD8A
zD&M`-3dGiY!mRDjog=qIvwj3MW0D%u(?3pq$Bvy-)L#oLTiy9X<uhc@y~6(zU?XG|
z@+&_Isc4y2B^|`~+It3syhd0Xz<@@no<u<Hk4h_UM(MPP7QBs1%Z<B>mm&3<#oL(3
zD;fvWy?E&@p2$<@j2Z@~3Ku5{;^V=lLtG{ALWPV+D=S8YCg%J{CqgZLiX|9iTUmlk
zv}IgmrSsQ|?w}!;;>>PaNz(ma9KUFPBySNDA7B<$kRNMtnIYZ2W^<#P(W+J0HEH$F
z|CnIgVOAKFdKw`-rKO|N|5~hGdkY0jAq~)O#-us`assrhZ)m;;$9KR&=L~%?<uW7_
z=b~+<lV;ZRhGp*l9`7P&=TrV8gS4?wH6Up7UY8<bn(kgpM{=&@l<@=A?w~9<|Cc_*
z+|JRXa(h|)TwAJ8t?#(vx%Sn-{6j<4Pk~yd=HCfY6|VOdi65JnZVvVo%rAa1exGXl
zVB>M32ffjat+Kw+V-OuTOzAEn?}cSMD4;Zht9SX7{})3BdbVDY6yZE?9X9sNE(b;5
z<s&$M2H||+CTMbv@%+Lzu{FG!`VysA@<Pk#8{^RK6U$oQ**)zeaK*;c7eiDVh|<z#
z`;pq#?HhWtt7*Le3}H@BO(Y;c>e-go?C>1iMt6ym($keoiz~qYNWT!b-jVj3f6?$&
zX6xh-^#4DB#!BU2{V&`9<&1X*1e*hu$X(?4`Y=e$+Trwgy*7nzzGY?M)psQ}D=iku
zerL7_9~nbcZP=o~hFdtw^a8x|L_xqHZ$*W0Nf(rlV1abtloZEE%3GhdMDT$-F1q6T
z36l!CNHHUtmW(vA)7mnwo0OYrhV9brH=eVQ(bw*8QoZAAejMs)=oA9Ro_hPZST*2I
zvubp0HsCg+D`P%W$CBIe<N_U7MtfEl#TTzxpGe0S#0mqItF7k88QH=%?x{|HR=plq
zy#m~tt8M<W+~4-;jO@#d>}w?c>Ms|*vCEKU%m0BDx$NTnoZ{R3VR5EK{|Mq=zM9U+
z|I3NjNaEElVg8YGe?Qp0Z8KW<i*>G(v$>qoI-Jtm1aa8<t|5j0PrzhGzD7(w)@M^y
zM>%i%(M7RKQx|#GU*uy9s_*ae#F_fvKxq6g5Z3+!g!L|yQw9+H{{Ufl_J0B4+`m8w
z_!kKO4)GzaZ{{B$H2Mb!k4`cQy9MI@_H~6K{Wt7i-z;Wz)$?k$M%*;>39KOZl-&=>
zvO~Y{#fh7|CS#g^aPKz;n7DT>$$NimS{mtGORRG~uPwZ{)+v0SOLiB(vNda<TkV&3
z=WZJKIZLsyf&96m8PNoXv=(*xUCYj=uaVIgYp8rh-yq|uHR-{;)UPg00Oj&sys(w6
zt$Eaq*?QH?;el*f=r_Ijg~4*g<|+<n5&c5fiLtM8C&x158V={zD%odAGsZvkOIK;s
zZD@D2CMOYG+kyEo7x2dsa{nmpwse^Or)}Uw8>cv+3h|TYV19@vxiEL&`-Yf(UK)8v
ztc25ftroy|{nN-=dT!bmyOEAf$Bm2SMj?W6saY#z+8gu1H~7;EoS=tq1*Yh0)tngc
zhPB{}V!|q^9!lkxlYDK5KbAWfDGNFs(V9P2UeJK(Z7<A-`^Yx6@z3H(K7-o+fpd`b
zPyFwGSR0HdYapmg|A2IdUJDu8ID9)Qn0P-3Zl#2I#s&|us024Kyf+%1F}R*BQoCFk
zz{FF$+^vy`E(^{>-mdJ$hqODF&f4&JZFt?)lr;AprM9FTm+@Ouhd}b=Wn+9Q{)Jca
zr<N_~%&*lpXH<QqgKs(|#^lEu`f?|dyMf6OwXTU=Gx<t1>vVq)m2Sz+tZLIC<W2<f
z7X8#%HLmvFp}4iyMEG!yOr=+UXS=BnaG+s1m6B##&hz27=yi^qFjx0=z?F_|dwdS_
zBT;tts-ov0<5V8g6YGhd$&;`IlwH}iZAl)WcTG_CQD*Ln%{%3cm|K&d>Novh)dNo*
z;I_PId!CYWYoe|6=J(!nihpr~&sf)3$sV1T;6db_tmK2Hl(;Ty)u5J#`bC6-8N&Vh
z82&oO%Nfyo3g@@!4R$)5)!@jhUjypI%5hhW+r?dAaq3Psyzc%3O$H_#O;~~&lQZnb
zmV$<s<^PEHUli}d<%=~Kf1BZ_IVMQyw;p_;GGQ{TKpV2e5KziIw(PS+ZVJY>!hMF^
zZe*?IA6n6$nzs*08uWRY;TfJ-Jany|7Fj>f-P-v3H$q^VVOQQDhpxp6S7j)(stn0@
z1Pgz~5sIn1toV?ssfUXDJJF;zEyea_it576y-FYc`K*E9g^gA^f>$K*eTq06Fs@vP
zFu6D=tXpW2e#V$=!3p&?9<kX}XQCOCjTIW`xDLawQxdFEWKs@>Tw6RpTaSM>{}?UL
zn&^%bnzT1-v5irm=p%H=eE(n6+*MFr@q*|3;10oU1Hs)TH~|71ClFkNySozzE*p1u
zg1fszaCdjNjl(6UPtWO@J5$r$RrldO!OMc8ioL4V|Fgb7>rREpZkK7qN9*foVL*rd
zf{?`tdBgJz-WK;Lrw-=CuS<wJCkPN<(>c8*=KZI0|4V3fxJ-SVSn+{-ZjtP5m7a_U
zC=cV6p^UaJ_*3O1?F5YyVpUC(foSZ(0SiHQm9MJ9X6tt{zi|;F?j?!SrDl7gw+^0i
z<#n>;ABk=b5snnF%682pc-;D(b|G4r_?I0qE@9jcPA3`}X2U3I&_`nOL9?&~u_z11
zQCXne@0)`GfnW$r=g9?c9WIRk1L8;drkURF#tp7J)uS4h=;9w##4%upJ3C)D$v%5n
znr+_e%_uH&n=6&_8Vbe?4sbBVSNO3~vYOvjL&$5n23-@_j);(V7Fv=tw+Cl2$5~zj
zsX*N8{NITm)-4aOQWE65s>IF4n@5jly7(L*Q8`&40|>zAk#tZlyazs@1$#f@TfE!;
zui9$AVDqF|G_O)!*SqJKfRAs#>``CZzC|A9o>k-IrIYdytTUCtaK#!RdDQx2yh7Yj
zgI2q}&kX+L8?63}xP`-<`Gv3y>(Ll;Z!V}2#W-pytA+$#>#otFp%YFt!E~x09RTEE
zhf9RYe5>=kYh}Zh#Hz7F^QoX(dPK}6(oY<)hx!wlmaD9WI^q>}N3ry6Zp<su{if7%
z6))dfzYYGD6{=!EB*9;;TRZ`3@o~_>;0DAn14XiX#u&U?o%aru@`N5IMHb;CW0SlL
zBYSNLOX%^%QMXRt&>k3j9)RJ9Q*S#a>G6bOGkywR9d;kx+Gn;D-vZ|`w12!@KV|t{
zBAdM5>NJH>j>stA1i&|y*cG-r2N?TCI@oD-7ux+i#@T%8`>Iv(+9_6BT>Tn)gt3L3
zVjMQ)?anX)3`8Y*UVmi2t29SM%d?ZBDKIxNxaAqBf5o_3Qtnz8Wenm+Q_B867i64H
zQ6o5-zr{>Wv+ogU$ADh_N93-7m3R%YAy37$xpAL<j;qRqwf#ZPZHH5Lcy`V#`NC+(
z+fQ+(^33&@n;_b*Mb7@NT^MpE*Vqku4aTPimM&x%SHrMF53}lUi>~<yqOZ>>ryls>
z@WT15=`WBTOtx<-mgCBY5e9!QIjY|36v$BBn>T~aktiYUy!p6-KD^%9t@nL>7W)>l
z-hyi>2}Z`N!Xf@iJL;h&ntDgOO+&&ydkWcnaVsUY3Y1^i0*`L<v^(ehjx>%mm0=DY
z10>H6o*WI{Y5sZj+0p>khzM4l;{f5!W1Rf~)97m?D&I@hBG4sN2Z&d<$V~*t;PCqx
zD*YF|F|WxkBs=tY`klgtSSb}snPLx41cHB?j-BpPrh>cgxsgVX*xbo>o#d-HC(g#D
z<&*|$lZu!KCmT!Kd!lO{h;5M7J;okBCq<uc1#8lj1#Oxhw8hECO6G^CcTyQcZB7g8
z4vL30!aq>3hD-0D>kD-~@?7kS1tb-{9T>pwB0iCLJ_Td&4rDZ8w>VF|HA-V|{6p~6
z#E46V6EO^hsqiN6x&4r<?x8^+740X{%=9xSUDK2PP4IH#{~>t0_h}ij7@0zJ84(R2
z#TVN`{Iko#u2G-B7A>gEHwKJ%q&-)sh?-~S1L=5+NWE~xJepBZl-=`hJrV_txDyYN
zWk6jj57&^)Ffe40dWG)?t@u={yuR{V7;+|YKa>Wigjq0^H@Vg|F6eOJ%KsmoaN&_O
z60r?2a~b2JE_36!W`@!KZ0I0Cd$-mBHEflbDG<2d0Q=<R?mbe@bFjta4`H=j*yo@o
zBcJ)+pB;9iEtGxRc=mSZcz)Oe&kHUW=F8qaTshvyq=Qzc4<TB@rv1+ztfomc+hH5L
z-wiGmQ`RN(&c@mUSZIsH7X+ntgLWzRxNPYH-o{H=3z9uCe3>>_3V_I#Z+Hr-#p@QZ
zu{qzUIAvlAJYm4BH^Gfu_jfR#?f}0zdt3C_O9;uEP6ln?xx-c7AAbbfE!rn~XLKLF
z@}7Ju5AuS(!CR{T%>T<NsE1w6(aetfE5r#hXrpp9opfd7GK|!!NB5}^XdCnLJv6Ar
zbn(l$_W<P_q<rRGplu->P`iHG=i*n-f0RPDfFMZ`g%{HA@$(XuLad7X-Fh3DNn7%^
ziw4@K>nsbu9Q%Sc!iB0}PpS>jl33TRkQaFPqA<*m3xmcJT~7CbG!4@b-Gd*Yo-<q;
z&}Y^hlv>|r@1xVvbQe~RwG`Dof?z6M4mb-2fODgaPuv|&1&tSG+0!DiIF^h4FY4>n
zg^fSJu6M3*35Cl+3q`LLm9D7zGlhKV?9QjKn)->$hnCw*XI!sG#H%aBV7V8~Q=o^t
z`&1#bKU8~xLwARaOGd7rEGTku7ZlmzI%ei`evcuJ@`!@_lEt`31OQ|=gV9jUAX;AW
z2QtlMO8N%)JpVvP*mC5J&egKsO>|ccl1zv<n-+GFiLB)ybR<Vs<H)1wvcL$fcz3~{
z?F%&iS@d;GA9xJ1T$e|2e*_5%@PE~cED{2o{-`)1M%JrxfTw6gDqMcR=Q<~@Mgsi1
z4A#>zZaJt%%e3M|BWRK(*KTQT{cmJ2scG8E+2-Or5i0PCd8Yf2!!DP%W0JGHOLG%t
zz*?m+b++1u%dYKEC$VYAW7X?Z^ciy{+%0FsZP}aYv%p-#EAgSG58C70y4%R>n3LdI
z#cRf7p~nFRox9XTs$8I{u3TVI_q4Z5P@Pxj!C?9;8#ft!4W>HsYrn)|EzvxFat1@E
zf4PrEed(cK5x*z?3F^yo0V@&m_{L3o8xc$L1t^+Rv>a1oloO!r)0ZN-ox2}>mmD#{
z>qm9B_~_$1KzU+dLTDv!lFGT5R$GvM2EkF|;0T@G;qc343jb_NPV{7xh$YT)Qh3ps
zS7=**KPl6f!OFxpuy|V3l~>^8Jl#Np$1pJaREEzmmGA!*Oop0`g^Y2eFd}oy-u3p1
z$-`ZpgON3oDwA+T_A-?RH^O7?@ZRf9_AMs5h5uO$hdg`t)qHVN0wTl~WY{Vv;}%?X
zkO>)~2*q84Bh2E&ferA)i5qqZ88}_jWUd<1dYO~OpZ{i(2&mCrX_=c*KD4kK?5IDu
zZ&Qgg2v*q<srXoP%ofy5n_=GDNJgahNOENnEca`|4(lW86VVb|o)Okm5MT-wJa8R|
z8%~IkMMhrLCy{1e+~3ru5|V$WS##PDepqt*uQn1?wVVXfZOv5eVAO+?`jI;dmr9Lb
zc~&IEULQ9CG$o(xp|b$3Gz=e%$2=uu-7w``!#aunY54SSWhA<hM~y9;N}PW<%%)@L
zVgcw5huUF}TR>q83q1T393$uu#}!Xwhr<k-oI2IO5#b^=#1sLi-<W2aj0vtvk4u1_
zPPn=ZF=d##b}(ejLHr2sx%)JnOa=4*O$=`@A@{Lc5zO2E+hPdVeHX*m|C<=Xbstoz
z{uM)-3hFPU4KDvlF=P<-?b`j{ilJh$(SK14QK$bqVrZ=6u2yx$Rg@gD`w6ia3kdsU
zP##u5NE`$w^Fuh_o=U5D<HrYh%)-7CW8r)P7Hu{vOmKeeNz`IeNclqHQq{9vCb#u3
zRUO@%GQ>&*-fi=MwE-E0V<ZAX#15$YaKOjQ{B?BV;Cr@>=WoPY=~*nKe5Em@ux$#i
z<QV<375wY!`~m0o;iFZYxKZ2Ux#Sp+`O&3YpGpMy0JM^D)9l0k_>BYg!lsHr%z*`h
zv?uE@N2d3~8%}Vt{nU<YTdnZ&VGoSd{&QrtOTTa)@uJVlRyl3!;OYHf&$9d4u)`%k
zFyM{kV3a!6S`xRjy~VP9vg#h$^yLAt2GM^iCiKLAm<q;l9r5?7z3J;@z&+^S)(UT+
zD|idDHNqH${B1%&H|xM10~q80K-FV$<ERV=3F~_MV+Q=^O2}?(xkbMJv>0B8`1Ew=
z{h!3pt<ULx#SWiOA<b7D`>7hIhSe|>lS>>wA&Xj_BgtVuRNJoTYp4K&ocw>TgDmjp
zbhb7Cq&lZ5@imP!54%?Qk_L9tkt1z4{mSbSARLeCZezWJq*WTOA`NFm7{`ID6$^H6
zof+Ita%Ws5>q-21AH$G>53TzPX_3UX1gh>7YY91aGFJ*~y?E6&W3u3Z&yBopG*lTq
zKnCrO%}Z74{cyEjoE3Lkk+$BT7)I|}7$X%FQ)$F+BOB=H>+Hb|u;m-`@fVo)64AJG
z`00&B!GV*tzt<R*Aw4OVwU=C6jdi!Wb=D%bkV32%CXJa`G?iB_6)(OGV=5|)mvz$8
zck3!3Lfcn6S5{PaAiif-M%yRe4TqhKmiO}2Fb_AFT+a$b?*NL!|0y=RLN)P07ut?w
zR9FF98%-&0kEDHo`V#9R{VS7d_{-thPPu^(0I1o`YCY3QLZm<I)a(~JrmG|zaCPXt
zQBCe0bNrgLeMdh!ymc6H@WeaZcrG|F3EDnrVOJbzX;v^-=1|)F$OEacjr~>FK9rxN
zC+Me%MqZ#Wfe=dCq`tvz)DbN{9L%o}PmjWzsOo3#LDN81Uf%nykqwHR3WCEDB{5CF
zwtiR3q+-Dai+KuZH<|NivDZ;1)XF6gr}A)U{(xEE(C?K2-|4NOe!k+=@0^lc+9`A=
zd7Gj}rH1Bx0+dQmhRh`D4^54}^!ZwV1D+h9uFRuIJBGvEkg6Z^K@cwikF;q!oQU!e
z7Iu$P`0d>V={(}sO=B`HtqWs(G+Kvp2>qgAyWgj112al`M|Cq}CaG{LRv8Hve1;v9
zkzsKP)Q&u2h?1cv*VxxEfrX1;>a}7d6lm;j*79U>qk!0*F8WjX6m)rMi~%4mx20)6
zQV;dBllv!s5c^m;!9E?HQ))>rrYf-E7q&~>N>OoRG)A}bQ5H0Nzezh@8iAeJR?K5r
zk8G%A!?ZAxfP}lBkYFO&_v@42{!5hM2z}w`%&omejMH+Il--ZSy6d^O!gpe#t2@rk
zlBG1wtT9`H0SUj|*<6|Kp2Y9_xp|DHEL!{Y<Ev>govw<%MPv*L91g45TlU#gl~47}
zQ1=lzCDTMC@NNkEv%_PMNIofIO$o|Q4rr~Vk;)<)BuU9Em$?!Tt6~)L2FAhu;^(yc
zT?*pgWc>5R?j^$b3%r-JzievoEG^tC6l(k8fePN0Rl?*lq+N3^bkeQ1n02w*t7viE
z)6P(0*P%xwrhx2V4-Zx1Go$I&V#H6#L{(Kf37XJ6Ma^vqce{Pem%F_31iz^ryMiHV
z<VwmTIuhrwDd9{$xhi-BTAr9*i#`We+JH7Q$VK>Pzt1-xu+m7FRR7cUqo?ML(V3^d
zmvgLEwNgWVsh7X+=bR?-WWaX<geX;B8Fv*ocx&;G&v4?$ns|$e{U1vjq?8ijxHTSL
z&321lH+*p>gk^$VWyK~TQGaeBUWz&12%<r3;Fx(*U|<eD?DEsXGbd&Ug6lh1B((7#
z5*<2Dw)`n#@GbV?)N*XUh9Vmq_RfhKjAN<r)0nUM(!R`c|A?uJX`OXdt5{z>llKT(
zO<j)n04p;SEGhWk|I&ru%%c)(NXxe_o}=(j>BaF74$G(lyLLg*^sgK?h4Ra9zR!#j
zlR&zMfhL|J6mng)#B9!SVfp!g!^?Shwx4%h2QJ*D53&!Fr8-uH2>HO&xA|Hvt&Zoa
z@&_zHzqwC_!eq`Hr5m{45uEELXztwF$E@UN{vH~oXiBbI86LEX9kJ>p`3<aaijcNp
zkBT<Wtbu$|iY=1HaIy$Vh(2(iz$7(+KU1H9eT??ZS#VCMDofd6h)GkHE|=290<FLJ
z5+e^tAB7!xjUk>mP&?1wYUCMxC3sX}5?rNcPyxF|YW!QQsKpgBQl76Yrs3sC);6f+
zm`QE+Z+&ClEQJ}Tg_&}+3Mhk{c!M+#<DXR_V~k2kf(pKwz~0o-E~bJIS%+A9?0}Zw
zZ-FuUyCSL)uAyq}k?L@Uj?L<mje{d2mbr5{%w#4%9qNiQDr(Tq&a6Ehd12kBM*nsZ
z`IeY>EDQrg1-<|6mNIOF^K2#hv))z{vHeE>g^UrMKcG1*$JKupS|3`WD5CX+!Nj9Z
z{D^-VX{4Oki&Atx_=U}&qjaI-yC>w6BO2y{^309dSJ{e19jG+pE64$PGyYFXD&LDq
ztz-%D6;OT#^}v}Y<o_i84aysOj|h=xS}{7uDSXK%LY>S$zTt7{=i!&9I=>lYT1DDe
zjsT6j`L~f6k&c){?CBwDhamhI$vr=eVUN+o)=elhj93ak<xYLp*i^|q<=WO{<Xhpj
z-!Vup6-~ttI<gNl9QsGWN19<9@M&(nr~BK&=b)si<j>tWDX?#aT$+@PZrX;67b5Ws
za10X7=4lu2W&gG0kf}xasvK!-!Ib;jCjW@>sc(ts>ll>8H*=Me&<{!{{3DR2jNMBl
zm$K%#=sy^gEtX_QCMnsiVO{tL8u<CztFSGd_ziOSh6$7Jx1&GH={KiQF-QZBwE)or
zZdF!1{jsZY+uWCgsAclf7G~962xXw{!6p1ELE9hSmlnbs`)yq{#!cR^IWp*1pN;4t
z@yGLqWU^Ckz!#qiKN(^3j91fI1dpG$M$<?a>s!@~5@ZzUCqVd^(Q#D+_ax*MSO+AJ
ztc&!w5uRy<Kh#;%$_&0(yjIU_C}hEsxS=ESM?7c?s#6IFk~gyobLsh`eC|<5^b#|6
z%u$LZXe`za@t6Y}(f%A{In`GdoXgzeDE_oG_#IA^y$w3AYxFqR`8!Fz6Qh^@Lqhso
z-AYK|p_w+lo4&TS0%?TW;odmeL(B5<+KTwv^-H7h`Ga6*_4a~h*1H)lC)xhTgmFEy
z!S^$LD~uNX1I~pI#~;P>5Ar(M-cYVDu(t8W4Cg2t{kX5NwpI`}AAGPjH_PBL$?Xz1
zPHebM?o7$7{}soDe;*wy86SfxZSm6n%S`X7+T4ld{E4T{f*a$HeT!CNiMD?pzNg_Y
z72`jyu{u(a|0CadH<&r(vUy}L8$9=P_8aNAS83;nd6eA0_+%0JTFSHZo9A)Ow)P^2
z9PqCTtmcs0^pVRB8vd4Xi<~CR={2|5n<vdz*63+O;Jy}H=YIz;jDLU^=pW#fZsGij
zvC#Cli2EDT{SAn268rxFUIzc8h{O6@#GRr24e2<6^MIFKyLa$%@!@_C=}P{39dpFr
zx=z{PfN2@Y5WDL;0rie$?2r8xTT5>=yX1lfLYybUn(I2_DbeaW@w3^;xOhzvQM&3f
z!D?bgEup;$eXjBlv2k@v6=iuiarU0`!9c%At*q&UicHY3`7&m?)n*BcP9OMjs&+g+
z?=#<NY}<~T={gC)X#8ZGY`KTR2$M%HS`%1Uo7Ld0+4yv<y@*sh8!U#q9d?x8WIfql
zL~Ezq%uHvYpk3wM!9N)hC66x(tbIU57Wkpc4b(FK+bs&R-5i-98P1ixCyYi11EiT;
z{mFnAioh)#d#E<7O)H8}P0;8g=JFZZcfJ_!y_zEB=*8bPKY!YYSXQ-<l6lVMX6+N6
z$RjO$nClY!#Ap|fa}@+aXSDVE1AFLh%J>Cs*8RH}RAyG*+Lm^2PIo`(B5|V6M-R_M
zEzGNxBxZ$v?g}Z6SgmS(@w{ZiT6q4I+Vl>p*Aq!lTFP$5l%2=*1hdi2FU>l;4O8(~
zDb(Nua8Xh20~prqN#@a@-bX7Gjd0tqDGLroRexWox@K`oP$yC#&dEivIhuNsPNN!q
zx?y)W+0wegt9Itcc<RZx{*&tYh|#r7efQq1TY(`UcPgRr%&PK<fL&;lK(iUUODEio
z7?x|3?R^ihnvvuK@a96tPQ|-#0)v@EaYSE`A1q@&N_e(PxqMAs&{j_&{GijF3q*0h
z(|W9W@Pcz(*I)L!@hm<+IQiq;O>b=n>r@32J}0PoBPBDH_A+6Ic|n>xQ}`_(<D58|
zt8~}cwZtl{P3xRL5d9LSOZk#PReWhYClCO{?kK#W&tLrxR_4N19XckJk5i?~8DJyW
z3bkV#Yp&4;TGDAIt*g#0WOS;YFBFyvyYFQvGz}8VIF3`)Uj^Gc+B`7b!r0b^YK~N0
zlh?lxY}c0^G?@oxpo<<uh-&^7-1q-^-2Mr(XN|d_HDR9hEPx7q!DoC22!jX~@zf2s
zC4L~LxI-mTtwc{F$$fc*+bU~6*gnIVJdR<lC+Pj?znrPi!wYsv=j~<>{IH4I3VjX3
zSg^_4Q*XxTnJo3_tHFKkAVO@Hu<uFX=u}{mo`@DRGitxoM(jp(iet07oV-UtoVPco
zyf|osM?hky1y%Wg#iC_TQ55|71A(>P$%$V5R3Cb1ecZ7x@pivBtL+X9y&YAk0y3*R
zRLy07@gIST)_f_QUBg$W8#8D69wUQVeVo0d<rNpGNPDdIRFw&rR1oskeSFWWx&cYk
zZejn&0kiVU6kn=vw4yR3*OehJUA@#!I5(1T>OtOntePYa-u?k}M4|=)uq9xDeOnHE
zf+B8O(S1@qqjB4~#LE|1`bALZda{PBSyAq(-9)iX!n9ZZ>Oko?aU>RSoy6-3(rfBG
zOJ4pPD94mhUUYbfy__Cx!D(}`AU2qJLewUx->6uJA~`}Wfx)YkYa+?Algl643y_Wo
zWa6kr`%x3t55BIRqyt<}_zdS$Ww5pMktbTsg@RZhE7@XLzQMwR(Qum<4Wy9<&~g)&
z`W15ZXGyQ4zX}%E6?<Zb(}pOtB%-)(b@Qdo`)cA4KnNN=P27q>du1D@3bDEMKI|o5
z4ZaymC(hS;B1UpbdCXbq6r!#rW7-z|M5mdl1lcO#H%<$`dGsb=-V_ZKeWp6<Cgb>|
z?Y9`S-ur}8@60pxMjhvHa2{YUL*kTJ_vsXQX!73KjIlBqpt+Ds6nBBvDCBm619l)W
zU?Mt3wJa4%%A%54;9BUB?Nz;=^s_WGw&>QVyu-=Z6;VKF=1(oEu@c<-RQBVmqM_gK
z-Io3MVu_rj2#<RY&dE7dywi^@{0W3VoOb|a-S%o{nyu86fz{4uEi<_$M#ud$PMF7H
zMIe0FKAdVmF9ugNU~mqYL73}fm`!|7<6<MskVyFsZmvrU20(3tXWPg*J$0`0x|zr#
znSjy?fA>Kf{4pVYrKneGq2tmOnjD;PvK0FH@#}n?#K+IA2?vnJnd@-a;U?S`B9@DY
z=HGB>AI_@^gUVb3$bpI6%-6p+N(0W>##oV7hj=?e%TA^&F&kLt5SNqq%V=}l2rujC
z5A+VzP>v0(Q>v$|!aY+E+=AVQjE2aGM#y5K=F1UZiOqR-+_DhUs|7iZDUxr^n6Q2D
z-#>>A;RYd5MmWtK5tkP#jg)mF(|rQERrMhPVV9#W!{ikYiskikBfprX!I*J5fldSD
zf3DqVbTI{(L7Im5wRQ!dr1%d_0SLx4o#dJ_>(a4CT1f5IC!e<FE!RhWe1)frI6T2!
zaQOMupzl#GWdksXEDU+6h`_L$m_`9AY0{tvUoThjKwor#o|ve%nElM3*V=waSDSPO
z*ImHNn5}ISEVatx10Rx#t#Ky;DpVDZ1l_xF8f=KiFYCyaoj4r68n`Ps?hUqz8><4_
z*Y&|~nDe|T*jFK54csE%VrvO^k|Hl&D^v9Q%)`Gl^MVg>NO-R81P5u1e-ixxrzYY4
zlz&7SVJ+8`iKGqDcS`<haxSbS=3p)KA0sx}3XjmTpUyry_Bp%-eP5jxgZqtci6ZmT
zuU!oWmmJ1Ay23qifh!{K5gSnSpAp+1NG?`F_)OGC5iX5hu|et(JuNcO$0vlj@t@4h
zbiH3+fF$fRsta!$%fPq^9`K*Cg}T_+0qx%(Iz)W_g!R&0sVTR|eL`aE3Vy#y7zX!I
zyZ(xb3G<e~d<ySB?nOSUt6jvwE7STFH@wL(uzzD2bN{SxY1!%(cND}oY6}uVv3_M=
z*(JDtHi<r%LU$V{``PNf=yr33sd!WvJtI#}YSy=pK7Y=wHFqt6a|sBpP|SGAGS<69
zJ~ZGL+IfwwySm#~Cok#h9WoJ{(b96T8|^FU>jURBxh|~>wl$re05aAe+us+E1s;t!
z8JDLD+q^sH(Cm~Bl<!v(d@Ej{bu;s=@!y9H8&2p7(;d`Kz_zUXyd9hi71P(6&ZEM(
zF8Ii%%Yuv<J^JpCtqu;-xZa)C0g(H5)47}DpBp;Bi-L?+&EtuMJ&7ltCmC{R3)*SK
zj^j3|uM1b4jknYLZHsB9%}=d94$FMJ0pL@Be(qC{EA@+2J>Qd^%rUHE%<CcRiq%^b
z>AwS*?7skv<1c{8Rk{Qx-dnuc>^E>R1l?T!!tvtvu5kGf)`6-=lE~<CxKEpQR?&kB
z>9GL(3d*+g=n++O;pX=jOT-Bhjq4n{fJa?k#qS;kbl`PY^pL(T2?_Yco@6S1SrMCt
z4F$iqa(M5pT+oxCP&YUe^}_K)LXH32dV9!Yd=Y_QdYk-sFTpEqo%g%eeeHx<wj@&h
zN!P24*9k3C($J#7+RxoXFsV-Y8$CmYsOasX+&*$e&9pGYsbjsJl?$Pd6Uf>c&7<(O
z44f+<AaFRH*ae|G&9kXnz&DQGeK^V5@;qsWJI-@@PDGCQ5*SwCjr5&@Pr@Oy01il#
z^4dkpxRE8a?VAz;wQ9g1eh)GIz=+($h+OF<XlLEsbAev)O@rUcXlvs*c*wub%l+OD
z;o9~I=6F=-yvt||mO_<yfw-F8+k7I|ItdvKJ)1n{WfIl=$K8kB76*6%Z(O;CYj(>u
z0YhAF?j<A@exbw785xaUYIFVZNwN7!v30Y)7CDbRcGgLkWmdrDY6G=(<ytkvW!tK?
z+g1y!*~3=LJR+!WshqYQG#Wx00hP{*!)vfoH^NYc2Nzmik`N?WgHVzZlI=F2l*=ys
zro2;r(ELWi2WGzSh<xj5@(zDs^XXy*x~031lUb|<)&G2P4{I7l(5UWUSqAehKhm)g
zUav2vPqxw*agl6Ykw3khCpB5uJVj)_jU%JozpY=kL}YTX`Je+{_WOM=eW>l@-(+fQ
z22(0gjLLt>V@ss9+%Ng4>c8!8cu&_4EYUrD*%rTYd(9(x!#xsXpA7BDGZ&BM92EgZ
z%e9Az5oIo<EJ~a+7t#7KIhIE9Dr;HW7d-R3o~dDMzNM1qh~L9q6(McJuM`d{Y`r~!
zt_UJUzyG1-*AcmYLpZyNsS0sL^hdmYlc+rsLOjjor3>k$1kHe0bbJ$je`5;xPy=CA
zk7GeswcfzfGh8a9%ER$WCP_AP*GTcYu7Po#@YSHZMqYS^T3I{T#J2y%*O`3v_W+}_
zLYn=#{MI1oCj-+0_I8=a>>&7?@ovNyEupP=)WnlJcp$lK>C{gtzz+kc5}RK@<4ayi
zcb!d&qL|#1@Q~wOtv;H;`mdSO;VzFszz7=pus7v7;W{~>%{Ltj9P!fNzTmr+#i!5L
zwPL@$N$i!AV4jf1BSkWHKb$+Xdw~T=CLLL&#XGW{-b^R15w@!O=NEV{7QB(RViNHc
zIdNxm7n=rQr(4L}2R+X(B1P5^Dli)e(C)9f$NFk2UN3yOa=f!b2b^}a+|tX{Oc=k7
zEaf}G7@rOw8@(lg9Wh)k-(RFJS^A8R|3YmZ6pnI=(wfZ+%H<!)v~x}ffA+abvA=z_
z>WfDSFrj}0MNcff^*%b>uBzM-sOBkVdo+7Fgf;c)qPoA|i@q5f^U6}VI%aIB8=R(9
zI@<(<WT?de3n@Se={u5IHqG-yYb$3?cccmnFEC(keElu3VQ@%|Sw~9`nB<?kvgA_h
zXa5+LtqCY;Dd(x$F`N{h@8IOHYDNp;5;gER|DOzWP^_oJj3i|RVS3U+t^74qSUO-}
zKN>dx#Bz%4PvKSnO7dlk>~dwX-_Y`ZK~UO(G4wS_lg2PU6xND`pU51VjvvS6_D!@-
z0N0q`5L?(g!)*$KE*#yHj7|X4!+LMe_3sQCcXO6DwJ0Ct7){Pg!iZX*9X~KZM^%RT
z@?77C1DjUhp}&<=DEipA%x<dJt*-%}?4v^J4>ETJ&D=pTb2>y<5B$f;*j!RS(7mq&
z-r7)iKba*J{(`BLB|mduM8eOAc`{3xTn^r5v5Q7%ko4cN*mnv|L49IgR2HA!jAGWy
z>Z;VP_xq=Vno1s)QT$hTVugj7Am{=g8<I*yV&ndUgBGY(tv4GzSrLd60&=}LCJd~t
z*Z*tbB(dR_`>*vI6=+E;ZmIsURXG4d2Kd77$h3lMnrVNMw{O0Nd|#kHCzpEsqp+lu
z)2rUfFD~6+q5gLp_I)7MWh&Q8C;IA8$6efg#iRj&D)nzRnfOD|V@EP+4l+sbqIEwf
ztzQOrrG`#}>tEUA`*->imbRw#_$^u+N=-<BQXfp295UZf){lc6ejo$%FO34YSqbqF
zAN!rrgZo`>!jVM=KmB4586^9~fO!9<v?>_Q8mb4oJ_k9F%vvyls@wr<R0%_wE6P1p
z)GWN(&jf#1anr$wCBm7N<ugwbUy^^}zGRqVPhpl$$cF%#b!FuKy(C-9(3$Y&9%G&y
zThr1Wf(N4dYHIV^P|}6Xhafr1R_GCwN@Jw{*)@6v3pYR=IJ|TRgOxrk`crd9)tMxD
zslqRY)Q?lZ82ymEOmRFn^_OXvu&iJFcwIxA^&cWLK=y6F!meh*i$U*?Ps-!GN!M@!
zPf_XIyPuAXH$Q32mtOrgwtyxjFz=igA2uI_?fv~&$W?2Pmj0*S|M#kzfgaI~!iSFn
zQ-eKZ6AfL4Gis<gQBrn>l|p;Xrp=pnuAR_0N^~aO@rbCfzASbcf2?toenU)C8nEf<
z@(!UUM<gGFdZon03!PyOL|hCXhqWt1a0#1el}$V)(VBY+8;EYYe8`N3B5S|l>xM$t
zg-s(F7H!4EdNv7CX6w!RjA-+<yYVo7pD(QNLOZgGr+yF8e#O-Bt4Ie&j55Hm&LoJO
z^6-bdxEl-9-2xZ9?9oNv#0+)7OUoA`6kYt*v#g=}Ut^kSxSOn{|Mxjqjs{vH#?+vS
zKW}-+5~&yV+&@r(e>M>%j$n1g+c|Qe*RFn$FnY1FjKDNsa^r57g&w|<X(fNH@jXfo
zl)42(QoymH4EQp_L<YoP=}jcAZ?uBehi|q-zTi0I&PwQ;z={J#SJcSK^+|BPpyq&s
z7!-MP{7602f^7rFyLtZ82#R*}h!XY_vC<IVYUWfj`Z<2-%Qz;#lrAMEY+vQ_{`ce_
zT=$`ckS%DA-wTO#NfD&q(<)hqE1?C3a=Oy^R8K-hDHTsPOV1s}Vy_k7pH58-p>6fT
z6}!;WQZvM4tq)*AtxFoHSou5lTr+6Q^?B7Aj<-H}-t$-FEe&bUDffH8AhSZN;H%?}
zQ&OOuExHC`okcX^N@OcWqUc#ly4G`NF^bG==Zg~$%vcQJ;I||9CK^uu`E*aFHTdeQ
zNtDDPmp#gYFYx>8W%?4zKE9PZi}?S4wqWz&o>~WhD>(mE1{FAa--69y?q=9-e*ysC
z*JIDeB#6$B#EX?DzXgcyz<-TU{RlHed*z8t0-IqLO#o~5%wa1Y7BMUOyU-AT^^xl8
zG<`~R_-gx8E`I;P!A2Ofe{68Od0vGgwv}CY7|gl2bh2Mhky#*o*upB-!{-+6PII9-
zC0hCf;zDfd5_=LTxiAuf^E_+0S&BtRv)`4mfoSnnF~!(-ao^sLBs~px_qQO+<LvFS
z_JfJNSPzjNr2ec@>msygksO)sFl0Xv!z)vMS3>^I`hMo$0HpmGkQ%aJ+m&j>pZBK;
zHaMGZptxQL@cny0@%x-jrWLVTotsAaY1E52ACsbZ`ou04Ed_1$xauEVybR;zOxHpt
zj*m+41C7R!aoBP(a22&eX5##ek^qwYjmm6%X5`F^`9sMd!LiL&^R$xx;uh@Ttm}Q}
zfXryke{l-d7Sj2i&iTKkpaM^*4sH&C9IWAbbgWhlTklWul4(bHiUc38W_pXO1+RWJ
zut4?QeolcPy;(%zvLaI{1l*%>-y7#@>J2uii`yXe9!6sPm~+8y%GZBD%bz=orPhTl
zJm-<Sx#vXNPtv-Y4+~2qgGSWS(Jua!O%o$!6+wR3mo9&^9^0&eC>1-9OqI#fnNU>X
zxxgp|=XJ*2M@EsiBY{zYJ-(Tb1uj4^B)>pfyV&RHlyPfnR)k$IPO+Q{LyLh{C_aoM
z;lWWbTgzMULZ>*T_Rr%#Acl+HgF{YO+c2<Za@fZ_u1VdA$!}&fIwgc|MIYrH-btF0
z@NJ#ao2nTWwPp9)5Mt3-?;dRD2vloR&S~cFoE*NBr<cElgRAvL)feftpbN#Zg~w72
zX3ojIbA$iTAhAszo%Cgld0$f4wdqFfya8rU{AhTvBkjoumxLnvj~bs)Cq@9dq9F@8
zc|9?KM?vM5y4E$U%wa`8L~LpNPFa7C`^%ZL&l2Cs+DiF~XuikfL?nSDo?wWQxuWBU
z`cZ<pEL)MJWfgAi0lozYm*1Dn(_ZVIKVYP@)zuRop5L|Nk}y~NvN3dRjTzb1^$T0t
zv#-~>e<7geoG&`a2gr4(kLmM@Xz2m5T-%ZftZ+rFLU_H+!A26S7~vhbIpbXGU}zZh
zSf*?eYfy?!m9%HKXo`<srb=A)6?%e?p{=0&Z&Z_QT>qzPQi@+f;r(HNAILbDnO|B#
z#gy7>G(Yv5tW1zMTE~#7c|5jO|3!WH(R|80rZ(jtIb~4rFJ@d^MTGf(Tu_E1S)wdm
z5?)K}V0L;HQbZEHo-;ONz(g`jY|`YR6l;`fxL9taE}pP%tL|*;c*}-&eCHn`gDYGS
zJFgNO3$(v7=ZTM=*7&$0teah^Jp|#_DE4>Gb=HuJDmlU;S-Hx*gX;iAr&<~5pT7`6
z(n3+XM5=NXE$A4E#7E!)HTR#DsC>+|0m?G`419*<s+dA+mpLS>AI*&Tns-7?n+Ca4
z#$kYwDyaRq(xTTpxk^{%WS+MD&Z;>G;e^iokt3&I-m(Hhj@|9&b|DAxN4k>YnW*}L
zi14DoPWnSg&XK8issK*yybC0(xemgH2}UQnX<su(un#gJ>lE_5Y|C5KIM=}E`dxB%
zy^PC%1en{zuTAkEmUcdLC7=s_*he&p8`3x^S~jZBdQmteEVC_{bWXi2OCYi);<>7l
zo68;z7j<e9t2R8-y-v*kDBmdZ@%S9({Lc_+f#Y68Wk5mKTk)t0g0&8=I#=sAN-zXc
zOgRCyC+hsMry|=@ibccs!Mx(y3>WV4YIVMiqo|JS4;XPNv%kqcVDpVvf6xf41Oq)E
zro^D}q-Rr{r|lhw!c)6eBl{W0-{BTUlW08YAo5C_>eY28A~uPs5McykfEO3IW2a-O
z8F^{=d6!%6IZ*U?dP*b3Qdi`7o4eXsyc7+v5XZ$uUd{^0n7dx|fr;3c`s}k9`lX-U
zjO}|7Rht0+AgzSD*jSC_Fu~R}pAE&Z;%~KO%bLK#;v}O^-IUOXlW3l)?@E8{Vwwqm
zFStVbNPYZLriqAY|CZ9h!0C!r)CJUmhLSh%PLxlO1dRt<%S6d8{OB3AoiWml#!si2
z&4aZzSKZ2}^HP`k#3s|C77w~D<d&ri=DM2cnvh|JFtfdWELxobnrTaCs<j&xIx)Iz
znYnAsOJ%4fYAQ*Pt295<NW7kb-{*E(+6Ei<pq00?qFv2sRYPnP)O~T+U>s0QeK@e6
zH{V)bNLwLTT~W5!6>`GmyKU_{n0$T*ht0xn%?A2J>8w0afZ2MWoS{dTdG^l(vUA(F
zp0RwQ54Q*DRmC1AV4A$28-=<ptn&?=ACMyTz=C7-y|@f)ee!0T2A<2S7UAf9PhM-=
zc8Oi+#wWojo}IHs+YK`>yY#=%1{8C_)b5-mzg9l7!o3QHwuRpSBAOc$n24M(TKD)P
zi3K)&BOePcE08I5TfuD3f&a^FYi`9TM}8unP(O8{_HJ4i=y}M*6~pVnE;mT{kfMR;
zs*OL?Y#m)t<2Y+_7x<>Rop6X|>2y(7c_Fn&vf$DfgWlS$KA_6&VJRP^>#wu6(mS7D
ze=)D*9^oAP|IKFWCVgW?W%z}_eol#!<2>YCOjZ0{+PH-diI}S^eRr?K3MKaw;n<=;
zO0eetAY_JZrJC!bn(Gh8*1CV^S>I_nlap_mlW(2SSqEw1fuI6TvEnbvreq751Uj1s
zw!0t`4N3faY{@Bl{9gMJ2I}l1zhLCQ#b|Fk^%U_m|D|Oii^zKO$o9_w#?}LP;J+?N
z<rJ;J6QzH&diP@o=g~#!eM>Jf?w0aaO=y!KuS=%Ck>%{)q1eRUvA?0D!Fyj>{<j0|
zfd0G6dWQCwoBtI`y1f1EEC0=n1OB@4Q_nj$YmZKH0)5;7?>}pm<Zoa350sVuYtYuS
zc+DeVNOiG`xPAnlWr~FdCP3=Cs8+$@X<*-L|A}W=@=cUwS+qwTxzX;N0dRk|Ui>bf
z`ZRE5aoAM3Iw<+X+A@Sb7g*Fp^jKB((F&8G7JE9odI|i?Li3(3T(zpLo#WJj@^Dn{
z{be>c;-$MFaR(zuo7iKWnd*_X1JQD?Z>oP5I?B~sYM9L>RExYO`|fkx&B`@vna!oN
z;?Kf{HI!6KSDED<$#=f@PoP`1Ly4j-V2`0E{!Q!JLC9^7i;&4q@0Fbv{-fPsVT2tO
zP$(q3F<|8jvzRMk$kv=e&1-M{Q_oCfUgiUXxw2vXk>!t#)Q_5_HZ|~>w=@S2uus#F
zT)s|~XcE#|`2o+H249iVKGg^}6Kfz%h}NwsO?Fb#6gRlyRuxrVl0Q>AUh3lY6V2(9
zoCi_dh1NYo=EJF+xPYJN8Z{^C-Vl~vKCrI)jmIbgIG1EliJzdYGd>k+nx6l{B0c`V
zcB4=qhw0lUyd)U=nsiE%xYg6&_=D*{xF>$$0d!(nUkge!Kr>oR2U+h()`lin_TE^z
zxlyGqTaYkvE?iSR)hxWuKG`qw$Cg#PIc3n|4;0l>mU|GMwD<p!Z|qO77OTWH$O!r=
z+bFY{TWC1~*NHCBs+tzB%Uatz5;Rktga#6b)wmAya2WzWhnQ6`N`7lkyf{$sUlU05
zw)9?ySZ!EvCK1p*;^XD3NxulyOcT~VF_>uSKk*5MUs7COlo5XOtMiCEip^U!x}ljB
zu&#8L`KB7GeQ0O+wxufBQ7CENn(QRMsnB;$`yyp_7v~i(+oycxK8U#+A9_#~9@wa_
z6;^Xrx`3WKjI(za&sk4>G0O+0wTQ}SGErixhQnNC4=Ll%0jz4aOM5{8%C_e0Po9IV
zI{F)}=-taxvQ2s?+&d>!{yBae!0JfQk#L*s@@?!hP0(!Up=`fary6<qB$aCh*ZAbG
zNq{9B;}pvoT)TyVj!Sq|TUy~NJZ0G3MYdgJa>>Y{PUg@0d7-u@FYiVu<}Od+Ftf(7
zN-GT{-GUgwES$0Dv?yTOE)NNUcKV_E{;ofTQ%#z41yBdr(x-XnT)-0oSKQ>F%ymT=
z*bfBEf3OsUK~4l9&~2gk25RE*c_fuN+WpF5+~v(0Xsl8|2Do)erG@BYiH5IrTDIEJ
zIgHcW6vwFYB>NC~rK}FyF5;CX1@P=~-L3lO0s_tpuy0oP$o;v!cPoA77J$|d#jL${
zDHg3%SqR%u=Eso4F81rH3se`vYjE2L)u@7f3{(-3mv+BI2Fd~j2;sng8e;`K{Lz<6
zSXvnBR0PWF$0Et;qd@bXH>JVlantSR$x6%^1edY=yz#$-$onSmfSIjdGH>_}rZH!7
zX*5QrqwKH3kB6b0wO#8EDYlSqCYH19)PLhj8j~kbN|ThJa8MCej(!p*0Ty;gBcYr^
zEu7?3y|h}^A`bB1WZD;p6#lWk95hO;+aeGDP#R8w9QX6*OOKc}cwwshvZpA&-fpHy
z&1uk+A{u0-MW+j6Af-QZW(1Ydc!9dlxsebc?km42=;(zmtV^@F2a+MSYj-Q)zv@_=
z+T~=-^fQQ<%5X}ZFZOrYg9e!YfJWpzAqObKo;gmu5m)U(3mmZ?MT9h0!lOte{;1#i
zeAymQrX6|nhOi5G^vDKG16Pe0CHO?`Ioh=~kSwVNaKNT;l+RFS^299xFDnB{(gV#<
zgvV%9MF^PnC}3O6%URFbPns!<DU~=UZ2=}t-<15A#;J}hQ$aiu^mx%&$yw$PXKqGx
zF*L^8#IHKCwYS*fe7#JdUDzZd9$-l&AmIt;T(UO1H2sO`_CS5HnN@1F*B$-H09L0e
zfB~-RR|ErG^=<Ti{{e}55uD)gs^-TLBT05m%nUB@g2FtWP+G2&TEeIYV;wo&chckI
zEaEH8z-tg5cHZ02I4}L@&?1=J;JxA7YtQ1o{8y9;%I40;{X1cpZTzv6kDB#<1T$h%
zuO26ddh|M*gOikFBdv7Y7Mm^M4u3>ks>dqy>03CKd#%Fjy1A(2EkaG##63?D-PL$i
zdLENf3rZac$uo^bh;wx8)vqK5s)H$Z>Pq}Nf;3Rwh>JuDOVLJoC5^pOr2BNmzm44i
zw$;cClLO9d4Bj^ktR>4P9PP$cORe<B^*yiCuLdpJ<4Ve=*)~UFUZC@f^j0t7uRV!}
z=f94RY~$byTBRM58&iI<*7L(hycoopzMC?{pZBXE<O{y%T)tz&L-&(1;C+GtYi+##
zok*P`97x(W(|@zb{R+5o>_MN!k%HcOcD6)?cs}254-2^u%OdD-VP1$i!DrED<cIv8
za@QEhb^g6yLD2a3GK}NqaeQ1eD5J9PG<Cqi^Hf5aXcRD+q=S{t!&VW!I2(Q5ko5A_
zTNh~@;HkgP82NFA;uMyrfV@?xv_P$jfIO-^QlPmHIOds##ulpn1Fq(UYENt<*1N|(
z{1_+JKq^N~n`&~=<D>Fq+VZ$G*J;3|9Zt%u(aqQ4;|zU^tkulKw4z8eRt0UY96NQd
zAab*P{x!U6aN4yE3;B8<dc!vjw4TWc586zsSSidJU)evf-G%<`Q4u(H=vR{Zc!`7L
zz21Je4jUt3h^z{4lLN@Z*kGZ1Fgm4Uq<{mx^J(6jT_Rm)#RSPUSyd>%N<1^=pq~s&
z&ZQkRzWyZwF=^gT4#`xnVD2R^tG@@MRXvP2JuYd!J@?2Zk3cGgppOzip$$p?<E9+@
z%C@|J>Kn346RhByIz_YN*R>jnPZhe(#BI;N;@-2$+i3|2jZ3%QTRzmpTmBd^ez>iY
zI1{0w_W@kd$M*r;rZ>Z~=NGrRL30+nD|K{E-1&D}zBIS1YBOcZ05{0nSo84(?KiYn
zY<8G|ywzn#>Tk>{NGfkd0od-?f(?`#7D8R~?{h=2tFe=eEyHH}0Gt*9Eo9J`4GQ@9
z<~c#zVWLkP7HX+pK5W182YB1<ZLi+t=B>fSYSe05MU6TAP!KG1Q_Q}(B?aj<9TBQ8
zwG~=RzAsyyZ&k2rZBnysCBEyk?tn_8o<*rsZI^1rFpfut;Qay3_uR8l0m~tYfmH&N
zSL!KBl{?9x7mFDm-hn&L7Jld|CzFb4kR`y|M*{eQ4uhX%NxBDrp81K2el2gkZ9NFF
z`wtP5R@Bx@tZQgjD)QC==Pv`7B=VJ1%P;cLd2ug1dsZ=&u1rJataweIh7Rn47LjcT
z(bzo`9<27t6{;()laaK}9lm-)OW)!mn(bOug|c&OIdb_)_|Ui9EwAZ%tWT|1-jUpc
zINaf8Z{c#j9N<B0!V^e|i#V>=tR?(re~DsYbb6DRw{u5Auir8CE)ajVYT`L7#26=S
z&4+yB(@KZzipaufxkgBP%lM{m%5erj<$rel!?*<^UBg<rYHWa(!r#~sXQq#tecN{^
zGgex^qQn-8<4AZipD7&vSP|^{I)&sL{u&@ufQ}~E*mn+fuvJiKhIO6nzR)97R0#va
zhii*lDfK}w{4yKn$Pe;jY5U=0PiyA4f~f|)(#l6+3qmc?B?0(iJybX^UcRI0NmFp8
z{FY~f$xo&$v7>f|!rkPqoM%O3)ftx1ML<mEDDanhUMvr5zfk0h%TALI)RUaa%)*q}
z7G&$@_>MZ>DgR=e!`!_frR;GpR$1IGhn$^iIg=uXt{d=J&(0b}^$#M8MvQQK*bYvL
z$y5gnTWiV>wt4A!#Sy!o&|5wyO`l;d$s`AP+2>PCM7QE5IhH?ClIlgGTe}7%4j@oO
zRa{N+ituim9$WT^Zd>+nQsCZ=;C;0Jxp%w`p6~MW<EN4f#q}Ca?dg!&49aDW?1me`
z+Ae4cYZ)ApE39#zkJ{igbU<!CV6kv<Kf9YE;V#V1ECGIWK0;UWNw2l?$$B*I-9JUs
zQgrQTc>S{abgT0S#r*Q~ZM8+e?X}Y$Rm0Gr{tis+If3XI_45Kp#5%0cUaH$2FE>+~
zy^$dw6$?$2sGIfcv<P${zMf@tB>ntA{IT7x`-yJmF8#On-Q6XK9&QCz13=qGI07QK
z0|av}&FJQ!Hs)E1?Pqvy1mxOKNwNKx_E_%cmseD@zNa=nTZpSgtQSw2^q<e>qU~m{
zU8iDi*`5a7<9#e+nQ7A3mJJ{ZZu1!E=`_D%9=ERf#gEe+8e^=VN9f^os(gd`J`G*S
zun}*P2b19|z^)ta>`l}5$=Ccul0x_pd9GEyw3FpD)$Yrk8NHjfBm}=%qjD5Oa*2Np
zl>^X_U4v8joGu$lXIU3~=tRb3%Z#ykCO|gP;*sjdrY_pfqI66)zNWCxn(0S1R5(GC
zqg}h5n)BR_x(JSvQfLissrbIKIpTDROLKY=c_6mZ7Tkhsch?unS9|K?)p#ei);8Eu
zA->kO)qTh}l-|B<UwHm!ue2TW$lGPY#pSq+vBM@=`RXG5`gTv@F8|@!0Ja&v&PFOb
zVM2lfb>q!Z3)RCuGU0v`97_`Nxt!3zGsO{8u34CcXYWFu!!i`}Itazf8teJ8&XgpQ
zr~UXr=;A&>Cp#n!Pu@C<ZQGa4Pk!?W8W^zo^bw^dR<pR1FkK7Ll48Rl*sgHSB1F1<
zdx%k7FB2tlo;03vj~pK%(;WFRYGXI^`^9Utb2I6sk9_~DtD_X?DOFAzxZEZ^yMB9E
z?{L^E<DD^^*I|XdBeZ-*_(<R)#e@Da5U>Ir@?np|_xzFP`v!0$H|^MDY&$bU9K@W2
z@XLsQXKdSn5H`AAL~Lxk#|p|6{_vC6f;@YQ6_FuAx_mYz#PG&Yiy*DPzJVd@xOft@
zWAr-2;0BSbEXw0K;EDLW7cbtUZ&N1<^+G{UGdPwHRjt)ZkV&*sWf7v3lmJOU%#5Kr
zCU7Gpe#$6N5NsSihiuEO=O~-^YTDkH#v*qnSf#Z6l5_SNOKk?dW#TIc5vUz_Xo%A2
zgV6fQ5dWeu`}+$W71{S(bJEaSw(8s8;J)Zq?il8oXyj~sCZOEHbg$3*n=k!>%6OH7
z8k&*5-Pu?RGn@VBWp*@9H?G1cRYKXfXc7MP3OuHSyb3mA?_V;oZ;#8ifC}R<LZNrg
zOWIO}kL|U>bDo?ZaD^{?XvBCO5ku9TZ!4*5lI$4>M}Hb`{%II_pBeHTRlVGY;H~fr
z@p4H$?H9M8=F+jFO<QxNjJTwLb6ik*e+Mq_!1O*oTrI0RSaTPCF^sjmKe&W-%RHQy
zR$mUlgXTYh*f=!Hc>}2-tzD|f%_dBP<uA8u>*oJytmi~#$CLjniX(*#4dD#rXGr&c
zU;2{bZ|93_wrHV23Y_VcGZ<tn{8L(=DWT+szYeo)$Gk_x<=GD_TRpZozZ4fGl)p~~
z6-EnV5m3qQ8-D#1ft>xYPMk@t<ojJxB3&@jZ*OJPJgz(rJt;tHzk1bh{_2vx%S=RF
zmh<E5brIDjJ_~uktbAF*oO*-{Lxwq95O37%EX3Y?ZBeYLa^de=MSt@stA(PB(DpEt
zLUcy{Fe=Gju}KJ&(0G4g&7D#I$)7*wOr3lcmh9)qzV`<~NV8giXafB=HM7)6C=p9>
zo<HP54RI^^8Yk7E89<R{t?WSbvxUtTpc`*@eW6rQByhG}{kEEiEXMvP9jyndEP{<v
z2D-LfIOcQD$`cFOgo>techZ+7c(o#QqvPDcFaA(S(n!l+5X`KQrcXTa4xqozB|)XK
zenG*uVWJ)J1Q6LCOwem?U3~HzgA2|V2k7*Ld^I&R1291%8Lj0&LfUz7>wRj}3ql;+
zWX%q!twg&&MU0G;v;E{DLpQ0VWebFXt1V)KOQZ3nLTwv2259&39Lo@fX=mnUj0o1C
z%WHRD|39?-XIN8D^fn3;A_hbg6ctM#SW!el5D7IwK~O~OfK&kiMT&F?AqgE6X(~mU
zV3&^clF)ls0#cHI6ipz3Kmw$_@%Nnb{=Yn5o^!7Afh&85nOXO&d)Atnl|6f9@A=#p
zk@;;ucjfna-ST5K`fqnv?&=u*34E_1FFMuiaLc{z!by9A>n)#M#ZpfzKFD_%LjCR<
zy)3+F<5Y9gVti(3)bH*k9ht{p9{dP1l`W?{>#X*;)k94mos+WoB4#p8JdYSx+IGnG
z){_Y>vxpC!I~lhn70SLbUx;s8GX3nFyJL3i<)JfL8gbwD_x4eJf$F+#UZZ)(=I;Am
z*My60(z<I>_xa}+4#aEPjt<3pEkwyTiFvsd{LX@nY$Qm`-@c+aKEB)@aF#IME#9YX
z9<gI##`tsPsDDwK-rNK7)$r?6HSYo=+(ggjlHS#7SSNTKyJ>z=ORE}eFq7mi{zY=d
z;{H->U;YE@c>6iWHoigLy+!DU#eR-lPnhR+0O~ju9J)9iy*t^&y_RE`@KEgg_I5e#
zm?hp?e>9+S;YGmcF1@-qtebRfziRu!=yNo+-9<~|(X)4wQ!#!X)6JO1Q3qjKWvi+;
z?N@&64RL$7T>kjO{>tTx8`cjV?<v$iua~mbSMHb4{pnVY;;#UDUeEWRwVy0oa+kCl
zSgLQU=-6Uq)EhI|os21KPEu^$p-&T`X}=q<h<=BajgN{avu2GNRd`FsoLo*`%t0IT
zo;9bu=`<KLIA<0(6AbObvTi8oBrvK^Nqy2wU<`hRq~t2xF(qVrRK{L5eZP6~aku-&
zJ)gtoGcF|GSyGY?@%XsX!f$u~r4*-tLIl;YnQx~jrYo!7X_&l5Ascu(lS=}pbq*ru
zz4`Sk<1L}8!SQv?Mj8hklZKuS9?j={Z!2aU#OSP}FGUE`X*UWpYzwnt+uY~@hp9->
z`bZ3Ek#O?v?v|*+L!Woo<4PsWO*7K^_;qLHK?|xW=~;c9is}|WLJV>qkPqqpb*_!Q
z?kKhV=hUk!h#jB%h3|EKOLc_|4_B^<hJklp8a{aan|H}d#dhJVJUpEaZ!d6xv7D4E
zAL(^D@6a6EGhmZq-ruQzXm6QyJj?p}TjwpSz{g8Z2lDf!dEs|`VCPxJ5}aGH!en%?
zu(3~u(L;j*Ul@(`4LqB8>Q1cw1BTeAk9L;wOWQ0&L;&%^p4{E)7v>5NmAL%>DpC0V
zRXH$o>Bq#A2$chow<c~3e0oe0*eoIMh#wh3IMU)xEK;vL#Aun8Qk}CGCf(=H`CMC0
ziT=GZb2{jTZ8GtB)o)#9QiyYZdU2Fza~QMuVq$<-Myj@j+Y$P^p9Y2T>I=EqYDO`Q
z#8aE)C|0q$)TWGq*5`^u_DyQu)7hAA`n&J`7&<>`d$3lgmFv3j+n*h~Uz1vMIva>I
zL|%A(<OjL#lMme2KKp&2t514Dzj%AI{w#;>J2-M_{PMTw84u2Uj~~PTM1GHa0I4`R
zx|K`%%&Yo-llt(>^XwZIl^y({{cnOTs18@(9}Hb;z?ezu+89`m^bc_t_*QSn%3tSa
z{6E*B7EdIn;h=nzqAWp3Oyd6#v-{is^G5%trjVEs>XR1v16Ydt)Vl?b1_8=X7A~IG
z_wM6KhWNn1flb?nD5)Y@gqfG@m+u1s9Uo?)S$U4WF)WQb*TZPToS}|BRL99+YM^QT
zbsG%nqPOK2pP9Yg4T86jrD=D(gj$v7q~QDQ4uc}UB6h?4om#|X<1<0awIwxwN+{if
zIcbF_Hf+7MGZsvByQ^&lf9mV-Fa9MgGRYl2IC^`FcoWURM0wL2=S^#sb1_=w6;}_Z
zK7jnOd7@{<7F$#9^X#=YX<xQsLLXC4_qos;vHZ)<_b1Y{FRX{u^j{7}6mDMHV)s3{
zq5gpN$8D-Wct^~z%lQ?skW_xbJ<Dnr2lzkQ_&>aQu`a2t9(Hg;&r(%HWKY8XaKkVD
zKWyoLTU^M^WKiJU^@fgT#nWG7VrY|3J{VZ;mHH-L7H!Isy`_KCVE?{-6%W4n72cni
z_@!6)6J_(@%u%U}skg@il1omgN#LZ)r1o4A9S>l9A#pa(x@4)Pq#sXd@mXMZa5UAV
z_aqTF`TgnXDerpD)xDtkHPisJ*0?V}Khd=5<Y3Ru^Y%3)OdM6@96mt{>097j`ZKrC
zK2m)1a%j~3Plv7U=>LT+IxN@SozF7*FkN#u1OEJ)wDavQ%41j6*HEb03DA{0x46cN
zaq%DXc;TO2Ng|4q6RA7cYwwtKm@5kbQ#Td@SeJAcCtq6KV)w)>3VhyRIs{1O0{C4X
z4Pr0Qw1G2S))dBJkLXQ;l<gePILkOe@J`)wSEul$I}Rjy#Vu@i^8O|mD?6HuLA-{h
z130Jedb^6!I&MGXo>GiE490K~%P?^Pij2|V(!&I@!+4k0&48wQWlMN6m>!z4ICCv;
zCgIPeWmE}Yx3WNR1)tI&DM(lxWSLXl{oCeJ?@SMoq+f?zyIY>A$LrB~Li}jhiEx?h
z%zl;f!td+GWkv0+ZB@$5UK|gY#^aC?`&iQFIkTTP2n}hx-3*s<ghhWS;yeen9D(+n
z5={0_9BoHDCQaN$v840p=DS}D;?FrWD*xv0Z+JefkKdgYdblBS`aMPlZTuKS`fzKg
zYV+a`%&|u%`*(hV?Lf2*v|-M$6AUWsZr<odeW#5-`*GsyWiompuII6rz{+*5LsEsA
zGYD}JARl!-k8B74zdoUNYH7R{A;EW;Qg)_>PO6O1*a2zi8kU+Mt7@inVv;3Mhj<-o
zAm|rJ+x(u-I8P-`ZZ-eWSiHodmJtWP1=3`<8p)1%-E*Ug{98sjvE6$4zb5fF-C3T4
zR*SNEOiWuSL-StFCQR%5(!MQD4(y$IHq6Yd9~PDO=Mn92P6Hw#nO%S5uJC=A)3Dq3
zeyJ90CTy=nz}qd@vr&gScxVTf<T#DK>~6^!p#Rkuo2a|JjPUfDP@wT^SN;%9o8X*%
zUv=BIUei*qSU%-^<xXqNHHMYV-6us)CsVhN9?*5>U;LRF9R8%tHnkzj*3fg1T_enw
z^wOtyRnVr_#yVeMx^L`0q5d~|i_)uoH=QNw9D0NX+fNC!`&hiW8j>K(OxGRdIk-<H
zyrwq(skZCv=lJ)Glf(p54U4q-Q{C26VK<ZnNfUY$c83m8mHPC63;spglInqS+e>()
za<^^nuU?r)%ALQ78;%?0>B_NVZ0B<+yyPbqLcC>{SE&bk22B=a?xygM-g~cSw67z_
z(B#X^*d9ehZCZEALQAOJBrK_eGv2JXqNi9fmhf}wc}+vm6rHNZi41`)3XcJC(}vA2
z<bU61KGdcBwq{A%U-0XWa+iI;H8p+>EocoIeE-$M!tKIyKM=Jfe#1c}Zovzg6?pfn
zNAPvg`ijt(jPl`}{O%v0KI@3hpMGwS7hxqF8AF|C+V<AF>_-eJo2CIcUp}hamBS=B
z5AQ;s`yu>uOP8tT3A6m{>1s4+&hyWubhigXUHi=MY^4%aP`(xav@sBNXa8i`g@q4!
zFI=3=4t^+q)0rK1nEKe_v99BjFEq=M(RV2(?i7@np6SoWX1CU3PYm;JE!*}JIhrB!
zCc01OU7AC+s+84X^z|Ad?O89$^8uS3=ZAMKw&3_QHtZfH=DD3hKVo*S4n#bgA2IeB
z4aqSu%k3~!sMyaGKHmH`xcNs)*@?UH?&_qx?<d;bd8G>Ox&U^^qvu~%4>)J1mj|&O
zrgb54=`K`s^Mx@p9M8bRu9pV6#t9so4JjUTNbtjW!t^+XW1!V4zfZjt?&p~*c0q+x
zda7JCa^^$t8+J-sH8(4<Pe#7W&HiNUPpd45tlYoyb$x%&^(r%^vWi6W*R@FB5WQxa
z0_8wLkkAwSUI?7j)$a}Y^AJjfP&Rq`Ivv{Z9p%a|*-<>C;-u%8%u?OkfzajloZDMp
z--IlWm{%t9i3Zu2+6!0I6<{fSMhS7lgRw!2yYnwhnvE<El)vc~zU@v=$@GI>R4>~y
z;a7&y&yIzr%ysd5vZw=nAG-5tPzW_7&ljP3Fg9IT^TK{NJsN~;fw8at4s+<$EhiCE
zi#ffG_9#gA1Jdr%vlL8L*R&2NX?6e|#N0{z9&M;F-#x#0&+UU=#QetHlsH7rmI~><
zV~_Uo59{c8!rOJJ4AXn33oIA)71<>favgz{qp*vUCMwJmz1fBAm68r09S@i`vc6H0
zlgHz2WaMy)Pj$yU2C}KA9oN@|vCpbLy_b^a!hsu#nMzwv{4%cM)OQK8`n6A%m~nzW
zpsGIc-LS)n#d*V6peWVDa=+#maE7kiJ{tC=gdSLqZYS_;FYIBpe9#@`R5#7&<V>)o
z+gSNwvQ>9j=7F{{BZv5|erLWlFlk0Hf~N;F%~TdZ1|62{{Pg-K4IKNnj6Z|Cdd112
z8LL+g-G9zcU*BET2<ekLVfN|T{o%V$Ki#%$4m?kvL~lEE<;03ZgL2&A<1{vVL1-RC
zUfrMGlRCv#6EBB}-{0-LfBnox)hlW!lGDjpnD)AnbvjI$e!y*e@z$8{oe`&xoYNzZ
z)HW)wc}6`xqpdZi>7g61O+8_%pYml`Tf*x1qhcyS*S5ApEuc3fGczd7hckv+oa~=|
z#p^S(pMH}wYel0zU?1xez+WVoX=t@)M15TSy8e;iubSbogPm}r^zQ4U+FJWe&#T{i
zrPKDVt2zCXW%)^O<~4fYg_q|3f$!n?=N-@VZpbY-a;VERsIsSzdABtQr>eqE*SW|k
z+&qs6A1~g~t#^t0DUNKFlBb?-o~7EgxY(+dv5gdarHge^Q>Q%xdGD(9&V~0YQB^Jl
zUxw=A44~}vhWY+p$GKkmUtt!khFDc%Q-}@fF=@2K4lT>If{$?uIAv$1SQ2<a>D7|a
zjF^**uj}nCqtozPs%|>PJE{z;vsw?ibcfoZo+{l7>gXPKo;ZJW;J{Sfuut)+XUFdy
zkLiXOF@uOMp@_Ge9<M6i<$KwYY-zwW&UI&)W0~C}#5AWOK^)@|;v3dhaSfu^^}w;<
zL>J9&HX-3+W6@lj<MjED_Egy=s#0NLJhJ8Cz@hdtu~Z1Bk;{AgRFAo|Qg7S3L`<VO
zKJe!3^zRMG2!e6MR~P0ThYr}jma9)*nmTh$_3Q8WJyyzYuWa`%Gj;NnVvAdTwlupB
zh$nUZx_#-wJ?PM(9Rk+^K=>(76PfO<ug7xO1-oQVJ@4b!R&GtY6qSxUPC=T4$~9ts
zU3>pbwII1Ses54ktK;hjy(8MquR^YuzrGh3^?Xm?F6ou=mzC<T%tM`GPCXQVE!#|^
z*QI!C_|KGQS7eVn_AFU3*(<7N-5S5!J@$b-g&pMgj_>i2qht~vd<dvcidMhwj9323
zFgSeT$*z6R7LHVIXO0!5!X8fGESMaI+l7Am-?rkB?q7r1b)2ydvJR%gIYB*V-{4e-
z-cfE%RrXX!brKP()Ax+zdS^8>FKM~JYNDP(osjn!@JN+)qEpt|H5;q-o>$c=%!Lvq
z7TAT==%~tRH`cd@_`iq#kd*ea#aE+!RMq6>Z(b-se_I2qua-o0US~_kEA9auKH<7c
zRjc`stD8vUMJYazp(8lwfu4Sfr(UE<&S*^)tkX1YO?K1N94>I?o3mx<mwghT7(}~H
zLc_b0v+wphnlK8l@4a^%tg8C9<;^F{36lfZbQk)>RDYIVK*8q1+<MJ|zwbiTfc-9e
zX8)3TkI?5f9_(k2Pe#+^H2HeVmks6I+FwA<_p$9s_Gb<~^{yP)6mjf7PkJ*(1hYRS
zk7@6s+o<}RUi2Z2%d(~nCfZm(wk}TUIQsdZmq#j?Tyn#)LAzj7^nsv)d%@eST9mZ6
zyPc)z5sY(|_*bIorFUaHx*^l1hNx#ARnKiOo^k2S+E{<zNVK`)*mNXk!Ts;?m+RcX
zgc`f9qFKj<;PuJ`d5jC2@#`&juJPs}wwmH(xF!kxEi>pL#v|34H06K+v&2+c7Obd?
zA=7QOm<nt5YmB`>eo4B;QQ5gsdEb&F(OwOe?t%GM-0+wcG~q$vG>@O9WfWa0PuMBQ
zb=KR-@}4)DS#IrROZnhGfl(8+#p5zh>(V<|xDmUm{1iLuFI?&$*jg}21jX#lAx&N0
zVkffR6>>5frr-2)q4G=5VW$OBT&gxxf8s6IWM;gV-5LzbLu1F4#}0{cPH9DVcsD*a
zd)y<qvk9WmXXToE&3{b+$Ea#}d{yRVcDze5C*Whq=x9GT$o`kiu4&LD>+u5tQq9g2
zwHcBvyTE=je~}}fuo+53m+STeBr6Sv(lPwhRmdp9iA|=RnR5#anXjp-i5e3wzZ93w
z)15Kyu+9f4o68A64^9LL0cA<d_+Op0N!zsbUJfr2K7&L0I_EaPTM3*F(D=GpJc59;
z2~?g)Pj4#huS>5E6Y?|EsDyRvC2L+?9)<oERRHkdRL{ba8{Fq->8icqC2UQS+i{fK
zD7O5q=bmEFPx@x=gq|stPF`Mw_4aCnHNfXdM|aGm(W|$U2eSEJZO}xKDs)CW0$bNM
z-8kKot{hKy%-U?YA7D<?u<H3ta=Y{&(p<Dk$EJ3H8mtVhG21YywZzMzYo@!EqEE6X
zUangLCt1&iFjxj#n^`%_J@KTN4z$WUrS{SAGY31NcWy(H-JR$>&9<O(Yi%#nc$xH{
z6pxMBO>+C;OS;>umqsG^{Hb^!6z5rEQ95Kz&!eOR^)Sb?3Nkzu3WZlJZ`~aIhmt_*
zYQU+%2K8R~=+>bZf+v^^o(sP2YP|+@Qj1(^w!!07gAmLknX^Qf<*E!CmFi&$^k`=`
zRW=m~6%{c>B2ihpvNlbIrdLC}9+lhoiwPMO#Y~|5UJfjTr*GEU_TYVK;{yYmA1|x8
z36Y!5duDCgu6qaQjB1a`RJ&sQeH7%h6FOhiR2H#g{>=|~RtO&n<JawvM2Xf3g*VbE
zHIIvH4)e242h1i=wS>JBE^k*`XliBOXG6?D?BM?T1hh;`sFl>h=0oo)8gPtQHd?vU
z7y928q5YZOU_GbtjpJ_>rr$3T)SrV~4u)_4@ho15hQcPzhTQP1W;23rEO4|n`IDw}
zlGd;1Y;WsQjY7cq;RhTmArp^91zM9?v;cqN$P}+X1k0vZT^y!rt$S${1GtB9uU`_y
zg(^4Thfboob=ni0tRRm1FTiw*U(59O?Y~2o0{_W+$Y#6A5qR}e4Mphzfv}E+252g+
zZvZ4)bTDAzlLVJanLedP^5V4rYqh44_@ZuU<U3!#CZN0;q|4+_*WtI(`ido-3QMik
zmj9zBL%W7kBS-qfepYEYMMX_rNg4CpQ<{Tzp52wCE~sQ~VwZIk*aVgymG{9O-=)c7
z>r@oI_Wxgcp#>~UZt9VpggK)rzdbOvusH#ge%pf0tERd;7fbLWhVk`rOaH8D3L!(4
zW&W@m7=Nf;TqT?>CZ_U52kPIJ!v^^$by;l3oD<$N%pM_kb(rn_0jfk>Gfk3ztEShh
zQhu}+eOM1;a@RgUgBGq*)nO~sUX%Mz7X#K4$*P+R^A%*yB^Gh%6hz?cz^dEK3CKx@
zG&f}P<2~u5sW<-^vTNZM&$4EGq~9`c6ndJv@coWuwz-@+$hrkqgS7_J+i4oT!YONW
z(yg%Z{;0J8s$S_?LNrWUuUV?HnAaomr~5u!;Oz2AP{O3tc_@$Qsio>>GyPB=ozp!%
z?3#1zJsG_A(uAFu#~uMPIqzJzpl$oiSKJ<r@t#5tuQKTQg<_P|fN`M+V7i;7p2;o_
zmbG<?V-u$n`Z9M36u*Fx7&D5(dwBRQ%=?8z$&nx>2&clwO&Gi?#z%qej9*zg)`S<G
z&r>#3ktwWJUAz$z8nhWZXMWeo)72*2r}>xYD^Jhowx%jsQ>!C`zn&J~ar|AGaP7nG
zFLxDR|9|(v<{)&SIFUW6Pxgz5Kt%rkp()I9(*HKr^nduJFf^5AlDN<OV24o~=#YB-
zi`ZD{BY?=n1ENO`dC8yGzv^p`j8EI<aN)Y#@hh>Alb1pa&gMLeG^{DHUX0Ta3om|s
z&ie13xZ$i~Mf<sSLlpj+fXb(LXrddJ^>lUF(<xsCC);#CnF;u$*~~^Cl+*hufo;ko
zOPhv^U&@@%P_^8GZIUhB)e`!=Z`?lJcY#@!dDbluefv>G+hlXiVl63guc2uv<UFL*
z--db^7<}jUWA(nN?)sQ=YVE?C!M7>Ft!(FyHPL}G=lu)k&H{_RJyE<6`jmBl*|7fk
z#^3b@KjeC4jADjx9<}RDr4*&t!DXgWt3bYf$#<#|=f~<CrrR1~f_A&X27<tSW{Mvy
zXoZtD&z5JFW_pPY$j1*Q&hGf!Yf-wl_G9g8-^5;2KNCvV8r5t}ubQXTn$f)$Gv41;
zY(ca(o|?BQU3TcNS9h^yYF{O4`3n1GTI#agn_a_6Q}gUQNaURxUI$MlI^7<xAD}&7
zZFm_33vljVDYSGeqMwXcoJw=LpFYNJYluY0XsZ$H)6W-ibZhpXIeSoa%IVdkf{*{e
zc4Vd31**B&3i{rEkUML#`sDFd{(hDg$){H>K$c2gp6z?_j7RTV*Bf31jKpQ^h_}1Q
zdLFC`L$&HG1bG=1gxAd%YBF_eBEKj5x+PosIgq+-J9K7bCpG)iHJRHTt=aB1(>YZC
z5BIK~+}x093;jIpQz*lNGBX#NWJ`<E!!_V)6WZQ|{S{HZF3FaLrJj#4_prjQ^fP;9
zpB1GS-_D$?CC=BLA%eWoUIgC&)5vy6Tca|aGv?Fp@_)Ajo{w>7jJ|l}PW3JvSU(co
zo}AgtD`h}_hR{}&XeWqj0j+gjMz+M>$f|iIx3NSw$0C$5`(*mvi|o25jgh*4b;kZH
zVn?+HPxo3D@mgf3QmLWCACC+fK^3*HydGo!YpQ6yBvlKaIVSVWWj;gWKu<6t!%giE
z?#M{^`?8#~Eu`1%8WIq-wwcLH=ef7a1x>ff_NsfTv}YjRP-ow-98a<47M9X`IE&1h
zX-Up&P1G8*+19}}xMn?0dnPkIT%9>JJTz)84R$y2-ogH~6H~-_SFen!e_J*kJU~mL
zM>e$9uRTj%*i9$DgoAt-+b5Q4^z^r!k}cm%dd8#sIk%aq&J7LIOL4lZx*gn;GcPS%
z=6fF$3GUR$b+128@Hcqn`SNI6CSty#&H={lY1P}}4HeXNq-NkRl^XREX}g)7|I`ki
zCcjWepCq<`b(4SpY?0#|qI1r$hM3Rr31g`l$|#++;0)abR4cbST6pSMeRRdODC=5X
za@dX)_8dQN%;^6v_J?_kuDP5<-z{W2{lAh7i=qLowZMP2LZ*M^y_<x6<v@k7A0;TD
z`B4QL3w2{Ay@*$SvspE2ZUqZIvglBebps*iMDDX`&n{kLHfbNHljAu44b4YediKk(
zlIfcB>Dn{U-loIUNz;t$s^6|H5o<_9rU@;dlsdMMp^WCw7g*61bZ7zdv~=x2$6aRg
z!B33WzTbLNT2O&M#isHaGuLw$7RDCpmDL6M4*01S{&`5ieEhoJ>|plsPhOZ1F1mvB
znqyRchNGuRrPL0T4yeP>jXzK@)bF5q@U+E1IaoJ!@2f~eKj#+5kCfmOI>o-aoSF^u
zgOw)KCxn`<HVRe7L(KQRRoGhKZZa(e4@&IyjWO0$9#b+}jyh&rZH;J-QclHRGb*td
zW7~Irg!wb;GMhj8i~g5`uCWUawWA%Q(>se^pdR1&c5Rf^)-P31K07<>*ZyORl&R^n
z234Z$^J&B=TURqovYuJ`5%<mvaQHWu7ABR^_}`&4HodFXFNY%>B?-<An)v00UlRwz
z*597Bs%-szWZ0+V_aO&nLc>Y)^5MU{joyKKoFTS}P?IoSyfSNOA)ojSRcyPsvuSH-
zw9v|tiSXicgLEOsH;-HtC<nAZ`e}1#d>YmO%h~*+2FqOMcq8cSo9x{jd&EJUz+CHc
zNd{R^9u9*14ayz!Q7`>?l)qIcSIZvZgxUxN&DKZz2*sSB_sG$^#JTuiLS%KZv9V*<
zov!O9EVh}Xo>Ol^uric{^Zme2&O_3GOXzBW(G!8iXdy3?bGcM-*OV`?{O&qL{9mF{
ze||Hk1{ol-v$J37noGNt(Rd-;^_qlfb~xe*Vk`7_=pWAS^!^)0G=wZp&?wX;$#bHz
zP~v2;+|bHYUBE!AwIA{QCe?)>OkeoRvuVg-bfm+c2>Gq16S=!N18WkPr00U`L4~LJ
zjSct@?2Z~pFv4UhRTs9v*WJBg<`-qJ-exy&Y-?hs?fF|W3}tG_tzj$A(Esu-b8RYe
zd@zU=%p7gJJosEXCUJu~ZWe>+=WR@O40G1c2<(<fc)Nk0tvc?^g$4wsL;ED=2j7Jg
zBY;(UlK2iTA@}ca*L73_thkpK_~-bB4G4CqDWe!w9J0kpK4Fs9Wv0DuZ?Wk}S9U=-
zTGFQ>3!!hO5hn!}?Cwv3{w!^cm7FZD<Mg+qeXTmNm^$`XV`ZDhkb_Ka&}@5rtiWp}
zH8UY7V~lfxzeM_WG%GDLvsGspa-84Nmcz!H_9bmp`4lUWX33jMg32ZJ6a2aMvcT3_
zb_z4BlK5(4;?dH=qnCCtxR?AI{qKg#O3o(g5pS|b=A3m2*Xp^^9>I7EDL7j}uw3jm
zS`e#lJ43SLJ_<1@mUl69FFw7Z+(ByoxrdYY71erjqhn(^hw91yU5DAKv)gzJAq8MI
z#sxm*FYu9tZN~!d|6pb85qNVQvt%i?zijW!*u4}CKxDP{hT~}`qy_w07&Tq5KX&53
z|6mGycv_Ap__3H7lAPecz<S|8jB_03mtbPk9PN>^0og97W!Ltq&9q`zE5%!r+h^#p
zRSh3o@j8N96rq;1`x<v*&Mp8Vzj<s0X3O{MxXPVauycWQVvY^|gYkq*r2#e{{0Ac*
z^Yi=~Nxbr3^+N&+(v%y@G!~=IzsJS~F2CJilSl$p!B^Uvr@?VPkI5FC-{?4y58Hsu
z&@rz5_gL8QPRxbFFG+$S+E*Gty(;uKq5v=WfFTuj!_SAplz4m+jjW7k-)9TOr9@)3
zcxD;`^J~k1xhf<lUU*9?3RTtBE$mQa2BlZ^FTAx#Nf3TZ-{8(w5je5JOMz{d=O(qZ
zn+CAS@kE(E+JJC4h*018-h;aL(@E?%0l3A1s)m!2f;E1Q`Faajpyk*WtZ{RME;zWx
z#@s`j{}QH~)GO@Rn)~tN)m-^%9kM7O{GQF7J9NP{x`0XL$nt}gHZk{#9X)sPO&bKt
z4T6zkdX|=7_j;&VkIS5F?mfiB?~F9TQ?!8ogVGX>5B=Sl1T;#JL;o2obtEk=zU@eW
z2}Wb36&k08dKApExi5`4!ojOR7exNHsZqAB4_v8}eY1bXLI>QVq2cD~a)abjNE9{o
z8+}g_SUH;(t?ydp;2j%*AyOpeb8d_k<-dUE!6J6M2$u5R*fjtg@(PO8A~QvTFO45A
zgOJ5G-Wj7_?1-1P8l}da$gYdQd<HSHfX2JXE`TB^wgLMCFb%2I!J=_JASup~uA<#P
zf&E?012}7SLIp(F=mNI103{cpMc@EZK~8q^a*5UiJ){j5W|4CB<YWFDIl?`lwWbJK
zo((9IA!`vDM7H$5hz&!G%}#MdSxx-~RA3+=6sr(rY8L3UNi%MeL$XB-)d|A24>++w
zq#yVm?4?Xz6<H8n!RefFfYG{ZgK)Z(k=ZMe?XhR6er<aOWM^Ps0dPqJn8hpL?{ulZ
zB5cu2&>+O#4l9I`7$TxffG{FV>Yj9K6#3E8h<omM>z?B`(nYJEV3k3OdLjectmViL
z<aOWCBoigA^%86UuNnz=UwPIa_!?|l8bKZ8_WTg(2N<FWXS;!Jx^juxH(6qzLCKP1
zAZr-|22clJw^rwW;W*Ly*e2%PA^R>sQ3FABRU})oL<r@-&K7|ffK}WNF_!46C@QQW
zY481m{5Q~C{NYfaQC;Z=Uv#h!gxW(lS&Y6Fw)$DJQ)1S#7?v9=>aO;Ak!jlP?Oxi-
ztlhBMLS_Lsbg)zrI;a)c3A`;iDBazNWJP@jIa(4youJz?W0IYcjBO!Jhs<_--H9;=
zi~j_yiT(mE3d@&i80dO!BL9lkW9tc_*o!H0#;RAD0ZadJguEr|6@=vgR<N{P&qL(>
za1G>DpfBR_o#d;>*6h-uxm0W$a1s&&x_=5*a1@q%&Q-YtcJa4^UCFZPj_zi>tMY#(
z%9Jg>hLot*sltpCJ~xOUj0dXEQP;)lDI;mO3>_~V`H!@vf23ssLmIF>z%iN5>6?E6
zA@XQkKWO+^v*eht-BIBJn7Y25d;WUu#IYV1qaT<esQrpaFdzpMXh*0KK?6AUL8EV_
z$0QjovX&Y~1epTul<TW2Ril6@$&DufmY9}I38Y$x(byIXBTH<AH8eeVz4z5?d_Ulk
zF1F#wg6NtU13(n}y)96m1IWhSSDealN}jj72(r+{G67kVWMGlxOn956=YRE=BwaV9
zo<thG8F)kST*oehiu!|5qAbAH<)!eU0u5zX1VMUm82je+UDx^piYU|bk#1fq33Gsc
zF@}h@(0a5KHn<{YCCX#4tNxZ7GV7v`bO<$s1>llcst}#)k9UXvp=Zxw^K{4)haZIG
zHV}dUp@Z4jBKBg```ow<;yPAY#uc_VNS}k#eb>-bd#>xfD)+1$<0o(lm<6%69SpS>
zwr@c^{`z(dMi#Q+jr(?vAgLuB{xk_zB&$Qk_14vW@#`2u(OCnjwHpW#k&|8jvll*P
z`+vCY{ofY<w-*iwy!&X2&;C>tF^Hn;CPv&?l->5n&;UvM98-|zVEM?s&Mm;^@X;M%
zry8~mtEo4Lyi=4iew*xcQsM~UunwMc)@A<qb&;;)EdgqnNppX}*t23PQ<KfCcj-Ua
z^?v2Zo$rvj5`!*-(V4E7$mU1Q24xnPlBAM)ft{i0_DL_$s;DI0ibZITN4ibLBJD=H
zY(U&oe9Ew8<(IPc8(R6ljnZLGm6KOLy$-SauqnBlrySk;fZ;OKX<!1Y2I-lj<ZZ~{
zh_BF73;G8^tXeO7q91hZt2d#Cq{hv^%RN@qL^6cDTCn>F@`%9y0w@a<XEA4&_w{4r
ziqPgf`jhR|ix~R4UHC>Sp4GskYx3u6@eelZ9Qo#j!0h=ujvJo@2l+pV0O2){YqJ#k
zx1qRp!_%jUbPp8BAr*3Mmia`o-Mk!fAH$5jTCB%E=ZsCocN`4bL~RO4WV>lxH%&mS
z70gZv>bVWvsR6-B7Dibh7&%ftDg*wpUGU46ubzZ20?(K@@6thf=ea$v-dY~$5n#?b
zjS_)^+C`jJCjRuH*f5Y%QpM@Ra@D_!o$1^88s7ydy5Rn9L9mm<&<{hhzOO6=`QF%n
zn~pfUMo;inPY|?e4iIcNM#FEyP$E7eu~DE~euDtN2;gpGxQ=-7F5!<23-|^H{xlu)
zH3M^bx2(?K=)yzrZvrm!RC<h-&Ui0Q#f}jTdZ+hjGu6e;Qdm&%H5&i!{Dwv{SpJtV
zsw@i0qe}3T{LTXt8Cpl*HR@O%(X*$|Zv9dg(1_WG$9VPJZ&~=Z!3d?`d6B$u(66X8
zV9fS^K-MWp-KHFEY{X9rMb|60(wFdQmBd4BqGIqdmc$5c`{9epTtn}?L%Be*zLRbj
z{v1o>_+D_t#tsmRI1hHZJ5O7fv*U5&`^d;U)&iwrz}Z@43)D427rSC~l!7!sXVVzL
z$%s8+P{uxPxS?toy1UHW4?jDO_({XhGfBj)<qi%n;qZnANZ>pM%#Z;2s%Es46AW5#
zr;eAsGd2J+qJa;U)~kW(Uh+N|*;e8og5EsVtc}>1Dkzj<c*cIm0vX5wDAG5k^7K4X
znu6RbDO9p+jIIbcLR|_R(TqzB9s~T1*~W>Zg#%G;;rIkp9jS-J<L<gf#x-g#cn){F
z3TcyGM&1zeW^l|G>m`3?uf&%D*-ow8HIcFb8Cek!>lApy8ogvkZ}b!^7uH5|3)}DO
z><8v~#!~TO8!dpsn5#fg-dhKVYnbMs0bB|fp$0nS07(QxKP%r!U=E01VI4uCYk&vq
z+(yQ2^f1f6NWocdq#pb#*YQC(#%&Ghwv4rWhWus#Pi%q4T>7Ux=@QT5d)lGMONU|z
zxL#f_t~LzOBWLF;5^Qg-2`UO$N?owy|0;Iif4WGbJl6Y3z@HtfO5#3RV<15g-?J(}
zsa<=fu19DO?p4ETVlAacjgQ~e^2!B_9c%KEx6mr70&le7rWz-4KH8L~mk(ak`HJ&7
z6?L(nHtzxjEPE0-N$!AD!wm8a!}IiJd9Gw$##R0>PUR2sv+;f7%QRdsZmSuG7Q*s<
zsg0%W0cTF#dOJ)Z_|*wx8-g#URd+_Got;KFK^bNh44EYGfnVOB-rCVjyyY2(;qQCF
z2I?Tp{y+j$+*483^@VW6q=Fe>ERgdThbXHmBIB*#UjCna?NNnRPw0q>KVYr8Lb*A&
z+Va2*v!Bd_C2^K)1h4?K^Qbf0N@yOxMw<b<pfMh?vbkcS{M|}%K+_VS{JVHD0q|%q
zZYMmk6f!OeV*Qxo4lG~tzXq!0z-3Pj;4|2bcisx(FfSo<SN!L>%)B}wT=f+hu*D0d
zrAXkGH`w(-J}?4!avtd<NkInsj00XE_B0P$ta=(8>;?40#(^MzHPGK{puY?-8T{1o
z%iK^+0y?4h-O?$$iImdfBV279T45R6Hg*ec@2V^cQqrPeuRd1+<x!A#O|mm%_KyJ$
zV2&4jiJ2kb+A}fED$e>P-MFdsCcK<8(xwZhTOtd_B6`vbD|Ec)UDOaLPak#8p=Nx!
z@pTspz=t3LT2opY0xSaNv6W{xs&*%Ve;dR(sX>}Ey5O>%lf~Zlc9$8RSXrS(`d<b4
z!)3{`CHLl#>V61EjE520Hx%+;bI*-mbOXY@2s=T&TJT&qg}lc-j&UC0vbp=ls`h{l
z<SAbAQ0`^JvP&5=Py!Sg7709vIy?g%L^8~wW2a@umL>U$pwNPQXStROIh*e;&Ry}l
z=NuY8lXH0DwGkY9K>OkLP699*_~4B{nBR;WZN`be20rXNfuSHr0pma~dB+^#UxKf_
zOxYcH;!CixIaEZj2Fn899j_4Tb1CFa32Tw}t%#S=geW?^5dg%%rAGEm*2!lzHKxSK
z4Qa}%JhpfbyyYcdRuPjAd|(fG`qCH#CIF@Z@lAGGB}Q5zd>2cy2vsB+Xmt<Fz4`5L
z5&9yNU2zJ*blZz@vkut3jMP}h{-hvXn(Roo;JHg!T`6w<#E~1I>q!jxJoPb08F9BG
z%}_EoiDw!FaykX3y#~`fPHi63ZiPz8%FF)48?yL)2K#p0h4$29$j{DkW1)RrB;d{R
z(M-G-;RzPB)t?5)^5O;TeF%0Hf>m;V##7_F6Y#IS<iAHAdx1H*kg!byAZR~|I`nY6
zJ#6gf&7S}-`?^W7%A<(U_MP@ZDPFU^W`dzqHUINWw*!Lgqou>)2i|yqK}x}YuwbyO
zcAV>=*ibVL3Qv?Cf5ViK#YUV5J($^Y;ab^4UV*oLgN{ucaKT;46x7)4B8$u>Je2{L
zfZ;$;zY2(uGnJ~@(6$R~KV6Q0!+1%oW4Ub?Y}k?$T3N3ia)*HJ9%+D>ci+5_#Pp7W
zNYEh)Rst+8<eiXnLi_u^4HM9%jBvcKJxr4I&oT<(xxQFmdn?&#02owR-2oWoXmYG<
z+~$d=_(3b7WV!+rS>S~Yz?N!n9NP}hHP_7D2Tp@y(oY3S0YFDH!A{W??O<_`1qAAP
zt{E!sua#J8v?Ri6%G#9)e%g{QB9z_8eQ|PwGr4w-Fy^xSBqX2aCN7vx#>2);gs=*>
zzXsNxO9Q?@<l%rrvL`Ud%Nl-@_N(RYTF8;z0lM^{OOvpX5BhH`TfPTg1JRtoKNh$n
zRmb9hh&KV`fdCPXtpeIjwrTyB8ve@4Ks?)M3yU!EEHQEj?jiskA;Q^dYk3S<hr^f)
zeb6De)YL&PMSrhOjBt9-1C_N{lJ#8=%PkWCYlKjms1hf`5#+j^-xmR_`w5oUK7_@f
zKLM@I{J~~!j>g<5T3-42lrMoJ!{qmXWE?1f0Fm=sD;7I-cL(Ou0emF}5(v3q6L`Hh
z{PT@5F)as({bLCK8d?8<Pp<l-=I~zLS;H_Z0&cz;ryB?Cyb)LebP;XCxp&@w4Mc44
z|DY3XI3I!T{Ofe$0e+AKItlp+Gz>UnavfGi07?errp7dW@)ben?vqQ)tGj)}SB619
zz9-he_dP8J+>BEh6?1RG@9Gz2h;YoEy-ehyfluWK;XZX4WTioIl#u7QGf3U#rgRvU
zdm6OSOPc!aJVxBfkL5Mxe1L7RPq>B1Th59SNOlpL_OHMti;=+nS&O16mD>JI9$q54
zQevQ^+S)BpajN_XxXxZ%s7?+=Uqwee3dH%LJGfU?gkSdosBLj5PHyaPGhptzS3K1=
z=*=<Zttgk!U9$*r8faz<hFyyWGUSV3P^@I>VuVGt6cF+v25{~GIZRG&F_^Mt<f}^&
zvBm=DUEdT<0;M<J9<?ANt9|A7+DM5)SS{f^_j}-6kT4&TWv=G%PVU>8t!SWjB=FB9
zZu?71!sa)q*H6HV<;w@FmM;<NLcby4zCZFR^}-khV8#_`{d39Vsa<!x#P8BWTnZAl
zTh@ATQvxzr?7WL9;sYRNoc?-POT_1hXqE`1tVJs>qB%T~H3r%ZPu!<eEAMD`Mymrl
z)+Y+`KMlUnDAZv&^oYA>CEHC(@UZh{x7p4xpx2Sz=UU4N(Qsc;h;>XP==`FXqkKY}
z6jl!D%bx_6zY#m<O1^YK>&RDyjQ|izXhx9cIxIp*d?|tuo(LY3nD`cq6i$VoK6K-p
zn{l#nz+FZDWYBs#L6bSoeEZtTPHp4Ui}&2$yF0xomu7ClMP&9^M0v@0&u)Jxx3g8}
z0G0q))<?hxWh8uqbC-Z1DN|1xFw`qohNVBR4-pn`u6wv&1yU;xAtUEQ+k}oC^>3m`
z1NIgB+oQz5Prvlhca>84lm4j#-hrESWKR=WAm~cC;1mOPIvEZH@4Ou4pdj81Mb1PK
za7|>0jKlW010XC0v+dj>CG4CO2PvQK3mB8&h=UR{jsJD<C|N%+;NSY#rUc@D3FO;)
zqf>`)qXZ&Ha|eRRO>!S3yR{OF;DiCO^~K)%3s5)@eKxchF2m{YYb36$1~6a=FT#P2
zyQ>p*4sk?!j2STf8m}W7fLPdeP*`Ar@N$2Hlv<2;JflfOW@oRM0R-mD8w8?<fPff8
zdOrEmUC(IH_`|8wl3=Gue=$M7v4Rjy)6zZg_hhj36j12Z7#-`L5>&5aBx>NM37-|7
zl62L{EK3}6l7L(BvM(^+6Ji`88Jn*nNxsY#-!>)HVU2Vcf}Hm{x_!zX`hrQ3hhrB(
z%{cy|Xd7@%NH0L{x4~BIIdM7w^d59%F$|cM1>RQmjR2rUXuk(e^el(JK=<=Uomy&+
zo6YWLsfP<FU!1ffpqXKlVxW_E^sGD0aFSjB#D+=6I+6j_#>g4U<w;|>il~hWRPEt`
z`6d+xEN)v9?g$vtTMO_K5e_T;yYegm20myNDAiw^3vveNi2sSx925iRN``5BfmB&d
zINicFy%H;}k~`0Glf3M$fT28>nO@fx%ZNVoG;jGCVI<@$PO!5~Hv62n>mg|3x%W44
zz<?tuWWawwW+?=CQpEW*38j1r;S^lu_HpNJmV0G(dg&v!fp<m$z4S*Q9Hbpw1adSI
zG)`$C*ufKJ!_ImAL7w>p3;|#8>&2-}{AHrrpArW;DvBBGe!<~cJ1?hm``-8IJ2yq9
zJ0UIrnf6+$@f~k=z;2T-tJWSlIF)!oqOQpRt^xqdulz&cq_icX*V+`Fc54#&ZTFNt
zQsdva69)^e4@gF0{z(U_NNy1Th&7}c5ll+~)9z*MN(a6KyCO$FdsMSfk|2w{{MW*i
zN6atV`I3p7WCSGL#J}1e@dBC=Plyg*jK1r8S<%|K<}ez2+)Su}CcvAn&f~NgX$oN*
zBN|dVJ1FKz7O_@o76WSOBcQq}d%-*X0~4jYphy`5;a<a6KaCa^2esGD%>e$HIFzhe
zAU(d5PhMX5UA#m&2h!>L^mh|x6)|ZfW+Zcz?G2+nW)pBh;vw+k)+Mr;rkfy_Fu_hI
zEvS8Z6e0S39fh!W`Lai|Q=o{<Oybyzkn95;_Hr+dmWLi!-6^;28RYs9uvFUNq!&n(
zjLi84?Mwzck*($XBOSmc<wzE;(n(|c&BjJZEvdKwXv^1V#V@wrR@AC|wIg9FLFr?2
ze&_BDuWfqorzB;_Ee|)mT=i5Ta=rR8o$m>;d*Mkt6_Ux9XD+}A)@gC$BCwg0`Jj_W
zq~C+`OQlsHSFng$q?(ZaFkqTQfn+ssF-!%qjJ$sk^h+CRuO=^zwpZK6a3i&EP*EoQ
zhRN}N0fJpzn4#Re=2^%QNHY#J2r?NGV;?i=40YNKMS_(L$SoAHF=;vuWc)Y8Ne-I!
zycY*z90}U*+;SbRA`8BLS>}3MXrPQJ#DR=^28AMFCFceoo^%Z(W6LGIkBZ&{uPMzU
zw?p9ybjTc)+a)(*)@HYa9R}#Oq6u$FV!Hp-FIYRiFFQokGD_!4B;y@<<CquuAp;|h
zgBVd-{+Bk+MgcpIz@NFk>uf?%fc3_a(x;gef;C(Pgp-U_ivT@i$nTeHB>-o)!BzHy
zUo_ix*4iIPhuUDLghon`T{N|P1YH#8yLWy4V=7h(>~f|_@^R$izaotwmNyG|5cFeZ
z)rC7C;9h`Q;B*WP6;h9x<w3|Tu^mGo24y3EaocN%kzA`aR{F6CMbvvE6fU9+7J_<y
z3bC&D+jgx<_f@J63`~e@9rv<7S!=JQQ+*hPV-Osm%C*SobAwuS@(S0Mqn^rrhIZ~;
zHkda?D8?KPbHj)@3m<F&|E}{!j-)I+;cNWH3AdBFf{}rt%{$*dFa9XMyT(*wkUcv`
zCZjKH+zb@uc|rA{C@5IzvRpW5F(#iu5uVYM(g~E3laGGma8U^d^1=lrib9fQ2!Oyu
zIH-=I1XsB{M{p9ZE}LB8*mHyS$}w`HpzY1L&|#e3tAovrh`)wnE0>=wfs%Cn6I%ew
zfC;f~?Of)W1tcWtF#Dey%EWYyn}6QueSSA2SbbI0I#3QLjX;<zV3PGRBOnSxyA{Z-
zv2)GQvx{-FX0!Ai@TS$hnHD!#`DXSm7TXo%!le@k(b1Haa0bBR{W%=u`JkjErB&C9
z>;ac|345r+YCb<E@#>sCc;_B)gd~XetmTW!!)--^cVxu-Okf^pbsuOi9}EiC+%#X*
z@0-P(X6{nFfw?zN_o6)VZ8q^D8#<}$1j;oqpgKq+71)h}_su!??>rvG+?2CBMm30Q
zJgVqq`8*^x1lRH*c_5ailoLCq_u<T_AelF2vKW>=Ci!s(<c0JQaF&c+6usjs6(E0@
zgFWlNZQsfG{IH0mMJ&Zj|Cv1_UK{)SG0L_Bx2zr~Jj7#yPe**5Dfh$)sGW#Jbp}~|
z4Uh@v?a)co<JKS2Wk7iuOgY}GwTn|4p5m$&PDS5#dOP13uje;k*gC<H_HE{Q*=r@J
zJMVxJuG`z>UY2n+3KK|mtG~?4d)0L}4K8=^nCyWgP;wo17|84LxT%?Y2fAqU5}zEW
z^I4O7xw1HKk95(v-NljOH-h2Fa>1qPnbUfllL0f_boIpuVxoDeLj84fZ^XyylGH_I
z%w~`2Ck#@%_5~(emU%?u;Y;JmoqEzr?R8|rwrSw~SqsH$Syv)Y-}$T(l-!NGWl!%0
ztYP2fOiBPV!Jn6R!8h%Q0HK%m=D#pZz0vol&)h52Ot4q7=}K+l$`Knl8@%cXXDGuN
z7e!(H^%63+pL%L?ZaTE~yq-Y>TWD;xa3%aB9=*!(Rc5t_dD)+1U8cp9j#LW;A6L|Z
zA5Iq&MoW&j;Yt-jQL&a%EzV8IF47;Y=^xR;YNND(S+FpWQNb`DerF^7i^tocB=+_^
zX-MoKOrCycrG}jtKnY_|u1&jm%n0l<VQ(B2oKO@8S_Hcp7vyDs(Le0Z-88M7xRL!g
zhu+mJV$+9lUxslK&A33glwCg{KTYi7MaeaEj~4b^r=t(j=kkBjR?Sz35HfpgCJjzl
z<_t&nnk6PsnvBY1ra=N_0^XZ9uu;xehS%=VCR~P}(Zl*Q?SYp_aTjjs|GlKB7mUb4
zguuKZTEbwGoN=T~R~aaI5GbJta+tf(>5NO7g4O;4OvJ-)wrrL^-3&em-2j6_UWQyb
zk=`pdDv>kznp@#f!#^l!cKqgb#X3rDCVvt*FESya48`o_!rspu^O9fEQ`D+Q!m*Y&
zUo}J1{-%kE?StF<2=8+tdxs=It6nENN38#GwsJtQwH~raqXaAlOav=-@po8CLK+Rc
zvdateTaDS)0z-xSDY{FiX#lfkTt}kO=Z)4ocSE=mihB3^ahLaLsx<vUX8wH|V`+b8
z1o#3Y5`gr@o&nW!Un^z9BAApbihY*RA!qr{#rbDg>&vs4)qQ2Kp#9*22w=~bjo)f_
zS4Y%*DXlyH5FEY8@4m?z--Ak(=OU+oalrGJJ3qqht4|B3;f{Q+Ah=^`&1c4H&4{mY
zgD*ag2VmH1g3X*!`|Y{Fmj!uD1-sngJIX_mIXU@~aX^J^r3J##KLnjWgbT|^+%H^p
zQZvq?4W~Z`6(@sZ)Np4&(7N1Dn}M$oJ<*<Lk)Wr-@a$weDUK>w(D*{$IKv>*?`H5J
zikiFtMG9C40?Pg%sgQF80+)qj3}_LLP`UL$$3VzlFHjT??@PEj4G`?uV6UyFrfZ``
z@CrAdd_BZpFyl#s>-9gt&EI6BpI;lOzwzOdT_jMs1t$XqTPK1uUBK^a{Bz}FHC1*C
zw>dgK_Q&_Nf)(D*V3Raekiu>J6zpp?P}u!hE=TP9Cla$U@40h{Mqw%!|KWJR3W9m=
zb++(qsB=ApFSs%t1!Pdx+`P#9J^>Z1;WwROX;%%qdU18{fwK|7XK}#GO_-QWpchiP
zf7@m~a4)FnuW(_t0vA*k9=_EyE*LHzU7y>)-+s);*z;3ot_Pi$?Ba@pLquezaX>+<
zQegPC{(>ZMK@{*C^dEBQEw)G<LIotNBJD%=pdcMAK{PPvE+@B9d9EBhSep&38M5WC
zQ-6N!&8H91pMox+TV!)Z*$fwUQt0`2R%Q4IU`{hmK)^vL1XPIJtM_=zew?@N&1u{Z
z&AVx;Es?<dQpn3!b;!gtU@yQ3v<<F>MH?l4+f7$^6!C+`%bDxPPwOL2G@7QmlWJcJ
z(kN3anpC*JM6(xH;YId=%h!=J(?BQwc(&k(r7{i*_TWsXg|JTqm`h9kS4%I~BhWOn
zNHFMQn83+Nbfa9Y$#um>|3$Ejm%j*Jc^m)DXWLf?hr|0LeLuapNnaoEwh%z_n&TDu
zhfkgU>dozVxp-u=c#f+~vF-{cYcOTfl<ju!bL@XrklKW4BxI{?B%cx>*4*abE`}|9
z07CLXat0u!-C(E_xHk$o@<rvUR-!z#q#OlI1F{Zif_BpC^7vfT(QoZJ_~_)_!zs*x
zopUP}a;(BIA+2Y~rB_!4S5^geZ#Tkqi_m3i)1=$CwmSC0YBL&nPNy6A&Atyd$KVfE
z7U98j>0j2ywIxqKA(p@5jDVUGCWXGqhYSG?%OTT%?Aw$kq2Js8&-tk7&m0duS>h-|
z_K}a$lrr%5MP_@5%o&K?+>UBj`mXUPa`Q;Tspq<w<+*`;$xy)4i?cf!g{aL5;t=e9
ziSiWk@s|1H@!?xP8r^59d{3(j-^|#`;(662GL$BbHMe#vFY<1NU7Vc73$TR<Xg*Ka
z)sYCGgml-r0GwzqE@bwm>f^C;L*D5#M-$7Mc}aTQpgSwz9T<(DBOxBbA7bk6<A$8K
z+G_uQl)YzEQ(y2dOpFkf5)=_B22=#3SLq~(9T89zQ3RAKs5FsI0*HVp#X=JZApQUa
zK@jOA6zLF(^pa3RPau#&%6ap@_g(9LxnJ(rv(7rR_srh2=gd6wY*VsYi4oESP<vj0
z-x(WP`y99(`d*i!uLKh}$2?Hi^_%@{jC~b(*@pmoiM8)Ax!*=L)k_vr%F*aegx#>S
z?yhWS&qeNRUH9IIi1;M3yi_LZGNnZTjQdfND<75~Hu2k}0>7_=Y@V?{wC6z-Df-HH
z<ahtW`~A9_#g1L!(X)~bj6Ab(fOmskKtJ!IajQUT=6?5JVoMeLouBqR<Kmq~o{rGC
zjc}WG{PBJqPtLNxIT^}%d~WcsN)|l9kPZ=Cx%oA^?rvt|2eDMkws-q)-ujv-zX4xP
z2lEzh)aKW^#AlnjE!b#E%B9%Bb8dE1zGddI!ogahfbtC^uNMc9cP@Dshz$+@FTK?3
zmRW7%ui5-gIi}QwpQ&a4nOaHr6=XjSNjoBBd@cCAdFO|kWaVv2=P})c5>|%hOVr~t
zQck#&N%N#rzlgnkgWwd74-e}sw=X(rcgMD(&<LP>Q8>RpI*+Sh<!OvtjNP7XmRHaX
zKEEWVl0W^L*X2E039XfeORb<K?tWdWcQ@;bG9JhW0V@bSES+|`G7ZpS6nZ#N^hi6a
zBaP=UX#1cI{ciJn#+Iq-X!xB@v{us5P}GQeYRP^IS8`Q!3W!yGy_ev<%wh|xA=6k1
zy4GtdofvDdW48fI=yO)c$o9?CNW6kCXz5(<w+6cSCo$d6eiMeP?Y=%v$3vV>N}arT
zkHOS->*b+_x~3|?U#SfYhj9uSl^Mpt_THHRfpAa!@;;HX{_S3=;YwkR?#k_bNf#nc
z#6$qH6$|D+5BD`$lJ7kLJSD(0N;DQ2q@(u*Tf4sV&$Cnr-tj^RR&!Hz8YUAsN5;*p
z+~ceRm7i92Xsjm3c!qxnE7JP&e#6uYCi3NdrUW~4+mhoD$rhH<qw`GMK4Bd(WbET4
zWWb{(^XkYl_8>QIo5NVNMxYgJbq@*cLYRntyu=J-o5J1RvB2dRfUnu{p+4um<ml<C
z;-nSb9s@;_>eNnd!-oZ14D&{bQhWWV+OW4BAQw6*LcrW?3axvMasH50=@fCm4bhMe
zs%`((d97PVofF-R<QYmU$>ibIg_rS_50&bY8RDN^!Y}_7->4lbD_Yr*|5qG7d?WmY
z{Yt`as8B!ZYEFJsYf?1O$_n;-*nP3$T)cR)G*WjNE8CrL>S}PiH39x^XV;8?MUY7y
z45H`0Orvohag+eChZQ+n698n76z7FDzSDF9R<eLZ>L5+oU;~>(WT)r;S)wu7(UkY8
z<xb$hW*Fae>@*nS_EF`rgn^SLJ*luEebC40ARTg;Qy1~+hv30bq#82g(H9@KBR%Ja
zg=CQ|@nPK){I$a=W%u3=cjnEk_|4X7vebyB-%>s|$am5mFEm=lT(!MDm$I~i>Y%e_
zLf4*D6e{xo`Yz^Kk+AfO4BvSsPdgkhd@1;isZTB@4OzP}O+Rh`g!VVdx`R{LhBr!N
z0gcws0nsS~jzd?Bax=QoR`Z36Wi?sNfU2u#&~4x(E){g>RQYSn9XHFItJJm(F0lYh
z-a2CAzKbkil1v#hWDtY7xu_aWg6^I8>pN;i2AqYfcN5{EuLu<ku%ZD+nGSz16qI#j
z7<L7s%nL9FK$NZ=ZGJ5$#ff?O6qQ?jNOd9_!U@hoN?THDb?8X>ea!1UIJY|TRS}2Y
zR$U@}o>N=!4uk?q%g|L;{M{9v3*elCz*Q^&Qq-%7JA%caHhJO$qX8GPG=%dNi#+Lu
zb=m-Zz>M}_lmV46PTCF5_0LN<G(gWJvCr8#)~8`&wvmbb((YPL`l8>Auf&XljzCrU
zfKXZBsSxJ^xjGW_o+8ad4R2)KowzH?iZx`Uxgq<|z)N{xUPKKA{{phK9!zny3A#Ur
z;b9S-An8}Auj@Nux(<Jlac_~RG-NQxtqVgoypi~GnqfJI)aYnriD1#n($?EuBOibx
z<PIU!sd8aMr^jmt6mRkY1*?jjuaC$y2SA9k!LqH0QH_H0i_WneMOpmQUBG3s<2kx?
zxRe&*v7$j69na74e?*pcc!O+8#fT7|np8^x=Ba`&YJoeaSP2tCJ{wOBNCYGhz#vNU
zS(${&BZy07F0}s&^^xryvA}I24vDl#8)w4x>0cJ)IFAY5uDBRX!94K#BVgK$k$K8T
zw>-*PxRT)B^4@`?%JAj`hE!M#$JFmj_#cyWG1(|2D@c#9gj`_F_yfp|yDZE+Fr&g!
z5v8ad0P>HMEpsA7kzdj<JsT6O-4PtphKq{(z62blS28f&8o)$7=yGhD4<|;56Ii=k
zM%keq2M)hjz|D$2y8O4p_7LYO^g2fm_}1&&j7(2M@=mO;G1PRo#Yn(>*%!L}4x@~W
zOnsZA!Re6&{@&tTA8ueNwc;%(7*9I#5JznwMg*4&pbr2YSirOw_&eWN{%Il#+48^F
zhb+*n$l*9}G@BpV?*gU$z@k3^&?P$xagugu(Fx<Bk;MdbFCG5~1aRM4I##!7*AG-r
zzc<>bCqz%`RI3Bf#SO>9980}fK~;m7(sUdIu=@bWlBLb5<0*;wgTgHEl6OlGqw>;b
zNDc;AUBq;Sv-$|$$6!`@U+539z(F^_cW+A#f7l`$!qI9*eh-pw5CBZ|n8l1xG~lEG
z3rZ<^y$$?aqf^$Ar`}6i*D{e&0JDMlcMJ0lxRvZj6>$f|=d0&W0yP>0coYD8O~B)N
zF!r}~yn=}JAK{^#q;S0L@_u#r<+E>I2Jmwit<utlEh&N@iw~wM#i0p{-U9QO!%It!
z7pn0Qk^o!&6%Lopf>1ZHXxbvn{~*v4>^OwN3Vq(6{#TS1-%EPy1&n|6qxQ^LHZp(U
zB7Su@I>ulYXxRT1ollTZUv7nAF%&ThGWsDsaC16K)eYH>2b@?)tx7x(^Hbd`&sW3f
zu334*7h&&zF0%NEoF!E^j3>=p8N+l8Af>VEFaH0Z+WzE%yk+}La;6)T=&e$b6C|CD
z$x(H!oWSJlAUUYD(c%CK+Kze4AZ`BwJBJ|$JXfD(Uxj_5g{w0sNX>e$;2RuY=OvP<
zph_TBj74c+Jtd?1FI~nDk%A?syU|P8I{tr|NHN_Z*A0$W4*Fg(&TFE<WP&iD=I;3I
zNfYfn(Zro5CKpc%&#Fgz(ro)995{)}cIuy2s#UV;x%UOSXXv#jZDs4eW`+FzzU$7Q
zd0Jq;n0&iUC8-$g6#0{E>(Ff2Ni_T8-Acy^Hs5DMCXKptbu_%;I15>%$=T}dJs-L9
zy-JT+`jRD=yO0AXam4ePBJ%p~$dYvSAI*An3&srtY$gPEzrrYepTTtjqfB;ewPQ}a
zf45jVdxzPiJu(}3m!?Duf769OaF-Opna7J17EiDMD*aME-gooi*jaohX1y7xfA|5J
zxplYh*}kr5<t%ZGbB0)ji;%RRV2aqxIBsv04YKM3UYK6icN|(e;8Pv;7e8|M;$KO{
z86a@hu9n8?ai80WqytqsfQdGq(cR0Qn6Gx4s6y*mkFZ-^X+IduF<A~EZJXdbetG46
zEDw+T&Hn>xQ%(B6qSXE$AO3Gpn<(*fyU0{z+pOM5zudN*ouS+FXV~Mzp&BwFzgE+Q
zg!*<JE|E^iAWlUtlzcQw`Y7LJur7EZcQ%>l@atrSV{4ly0v`KQ!(82(G7Pn7Hx6y-
z)gl{`1mw1txHu?dVi=`QK(Asf@~oGNr=X{fr|9SGmG&F|j+DDMm96ai{nJ#O+B@e|
z|A$+syW-<UI_Td2;-9L|wHCbd<msM>VYHXwozV;BM`vUv-%jQ$_-l>HvAmBye|;n0
zOVVoH^{)LRp>Nke=IM<>bw04)iN6Wq*E#t1E1#M5KZHWVxbzXVU9-K*dv99QLh|d(
znu#;H1bf>~_!#fV;T?b^*a?&vb2fNBZ-+Ohjzbki_6cy>mA#95ue)ntR!7GK6`#b{
z<<DLUh<v+p%l=ex_g6zi@bR@Lp$cC^vR4lbdKLZg%tqe*Hbf~ytcn4yCAg`e+%&<Z
z&ROhBe`C*6VwC7?Pn{4a(X>U#zdc$?hu>m@<aa+ygavFZrvOpZ1$BW_=(y012{`cs
z`sRpDN3MbGkFPfaFY|VOuk$}HD%0bu#IrF0$yE;KtB?l$_5C`IOP?#svyysm5+@M_
zJNX)RQYI)Dx<lgSb=*ZjbD{kK&M^8_`5KzIX2Lo@wZPQLd?krh3pp0TZMEFH#L0)$
zK%WeNHL`i34Ioi#Zy)GO0?cDf0BTeM6`&KJnLfYr{Gzn*^Glo$(xW{Q<?f}f5jBw!
zow5(&VH2p0q-YU4{S@oFs)HIFEmS=D46wBY>0AT%zhfC)Tq-aS=%0FU>Pg+XwOr+^
zLW3a1mf47PaHin5&L-IifltEFw!?P9ePLFYDhT<>yu>Ql$`Kl8Q2fT%&}d_ZO8HSv
zJa_;Mwm0JH-n;|sy%v=Fbw094C<|%W@7MtjVYx!n>AB)CpV`Sb85X}qt9bphJhVh9
zQG{6K`~VxO<bAb@YE#jlt7TGo<$>#ZL5y$2y}c*qtNLP%oNk<q(hIGBOT12f<k+Ft
zhdYr{p30uu6i`>x{obgR$J1c5BFNhUWxXUjpC;Wsgozts`?3pKQ{dSrKzI>nj0hia
z+(u-zgFir@yfhO2YlICOa3CWLo55x09n^bKv3&=0fbUa=d26FWpWq>p9`)_eHb~Kz
zOJ9nlH!ijD>vL+_!646DtraoR1W7#ojqN92U+_P?VK-gAvDHY1iE=ePyvqNC<zb=F
zl4;ep4(f~IXBR_`s6*nQ&$=R<%uUZhb;zg>U{OWrOCoH=b1l+u3$zx^&j-`Pk6Q2<
z-8^xKMcOcZEQo1L=#G!@$(Jq?^ZRqwIsf^~Qp1Ys{Bnos>5Gko-ihlJ>F;Nq_C8cM
zv~Z?MF}!2>7yLMe9@3pgWFZQoB`B5G*NFeg6;0?Ra~PrV^fCezl6$5L%4KId3mh$)
z>va4A2J=F9mjX~lg_B?n3258PdDbuaf;H}1Y_#Hc5TjS`p^*MYGCTN=DlrBr&u_Nf
zgmWTfZuFMwu2~}dbF`=7UNmL&MLUN3Q`YmLFz-Jr=})r9pf|O9qt28-E?bwtjys&>
z5j`A1xF#9P>X2b_Ykd)13ke!q9mRKOK`*Tbqx$$>npg!B0-C|kr86KNNw0`9Ydm8S
zYY^gQqrz@WsqrT*3iu<~c&^jba(cFq?6_gU`En8{$+k+>6k%9iwTRLDz#IQ@0OIWr
zeMf|0mE%sG3(XZ~q_s05DLa=I;V?$h66|C#74cy)viU9K7&BTWCsf8^0rUVOTWpM_
zV@?O5icWow5c++Y8WN6e4bI)4y?!Bf#`vAV>+pS3x!tIbp75n!R&G#>sgW-eb$Q2;
zfP8v^h80NVy{^oE<C}7v*cw;03X@6whZ1`}T9IUc!%kvGP8iuB)W7J0>hw!cE+Bs?
z(xDgf%K#QPO2JBIVdg=9A!Ws32xF&COff7-T*CZfd8|C_(4|zpv>D@;fRn~3EwK^P
z<(R2U%v$Oo()}=J@Jnvw4xV?Z*uNGs`?)e#`>CSmk)8v^u(rp>agxy+8pL6_XiB%u
zR%G=cSfm!x?hCzE<#Z4SIt$%(X5EFYxUEF`oyouyQ0=E763}u1k2eqj&D<vKgIj4a
zx1l!HP^r7iwE72dbxW(1DoeMZA7q?7E6^->^1v?ENCBe_Vc?#aV3idtGUY;ZKaEAt
z*Zi`|FyPHv)x!F^N>rfFgko<QB^#J_fFs_(iaz-x$cSgq9B_-huu5>QR!HvGc!-sl
zfD!p}Pmt5%0CfPxx-L}yz$c2-^Q%!`3{DQ0T;tvGN+5e*;=Jlc9;F~A!|tB=B!|1S
zt$Z!(rho1k*CRHry`~(k6wG2|v7bt^akAn_)L@DHwG&Lx(<BcW{8`UTNcTDDr3BWu
z2Qu76^%-(Q?ROTWGf>a@HgsdIAMw2T{4!_)5_z-p{_I%2(s>j!iVWWj;JC1-ytg8E
zdJ?Qj8qx(f?9{)wg7$daOy`eJO6qAX=PjQBCI}7ts-ryfk*f;JAj56M_YBAzCF94(
zmJuz8#j>E>+Z&MIusAoELg9QBijMe;PH;ypg1XTw32s4%Sok0ai09dpKI7ejrpu(W
zdlwu!jZJAG>4{3+l_!9=N<Fzz+&rNK)agV)Z>lljnj=rY7k>&6Rt5Qa113}Dq)Fo$
zjsCZ7v5SDehlH#Jp%{oIP_evCaBgE0q(5{XL=i}C>X1i=9XQ7MC64u#$T&j^Yk|LJ
z%EX15?tm3(vXs~7=-h+~?c&l1Ed%mXc@8FHS1xR<jt<E^?vKrBI`myr>@0Bkg|U-7
z?3#$bQx*X<1WhEsto9>sG)m>NWx7CPwYga#xmrYTLL|S+3g{H)Qnn%Mm_$!&?oai&
zQB)?&iY4A2j%`d9>qYWe!>u?vqXbgt70&990kBds;P>x#Z`8R8NVjSqcxr@}yWh_2
zM4s0x1Na!24&!nJ8LRy7<|x!j#R--!y9=HM55cS}LR5=hZX@7x4<ZSXCyk*mM`>8z
zLGYymKr9(WWeS*RisVw2`%n_ST=70OC-<UCLfgdhw8fM=;La+0h`%GBckOw8KzhNS
zZBLj5jNT~9XQi2q?0qZxI}J`qhYVH0(uH14fJrMMSXJt4momMSP#Now+}dFQJM!co
z%lo(fdYOJ|2U|bk^Kj!*ng*|caC8j)2T`4AY1)_tdK`6}qoXYrwbJV$$mvK1Ov+tL
zh_L&DqN{D-+s24d-BB4phwd-x88Zpi=0EZ4L2_dF<9(^ENb#c8m#PDMh_{JcsM<By
zNvC<x2INUP#D#Oo`U=dWYzOrlzSxcThpvPdO7r46G9%tlpuS6L?g(|F^y(vy#xa1s
z`!0ENGGrIDjg^=H_e-wl%7ub$@b*eUPTvTT=cd8@IxyjV&}Yc(Xpob(=WEXkB{!v_
zBnq|>(iSQnO_1))EKpfc8*&9Z<{^FG%)MELEzmi(BfaLzkp-YA6@Pw1rB~0veNjia
z8TRFnqOOUk4c)vr$f&94He%Fs3v?TrY7MnH3UwhrNX}bCMUNeL1icpeHy`rhjp?^;
zaKnjY2gFij4%gzv5Iw-73p|-q-beuDfQVllA-he(g_z+ZTLqi3%x|ZElWZws>j?E3
z>GFd>)Yo0*WEFqO9eLRJO=G7HqoUfXQw;CuwP%Ci^%P+8Dk0Lk0-{(9nZ3<*<vTpZ
zyy34S$|0Xb;`I^619sWq@8PHCGV@j<<y=sMw0&RtMb!6e_rOOk?io9oa;TDq<94@q
zQKu?%m$_%za%AR4=}L6pTgbH&)56B(S;2(N9}uZFu<dVH`eL1}a<I(x9jx#cXjY`1
zCT27sspuRJNr!ej*DJx|P8ehT5wp@Ceu|rmlS3*su`8jEYJ(IIGj`t~3IjXPN*N&9
z2{-xID7p63ckj)}OvBK;W$KG95ONp$>MwXgvKackVsvnJU>wXp6Dl*(3pN|kN35f`
z#<~1I2lvXEMa0!Ah^h}X%lpPD9~|EMS}2eFParHFS#uuHnVkoow;&T*fMth83^ypH
zn)b~^cfydfCydLZgJrJnVCCLJ3MQN~=0Td$J5oL;LlEO<^Q)Y+DLfOB*Cc*Y5m76V
z90S-&-ZnyP&%UVc7THJKX}7V7x`xYZQaI_!Z<z;)I*Gi?M=0ycv6?-+h-{Z&d7rw-
zVK&DH<X)hGiyA`@jln|*{SjN>-sq|NWaE9A4#(BQm!NT!xqvQk0+;Jr23DgOjHq?U
zrDB8rg1f$sbCNeAPaD9jLN1CyF9qJwMC@QM{}Li=E&1^Ut%3YGI5<uDxH~TbT&t=;
z84JSRW+hp360=&Zz61&|Ai0UT5FvkQDPbcl7Z-yX5J~4+Y$dYxr+)H(3!p5)=l{SF
zYf{&QsfU&X2>y}N{K*jmo#0cCv<MBiPWgNUUBP?(xM$h?zQCA#H#l7b2;<+DrvRm`
zRv#tfSm0Vn;+fD7;H%nps6Es!UxR#6dkDf<rjIWlq8SgN!$34tbdXGVQ!@Ap?4*2<
zQvr8hgSBCC>mBY<vte?=#xnmvXEH%=A=B=NVY<lQ@IQyKC-UahLQuW2_vS3OjvDX<
zjDP(11ot4Jsie;^(V)8mwU2UXHa*8*1bH=>gV^Tj0&VbD(!*BfK~0wFO~S^A{m2PV
zXo1C1BlXB7MECvjZ||1J4jXMF9IGHt==L5TK<|)DIX9F(fB}4YNgzv&WmQ=DAz-D?
z2csZ7fqrekx>-ewvMQ{Bs`L8=pICfrC-V=yi_me4Y=WpmG<9IfKP9ijg#8IS+Rb3^
z9<4}~81(rQ#>w>5b}-u&Y7-X<)-a9{5H`+rL1er{&N2RTMNI@UW8-KmwZh??xlsas
z|9lOsUKI8zB>)+<;|@EtY?5r`lf{%lP!TG{kZPJgbc5@;Egq{18xaaBPj6!nh(K>@
z(h&3Nwoq~g_wYlJj<&~%=Vl)^G`8VO53ha<s^sy6zr<lizpya8x*~x!WLqZTco`%z
z3Zl-xCAAD2FI)gUE=wMb9EpOY9){g72$Fe}&hG*xQ_iNzp)5MUpX1F&U~$qiL5L!-
zDomIWnhfeAAYH~}QTs^;IY$owhx!58Mu&|)7AB)+I2kh!?u}}=bHg!NX_vS5rjE_k
zTt4ROOv8)sW%eSc*RO2_o-;|-a*F=}xh7;EY)eKBK_`nK?ca^L`*kJKl<Sb@zZndC
z*(r-#&HD+q_Jn>CcR|?+sxw~%zL4S{ARC6L4tz%HVfb7EIC(7jIh+@0lRwG#R7caC
zUVl>LfK^b`x-2^&+=Md`%8Cf^!yAOW35woO`Vi*x1g5vt7#R}<anpj@=QQ_!^n;Fb
zMV}3wg&yPf{695PzfVQP*hBWqGpwj6D~OHwz^;T7b>^qzRxAr1wcp~@jv06dsC(^T
z<V%8&KR)u!lqXR=<b`+^z4A=@&Z}7=RQJYNo5U$;0zf^9m^ja>%o~TqfXB_|OuNpr
z!guVWowe2C#J;~MS!l$Dg=~(GfbXP3Zl8l*3vkl@!haW*yoN<A!cp!h8%W6c#$0Xa
zrD7&eCP>T($5TmqwHno|3DmXt^uy=Zo6&ymdpVkA0^f+oC4Iy3ag<p28#Nvv;Hy)+
z;Y=>F?7~3PCT<Rn+;*l<V36oFIFjKU;vJ0M!n`@ZfPby>gGGQorpPEo7>3!sHnt2n
zxyQGv<mIzf(r)tMAKzN!%cn?>3nWOP7PR2je~l=M5q@&5)$afDbZ;R~0@iqz@PD5R
zTDha#S;W2-^L6+H<b}eK*vXdRHGy(jZXxk(T8MoiK_C|Lt=VV0?8oE3`OTlSCZ;^8
zNGP*U*N}7MPpLS_R_#Um%|q?1L-{6OO=M_YPfn<F{0~CQndTeCAMK50xFj3?-2PzM
zatA~|PyV>cMLx2D@V9=HW-;^=qgrbd+{gJzy0z8CfS8*3>irYDdPH2<w2c2@0#Zgq
z|D(s*TG#<}RxmVVK_cQfuvy4OxyilTI5dcSXyMgno6bKHm}I6?i%vBXTkV$Gt2*#h
z?pn6&ZuCmz;-Ccdnz;Hw#Ue-$U)zpAoDNs!d9YClh>IAzzEJm_u6rEaY>+#L_EzES
z@uv3$xXbLdoH#jz*>q+~`4{d*RY*bx15ay1JUbh$7a7-F*a=#uliGiDKPm1;yU1Nj
z4FO#c84o&Dgf*}-T}Qx2`P<j_qMq=*bP>;Obu8mcXJ^a3TZ_UkdE0UPIf8!+V+6{D
zksGT`vf<fU2ZTr=kO%Nf?Yg#VPjJGtRf)+vnoy}c3W$Ol;vXOWPcZ<IbGZEw%qi6D
zf&)ac&46IdSF;L^;jU(yxM@Gnzp!hJF|k-*PKuT7@n&P;8#KO7{iEJHxEtf9q4jn7
zAJv~f(Jc!Tx^B5ZWH39-k3*B01zq+EBauB?&9{tmS9x!@dxmgBkQGoS=rH(mu+<mP
zPK0m|Pqhx#jPv!)D+0?b8Clco*o9K>2h8qQ{xf7ftd2NxL{kW1ab0+Gp|Q(>iVb{V
z+FdKC=YO!0IFUkJiVNUzzi3zrzT<JK{0Eh{h8xkv1)$s!Sq^+fUWg2PutCixsFwqa
zh@P6OGf*NYeY$J=QsTP`76`>fKr|NuvAHJ8$h3~q3u?yD{(6v}_XmbLYD-)xfbb{>
zbmy}J$_=ZRC*;H@t2lOOQ6SF0kEd0_r}+_II?^2+f{6-Ov7>ZiuJvIe`wz?#0k%8(
z%`3g(;4jrV_f#cbugf3zyvDb<ae#Q&1ZtlV!@7QoAfD`I1(|k3sGS3Ahqh>gr+1Kh
zdGs;SU^M3Q-oHDR7q9$tqkrF(b;DYlKzsB2^RnHr0y&@tA^R4C<TD#k`BfK0!za)a
z!Ik2yqkhGpq(peeVb3s@g^rhB;0A9npNn^gKza}L&ClLtacrpRZHQ`PX)*Zhl0eaM
zEtwanIzD0Or5{|~{NbzF<*``66I#%$h=Yx`;~n_>1!BF8zdvq?f26o?JeV`WY0B7j
zp>AY?%x-ET9)a!Yk}e1e<ExL5>=<7stl3i<X%-`u^A_)Xs0>7M=Ib5q;%So4?L*0?
zf<DK$#z?4NYsH8W@fHM<9dzIfOVVVg6e(gMJ{IF3FQb$D)<_9G{V?Qnw)AH$?)euj
zaPqQNuI8X+lSwBX>XQw&(Lv06-v&Q;2aud*D&`!3oYL<?WsN*E{=`9lzL5zIOJeRo
z4!6bMesJS(yg0%O^$9d7F<|)!oL~QCj3+ysCy61pzK1m7<iy!q{_Hg1dCwbZWs&nG
zqZtQXcSWqZSnwrF$hcyC7?ZhPTz%dF?}ngGEITePkZ$23Xj92wx<jy@71CMFxOb04
zH4*h{Vb{l@gz#K@unp_4MD8%yUV9Cc*^Fu1XE{c!$`8~olo3WQz+*^Gq|Hs7*(EAS
zvrBww{-c#yPTq}vK{qmG!FIy#^oVkDx0`7ysDQ8Bq+Ij*28geIi}y6rS4jEJ9?!n9
z(=Z+Vi%Ta0>}l1|R}HeRS`K<OFB*_H{}~HMgq`?(SuJP^TudAVCx^>uf*l;&H4#N1
zKWW5I{sPZ^tjj&H;u4IY(vNR|C{S`C&yI55&n^MI!=<OSUO@KAW*!L6VuLMsN1hmw
z%Tcxw%W1<TFd9{p9{2*4Z*#=X65nRK@Ga7ydW#b3Q@PMZ)-OJrOe#<KK~5%&fX)Is
zt0#v;t=6`(o!l`yj7Iq4Bb!}@ONuanS#RtHSn@rf?!!PUq1V0{CkXIK`3a^wTY(E;
zh&y1TEl{sJN{*uT3kInCxH6=rFvP&rMX5(@8GhQ&Jvs*0$FLK`B37z!^a)bhfPa+E
z;zU=*4|*dwAq6-SAXrmk3vSQ&$FpgIYQy}%I!=Kvqycv3J4t%`s)zvD1xMkiBTKA!
z=&ig1tHD^)@IIO$?3)RD<_W4>CPc1C(O;*+v0GDyZM?7i;I81c;UA!M(DmDb>Z_pL
z5X6i)DjULdCjetfKF{^f2+$#e{*jB~G5-HicP8CgvVNyS)YG&18-BgnL%kI8aRjeO
zr5mNvc+P%?$wb7tvjW_CEE(t#G8}ww3W(pIl3=}P9pPqN5BhUb%ICsK;qCBZdf8=8
zj~$=O$_P(9#s)0QaXyKhGXbT+WZIwc`DBAE7&Hxn;}aE5nhE~(gv1rE#$GOc<b`8Y
z?=MbAIqTyAG6s-ioJ@(_HIQ6!2<k>HB$*C<wWMaG^?F-32Q<AC=t+P044}n(M~M8q
z^1O9c@gIr{E&~dMZdD!v_Ojq^i1uSXT<LeAGMnIcqI73ya_%Z<!QLH}Iw#%T53Kef
z9hNQogWb<<v`r|sWg+*T9)cSTOR&1a+s16k;izv1kc|@9P@Z^X#Q?aDOk&a*p}{~>
z7SMlvlGkea4f5{6u;WH!1}rVx3W1tG0*nM>tbHU#9rEAJ0kZThMkQ?EGo}vmuNRfG
zlg5O5QE=4P7_9>l(kA#bEZqwC?g0TAR+>8JjWG`iv;;$Bg5HGafm!_Iz%-x}SWy2#
z#$TfGeE(T?JcOqhvN#s4>O}*|eP^t1fiJj1bBkGL-_d}QZVa}YH8voD2t~d7#%)ZL
zBUJnY-x}FPIcpo|zT(>`%>Z5Ic(eQ|I3SaZsm<5|4TL-0g(_@;S2n=n{X7PtJUap&
zZRioMNp<OT_%Y6Jh;S%k+dh37yyIo+2ldbdce8Y~5DDb!Sx#9eX3MJ;$v*|oH|Ckc
z*cU?nIq-)()B@AHw78LuA%p(#18}3;aTUZcN$y3~x$#LAYJo$rcqb^f124&{t}sRv
zLmK|u1Q8Ew5ne9Y2OaqWy2<z`F4Imi&i#)9cgEsz9hhR4_aQCRE9vAN<QNJ#8HuTZ
z3^=x8m<w}A029vo(u<590IzL#h!ZOGz&Gbv9sV4afm2jJ%*uS5#4YC4Uf}vQE_-b^
zX8*rkW$R4=nO;-~OYJFxneh|(^nZr(mZj^*`SHjd;V1X|JWB+o1sr<?csqAJl?Gzx
zF}Kk03{V+CASn1=&(2sa-+(>m{W%kkN)NJD2{{5wJ_<ZGV1>eObKbG-q=Drf-0aUD
z^cJY7|3G{wq8nEJkn=0)EvAk0&YKO_1}yeN5ib~)2u_O*s-q1v+kuH;cxz99f5VU!
z9LaDN;L6b5r*n(iv3)>?Eg+c!gi8Y_mPrOcZ4R%u8!&d8jr4xSp+v-cv#S-5mPBM4
z1J?np_rb^Onee_^&cXr^NB1u0aLSSZX9E7%7?N+@0r=s;ZD;2Wy5BLBZ42zfiD<d)
zv09FrBXDYivkzRBMtW@{(<##P40K*jHoy=g0DJmC{byRl5JRnzf(#*I<Z9r+5|!m{
zzAg3>(B7x(zQ;^3aJ-om^pOAKDkAdgAtDlD{DFdDhlcmjk&m{3eAbwRBU5LbsbK<;
z4g>$80qF}xr@{f~1mL9{Pw2hOIcmb`oA>quVt0*ddMs*zg~4e+%kw@2HHr7*%t>(E
zOa9`cnMjpZ$0h)mOT&DJGbaK_z)(F<7E6}GF5`hhGBlF!e>8-7u)@}M2Q0P#Z9H>u
zhkIvJ9|oXpYIqNYa@qDA?RU@=wViEm1LPbB$tCvN+G6|oQ*G0blq~>#;TIDwiZnL?
zn#r@4eH7`%3CmDYEL9ve*@$1Jpy6MoU8k(1N``@OHXTrP$+=AT9+;pFhI05Zv?&*=
z)-2N(Ln9>u6})B#IfajC$hW1fK%^n(UI<4doHKTWBY-`(UxnXgSYE*ao|BEpExPv^
z+(@H$8+_;(3mrCpxHE1dk%C@kcr$*o%;4#BIvjM?XebBWpCLU`={Rl7Z1tdeiF25L
zEEZX2%00lBqTiH<_Gmz!Vw(d5zof~QZ38{A2`fz0HzM-hV3qgohT4KVLK4or%4Av6
zDlvwfLZQA~8dZ~%;z>O*XL+2qYA{AbdNRhYwB%SgH&Vryh)8oBav6sVU~a(YCcH0f
z*E!}hXC2uah5x|;+JevSGxpU~BCy)g!?O{BLkfxJ08)p&(ggM<-HibGB?j;eA8QOu
zq;MqXY8~e*;g7s9J%%#&Iqb0t9Kh+2|KU3m<`K+sx9kNK4TW;*vC`{6NtZEK^hcA2
zJ$S(H5Ti^uhh0Ju5@uv3)lYKgVEJ*P>@W_nY%CYVjA?DKs#6<d=%h)#d%*yB4z0Gs
zojja-Ej*rP9lu+|El0IRAL9t!H=td|ML5cqg}1sRL-#j`awg1?ckp2E=d&Yb^v3Kz
zM!<QIx&!~RzpZDL1O||h2PF6{5#gM}L94d7;3W!sN~?74R<{s`Q_VJFidE8p&Tvcw
zbt(YO_8wyD%uV3kSjg>`q+AJRaanh3GBW)e14FKqMuaz?r&ql03#r;QM#!(ghiQNl
zDR6aKE@(d9JB0h)U3q+TMK61Uom5ThCPnyqG9%J9#Aey1WsIj@7;0^q3XM~i^9U(O
zkja{;#81pR7Mx><Y~F|S@kAtrJLwrI2<DSe;H!FwRy&1W1A)dm_M{iO^!XQL?0y4J
zkvs|6H_?S>FD*ApTLjaR=KRrp?r<#*_n+?^iLDfh8C|mc9W;x$))o*U|6FZG#E4e-
zx!m`gG;6Wey>Y*mkNN^^_3D@>CDAm1dV#uy6Ue4<I!O`!;iQDVrABGDD*VQbWlkeN
z>D=n#p;c94su<=WeCwOl50O3VXtSY`u9*oYzKe<ResPJ)zlUcBO%P_1{v&bxX<NDr
zv-o$&4ZQAr@5@xm*kkX<oO{DL@4tew@;U(po0<C-6NZ1i8eibqlhc97Oj^Vh7O+5s
zzs`)3u^_p(<C!F6RRt&P$^^5p5nd7KS+<wa#0)Oj*>Yc<!T55gTLuEucJZ-S*A&9e
zJJ9yFy?;Ka3OV9;>`$xhUp?l>iKHgx%*QIyYCV3&72`{CZ-XyUO1U@L_|Ud#_p>Jf
zNSK|}p@8558tp2TvRLJvhG+iFW)v7p|Ditjqn>3RtlS8+WzJA&l0&fEt?IBX{-qMj
zN(x@}$W_iP68QSLDxgXlaaUoT-5Cn3IN*#da7kWp3-?WVb1%FC^IH}E&S&-#Y=mR`
zi2jkC(qqlp?bUt6ETDdzKy&`xDN4e`KL}sHUL%eXY&yei3QQ_DmJTkn{LbZIk0{<M
z<YF3KkukG&jx)75R(8_R<k$|r&h2<9bGJb)eA!%qO>947-E6-xX2NOD)NPa}F@3ji
z6OkLlB0Q<Uj(MS&C^p)pV*P~;ep!M4Sc}U7$jzICNQyw1-HG<mre1?<)^q2p$h3w&
zfFX?(n{1#rjM;hZeU&x>Gi+%Cla|D$^`6xUn4e?WcjB-EakwP>%Q)?vz>W+vV3zTs
z6A<0EV^Tf9cVB-s{y_DAXE@h0n|<pix<)aXW9UthrzuSL96^ai?1Rg+#P*oUwKdc(
zibH1(8t$ytwD72#t^KO$&K6^_igb0X$)c{~?BnuvJ7@(p1ye)qG0tEf<lK>e6<ar?
z&MI+N2yOXDevW7ojOoZY5|>!}mL;x6p%pSYXhg`g5Pg+`rx*n0_bJ6X<HY!?YPP!e
zA~Y!)9EImwg-xyg=(6xSynHG9@r@bFf$4>-qY))9N6UT)A!!qq^p6f`hlOt&(R%(Z
zR0U1BBLjnwqO%*gkYEmT9`|FgzHoI9@Y=2?{B;Z=N|Y6d@<SdczCp}1I@eH>hALq$
zeU%$x{%FUA>!MsaY?oB#g{NhXZfex<u+0u_I5U1L?aeQu;geCVS!soCcuT<jtFhx~
z+8IEaO%^Yz5$A#4Fv~RoWD$Gf3Y|t5*E4j-N%v#fm0(N{YvDd6$?WcE$mII<jZ;JH
ztlNF^wP*uVr=$RQ@N2hC=Ph}stqz(kiT0QJa)R=hW%!-Jzp+Oj4h{8am!7-$hd1^l
zo)v9ifUQCYw<WM*1w*oPn4VaNPxgSkdw8clw!OmYe&x$?ZcYZrUcnQRj`Djfk9`bC
z+W1j<p?JbEBa&6<HE!uLMsua!*s3pK(<5GtMd;fxtzV2brNSTIs)gOA&z_xB^FZe|
z+K1j3^5J-;8a0s~4sBU1@T?)tC&q}<X1w`PEBeAaOd?RhR~Hz(of7e;YoQLC{cq-F
z|BpW7c4}~205Tx8clLBP(e{_LQn|dBV=w#kzL%rHg7jAh<`T>ugGHA(+jZ(9%Q7{r
zVS_JzKI*od7wbnZygm97v+%2U<c=`AI^Z|;4ktL)3X%d?vjT)giy@AWPh=VEKtmcB
zw-YqvX;O5-;8XqK6<H?3+ooODOZGY#6Ar&-MW|Zh%%1LAp>~)*@S!c@(+B2tvx)FW
zIVJ3qzMEW5`h1hn#*IRjYkz%)OZ$gcewzcG=p|5wm~Z3fd4bfA6=UzIh+BUCzr`xH
z3l(;N%Q>`vRwvs=!)Hy3a7$+#WH7O?8b_&{p{<amJYgC}(y3=HWuTDRLyjiFXJnU4
z2few@=)i?0xWMiH=R8FM`85B3%M}kSX@QF-$VC3otC^{d>g!z_pY1TWGiisIL46IR
z3aX)!7aoN%r=pRfz|Du?cFtSB_&@&7{qKS@K&@n^U`Lhk(=!``&yUixb!_K6@4i0E
zbaIjE4I*w`*_@jP!bs{IrpP~tOW6@AI&cS>n!~Hwt9Qs6X~%o?s(VZ7-3WE>E8gjP
z-$mX@PaJJe3IflH4~w6`Z9+qJlVU8-h+VwLdR3D|QMHK~_GeYmP6ql}BxMl`dxwdg
zq^-N-mV=LR%*IcH5>zvt4JM^Zt)FT9a;>S&F{#muDzbGBT^-57u(SMUf3^nw$W<Be
zv9<hSJ#}9iLF2!bu8<u3*Eu`+uA}gZ;)C^H%fwTt{dz+yL!C1iB95c)4KL%DFBV(L
z;T`2I5^y#0Ix}wRelkS%9hGdeIJQXJ%-g;B_z~xDQP){!Fvu!A6dE4!MJwFoA!XJ7
zQHXKyOkr@RNy@LXcVhmSGSAJ*>A*+KC9(IFjt<PC3jQ+FL3g#5Qh1nHg~_u$V9%@0
zT#Wy9C=ZX6_5Z=uUH|(32%-DGz8;h{lf8<(J5q~+zbbDW@hY1%`29fiV1dVTSHFu{
zR?P;M3NrVstgR8&B{!-jYk{_^=u+CBKez5$@s)OeQvawgf0*yko&M{$w$j4!fr6oB
zRn+=XhfqEFr*I?8$E?%(?r@w0Rfr?NXP5)!Gjuq_XJ~Qnjm>ojo}47S;|G>((W}zq
z4JQQx8jlaM?=<&h!4^eh(3dyZ8JtGO_$*qa;Y@B0g+G$DpN1`q%EP5NDzV-Ws=HMz
z0?HEbJ|VK06Rfwx?^z!-4ze7_)bwV8_(@{tn~CR}-IW@ZQ49B@oF9Z}<cl!koz9V?
zS}0M(lhK>FBM}0p(v8G!Z`!O_t!M|&@)(@I{>h{+;!Kb=`B>PC&uG68Sv8Rw*xB)f
z^ylOI$<NbSZwqb=e%&#@Ro$(25H7}gr|~bQA)j0^_t3z$F?xOI#@}kPK%yR8sQ8tU
z(tXe@`h-#IDuT?>+tFED4_IWA4Gr$*CLG&o0BH~IMQK_hpfktSPP`tnO`1*h7h20(
zK1q(*4a%p#58^e^+VMNMIxdC(H-2Rd=k;?TF(;mfI6_@&cAC0#gdASqH7s#=ftf5M
zI%r+!5<{@Nb$C{CL7#UeFTu0*VzrAaU(wI<_#iQB$t=%mQ$0*WOStBzj(UOdaJXp0
zaqEl!ew7$FhzFMT&Y5NVJu5p@{Kvdwy0ZTc_M>UlW|vcW!1*9YIV;SSoPwwF16#-&
zG7@XaMyJ;Ej~IKrey;G+*q!Vg8H_wOaJTyHgw<p)@!U+FeQOKJv7ou}s98zO6Syd4
zDz6{ECdDZ^K<67XBiJ3&KapdW8tI=Ket!BusW}!9Ip~9w{5$lzEX{B7T+hM$cp=f1
zygLU-8QV86U9UJJ7X8j$d_+}r9RJ!#UfAF2kPfZn<(>SXW<;;=94wxcJ+8O)TTdwX
z3RxRs4{iRnbO`sn#eeE{%(*Cez0xEtz6UT7<$J>T>Kib5@n<5xGZALbKb*TF*&~qq
zK2g5v!uQiX`q^@p>z_RMuIE_a&m%v3J#ppyEgiw$vI^<jRHx`C@0S<uC%TI2|4dWz
z88ds{sTsL;e_J5pBK91Zb%QTcEYfT7D8ZU{J6!3n^BjfmVM@oRhoAov?00I04h}S3
zGVK*R)mf3O_|n?7dVc*DlFuyX?VSfdW$ISK+7TOeH%=JXw$XG>FE1$NqvJ!pYT?J2
z-|`ZdUgJG(`lp^(c5RBqQiP>?PLEW+><zj(XH7Wj^Yg&$7wKb0&mO0EEIuAQ+~@RT
z@hJFP%7yQTb06)zHvI>ASV>qv<m(9Ob9(3ctLfyQD+X!noyPC(mu0<FPWrl8j6M>d
zwBXuOx^|y8tW~PVu6N+n52G{p4;l$H@%DX)jk+=1;)ED_XXH_MIdX=U-0qan-FhV9
zOp&UK<h#EHn@Jr-bt@>MhmmP{8+M~@vCHLOV>!dv*-l-!Li?ofy5?%yc1`mWAt|N)
zukSiFq;W@wXl7F6Z<Wc`PeV`|xxcWtwEkXv-+ZDg`oPzOdxtgLf1ge|hLGHR_5FF4
z>05Ge!r>x@`*o>_tMx;IdkwGry&DeNo97%77w8}_{Mwe%zh_@9lf;v7@l3~C_x9+h
z7^&ru$QH0^Rs5X^E$JGxhRo9stZSyVWp<{a{ouJxC#?!Zxq9N{l^-5oKYO;mdG=S#
z{La&T5AzNlXASvjb9lRY$F^`-j$Xn^heR_R|JjZ_qO|z5w7?O?&}9%UM&MdzGJ{<8
zHZ05p`ED`dy{U)nJP7<PX*}z3@2OJQZ)>BQ<=2`EMI@J-#{M{7fgK(kbM=?{@>ANT
z;X!)^AN_gz0!$xvQ0*w~axP`Y|DeAapKPcs<zQR5g+DLl_*TNOk)sa=n|_thsP!N>
zA(i(af53wUf!_m{aKXt4Dy2N8AUICTsfF6Z*B1uAl;k{?^?d*H!fHfxjbp!2bnSuN
zh@_Ur(_qEWWPOP>e<eB_7M##~E_X7QC{PFec2rWON!d!>@b;vK2JQFJNe8{dzmV<0
zs+nbCZ5mQnJGYJS2ljKsj_xldE===Jw-#;GJyf3}z}U>0%J=u&l;W7wqUj&%S6f5-
ztt6l3SoWTNy1n`|au9+>_A)Pavdd+rbt&OGo2g7!?b8&U-`meZ+<HZO=Uzt$VlCUA
zfr9YCg#=advolk|c+!=vdo|{BFAjW5xi9kM*3;V=>rSximxta>cRZb6Sx-7LbLa2G
zTlrsX%~-=6OCRZ=E6ac5ru(acMms9D?(a!4U<Yj#m$IvsrjLFK%$V*fDwtRO`o4Ze
z?i0h=DQVFmMnB%e-~40u^uguT&R5#Yu__Y-v`*SlvhNK>X$z}2<Ev?+<A{Ifv&3cF
z%K_o?K3CgRZv}UhMopc0l}R>HRIL9{18&HiKJSAQ2Z!2h{kr|jPfZ_RGb-zcpZAQN
zUP$kK?B!E+I76h_2ydRNSLIhtP`%ar%PH+hUDKnvKgm0Z4-e(WW*={eJrZ`Rg1=S;
zG}!hczVvXFU_VdTy(6U=#;R_H<yFq*Q;}aavRc1eOnbFGi>g0cR&P;gl1Q?unVH`o
z)mwSlVR7PxT%1HPmUh}O<)KukgYu(v3SU>YkLt0<j{@zJDet`=b+^v+dIeN>Khn@f
zYiy}EHDx;ogh|dZQ}qZg!VRaFp2G#Yr&1(U#`g&}vWDk$4Xr}PFV79fIyY{&IFl}J
zIfl(q#0`GDna;yvlEM@nN~cel8l)~TPQ-jsnLapswtLFW_Ya=Hvp@d0Fuy{iHiaMu
z9&A(XQsHZQUy#?C_>8P!cP_W1*lT0<d(y<o6PVMw?VziQOV72={yAHzEV~P;7X5rm
zH%{)wOS3<7ndbKE?;w|7RShqWI`I^34gKnGz4v0&*Wf}pkbGrNC_=9?ic>lUasB-K
zLk8cNWOmU?sg2F6-eW51tbwS>7!Blx`n3^@p-h)oPY4sX0b=sY0|{Ra>iQLJOJ3A(
zigY?bDbW(C$r;4#Js=qPc$aS)eB(W`$_a@nihuiN11t9RWA*_rtxIqhvqYV)XG(|j
z4cD{QV_Zu<oQ%+MTYIKk>+lu1u7edAHMMi*)4TGMoAsZ#bu_@*dLpv-I9=EJPyg9q
z^7FTA=6JB+`0^up-4!e6v96mRKzi-(Hs!`ZW1hRZ4re9e6h8L)BoD;q=2twrOx3vK
z?P}=#O#mIEb?0lohB?|_6Q`@alx(L#b{E*{jcncKp*;0#Yz{D(62+Re3A{ftfcMcU
zwB`-_a=%`EfBZ@w)u(hevMT1#RKq-YkD~Z%uWKB)L5>M5t_7ie-LAAve>-gw;r#5Y
zRik5T(V>0GpZ}g-`dLvaT5ABJ*G@g_9f;GZn8E&`rsEPGKTLbJ{x6_SB~qg>Sikap
zpUrp{j_~Yj%@v=3YkvtIt)ke^2Ja2nau3hh>I<tpt32xc^`;Z4v*p!W@mbkdg(%A1
zUhmE-gX;?yy1YAeeE(SIl746_9=s~~AOFv|VTIEYRu2ymA(UTXDmAb&X{~GBD^a!C
zkL8#9#a5|nE`O};KW{`R9asqMS>hWGkoeE*{ccCP(9be1?ZiT+R&Uk0)$J1dkL6ct
zYa}Kc4E%nZMM``rzP>7z*oD0poM2yU9q$`+z#!*}(*=<BwIF`>>9nibw{(*Qt!Nou
z@sa*BzA69GypdF5O@{+)e&NEKzZ4$2c5(7eF?*PeL^IK?8}lvtjUM~>dmH`>pK71{
z!0MyzO)>wePaDrStS)tZQDXZ9!FbSeqDs^Ci!00KylIinl2^n86;t_Igera$JX}tm
zU-(7cTUtDVyoR@H=$NsUXvQy_AJ2_exe%u+Ve?<4KmYsTPbYSgpQxUn40k@dl|WW~
z<Ouuv#8skDqkphZOfv4oA1mQU+C`U)cc(nE+;aCt{g3pUAE5YVo|*Lf*qz<J`%>6>
zaw!b8Q{pQg^v7Rz8D9tnPHdbix+$`?$B-?s?wB=wUQ`fJv$?Z-f7<WAlz}Eu(ghmf
zs-Mg7Wghw;zxZEQPOkIv4qm||@FmhFhn+emAHLPU{=D>E2S%Y|wh2U^$?Cag?_4(&
zRJ=U?H|4!pjNrIoO+mTPGnMff`ML3G1B25V4oS~t-#6dXmNPY^?QCswj<u(}zjOA2
zg{z8Gz>|k+aaFC(f^!^^(a(>cIIX+3Ot^NB?FBmUlTv{5247_FG*|I^Jg($7mwb4r
zB&l|?SS85*EJ^$txvN$?<;Iz`wyeq8`;bK6Ud(9ozt=*l-LuVZEyyQId%lh@8h&4C
zVmHVeip~1h5jpbwo-?Kw`J1dLS0kSjOnwL9FTdZLZ|Gpl>kw+?f&y~?Se2QuPI~mN
zRj2k>23fwsvAoltX<J=_JbhR+7nTxr#Q*j~{Mo<t%jjphLZ^y|qOH@HYEPGjBF9#<
z#TJd5jcM@no-%#*PA`-H@G>MKHs|u`!S1Js0!Dyxi?<2hF58{*lZmz=3jL?Zu~0T%
zbTIeWk33`VPAS#HEgH|*m))NAJ8DMr)=vHd#+}&MnlaWZvvA#YEdH^cpXE;bSy#wr
z_VO|e;$UAsRPfdvE;Bhq)*Eg2F3l1n8c$%nA?F{Ko1|97C}-b{(E;Wq^nS*8_w469
z{#mHUnz(f3yHG{{VUuUy14Qn$24mcs_<eJ{9f=DLiO31WreWv!^=6CXYqjcQf8PZ-
zk_1jdASU|`D~)uan`68<@n5PR3^EQ78GnM&<Og?ft?SuZIxBX@7wD{3|7h?!4k{EL
zlDFb?n<`|la{O0Xrzwz{utxaB;c7B%gYBdYHXLU**EVzZh`g~T3=ql5EK}&%^1<&S
zWw5tT58w>n;Hz7MRT{nHL!s=iv=Qg`LhZcFQHw)D7a_4RQ}(BLVqGo&=%2j{U9CSE
zK{`Pb&k~z^G*lTPK$kWPEDy0CsHfJO;^7$Ssk8I|_4CWKZ9igcEZWX5Xh@!))!Of0
zp6LMeZ-8Gyu8}ggP4OLKE^F8&c%aDPaKhdUa>+LbS3hM;!K_IQqJBjCZg~B%eJwPU
zrFALE@@>%_k*m6+&V7L;#J$Xg>!hwNGj2jfC||!hpg;Fd?yRj1zr*w6b<3f8s+~wh
zqJG0w&Q<=30Bpbb=w+h7E28S#O;EG*$Ek;>!rc8FP_zMn`Bne_qwKxHn);q_;luz@
zsaB8{EI&|@rqWAjiUk2HDo7AikRm1Y06TOP1gumEiVBK=2uLTPNRbjy>5$MvGXVl5
zG0EOv{@=~HIXCBda+TfInmv1F)|#33-Jx*g@2+O7$hhhQI)34xg)~3AAm!%J4@2^N
zL-J5Fp3bt0f-WC0l(7hsdUI-WG_4X&R{x_jxtyQ<F<GX)CSs_sZVB_eA*O6Jr!RkH
zAJ?}u5hpayFpoP_m$)T=ruulieB+zmq&3BV0K=;oUMdmp;H;25h5M1(J7mb-Xl}-l
zTu(WcTsj-e6k@#vc8oL?UQ_5(OWviBoK&KbUv8_iE@9F`wLm*ji_2Z5Tj)kjxLN2#
zlpC_<HlBOq>0Qs($TIs@i4algU*WIEfAl|nk#IlR&7ZMxXROTTAJl$J{xc$9m?5#-
zL8c`)ri$xa8nQ~5u6<X8_?%?R0e{io1<j0&0mTt+8;TtCDGyS4%!=BA@y1bw)#a|A
zPW|Nw-pZNUuD_J7V6)x>BlQ<+DaKb%J1;BEji&vHE=I1&(08$&){u`CHfCxlz9AH^
zf-Qjo)1OG!y(9urcilATIHlsQ#@_5}c+#@9eOZO%w8rh{Y&(7M@gQ#s>mBS=;n`Ai
z;Vk2M2aqK(j*X_6$1BR@zm4hldNf=mv9+DxNHWli1@HExFMAsEhTIfO6S$pPwG{mr
zcYLs2YmSu9!3fC-wo!(hZz>*lhHgUB)<!MOe;;H<jN?KNCvei!6z=8h*|});*X1<M
zo7zr8TCK^VBh__4^^ffzs%W1KTIeGu=S8<D8Vg~xfG7;|OqD51it&!yaCs%G=^6Q5
z1dJ}p)wS8xe!PcLN5KXgzu|^T@`(`!?OvpjD1DV9m1!l&a?$Wdj0D%1!wdX9uNbFN
zNR-uuw)2Nc6TaPlN9wW4!vFbdJT^G@yE(Iy8O`94`s5a(xxS#F!2#_Xac=7Cj$G+U
zD%etax4L1_BV0v~I_uxtaPfi<;NTqK?A>In5>hoe(K2F2s8;mtShdB%F>*TAdvURV
zF7*+ukzxV)L|lAm>eDn~c11K4r>81_Wnm;>(Euw98Pc`2Jup*pJvaM)09~uS&KbP^
z?fc{kX)wA08ScbDx{?|K&QDPhbh)(<U;A6PGh7OMheF?|;Aqw&w1ORfB)UZ$74Bt9
zZzy&vulkL%6r!;aK3|?Lta>KAh^-SU)eg*Vs!0z<L+tXD=8U5ap|^zwqCdpAIM`*g
zC9+wPkpYTS#)pTbY(^O&VsGi;!3iI=H{5C(SzoMuiZr@rM$0Z(5+61=`aW(}@xp2j
z-;=zLPE;v8gODGDgJTDc+Oj{Uh*2|hxG9{+O-%yh#U0A&;zRsv5(80{!rXs(m6HP&
ze=K+8whZ!Y)$PW)&wWftW4@_&?nuRweZM$c7#erVZN~2m?U%72o2s;w_qz@;R-e`=
zn7-r+$AW+oy%L<3z#LI?a=Sn^GazSSUzs=9x}cd0@uDMPpTp2!w;a#TLVFwHbbt5m
z_#WDnpHLMl>eXXE!L>`0Ffl=oAKpk9df})xyZYuQu>gzhE|nt>UK=oU6Ck_l{zPl5
zy!UR>(P+hy9)0a;Ss!*dKXUdFoAHPB<m%PX;M%SA+_oBnd6jjoQqFwu0Z+X*fT4c#
zXjJ?EW=6G88HYGa*x?Yv?d=XfCd_w-=3HH^*+I_C80)$EEmYuMe$*Mv*LaTBy`S-+
z-g8&@lP?Uzj4>&hE7jTbh|ZxzbU(w$cI5CR;Hh(;9iENDeebo!*916*cbC3h44KK!
zu~rBv*;S@!Gc!SAGaS%SDEo{e^ON80>|41nhhD78mQ%rE8R|a{?}(1B*3f6J_U>$`
zY!Z~aj1aDJej_JS@XcmqSRf3|-Qx+?QbV8Bojuh-T<mB~O3ZS9(`H)+=EZEJ*LD^p
z{y)Q|6h>;t*+1=_k8C}8&<L0s+SMB7Sb-fJJ`&^<O)Roa;b?yh{>{JnJcXFs+HY{Q
zHm)#1A|#Sk=k}Gq7yV$*S$y5Jp))R)9BF5bTyV5|?dnv*w4nttxaiqfFv9ShM^itd
zwtbD70Od-j?knLxPdyqxc7H!>+g44ZoP0Dlbuuvaw8GKGK;_5FTR#m5UmM5jv28+s
z0evls=`%HV;~X&}R^CuK3nzU~x@90?rlHy-W`UnLemFzHOEdFaD#!N?D!sb2%8*9+
zZvo@-msBNeCCSdBT+jbiV;fWcLVhntu?$gd%-cas=_B6#hTl?W47Kz0mTT%4sdL8b
zK=fp*yjnFaeoNOTS{n-RCDSakWKU}yN(2TQ17|p+N+l|1urTd;Vpn@vL2wS{MUE#3
zv4m1S#MCp*xfo(o@1}20JC<+4<88x<LQn2gqpb*?1{65uoAEX$+fmKSydL{pa%$4J
z;Tx4y>~qOpI1Bn<7a%GhZX1%!c2m|&d5V$VTsp$Xy}e_2(=fME%E_};i#S+Wnx9j^
zRPX1@Da<~ElV}?y2HC~=^@00RXUrA4x+C@`{EY~W_&Dq*g969NRKq<fV~@>Lk~141
zwv8`re%*wlYZ#E?S$m`y+hXVz&<CP1G=Ty3ffK0A-l`wfslWqSs_Ozh|H1nRK#_1D
zRkC;Oj?bI66g9(0!G0L~F4CBhqqu^wA0~m7@1k;g4;1NpPc~lC5!frDaCp5C4fpIS
z1`_^$S@&B+u5ULHc%7kYD^;cPu)Rz?)v$ZVgb$#)AdCrDKj*GAZ$uZ!y55|!f0dt0
zqG&Xx#;|Wq@lWo2n5WqA2W*IQRowcq5^~w~oc_lzh^hSLo~zML?(u}NR3@`=)MjkC
z-AYD*T_v@NFI|j)yn;&PYO5}@l;fp?Kw?TMb0OoYcM@YeUoRiEaXFgl!P8&h!{~^k
zED1)=<w-tTxSjS0ju}EQ|1@pRyqJItb<M)nHzHw<Q*7@SGe25%cl?ve_3atv(igjR
z9i&ZUUQ7aeT?tC-K8s491U+IVDVG!-69~cbW^IaRJy5v}mXaK6CWd?$O&+ugs{h4(
zUGY#JRS{i2h(<{VOh(GJQB1}Ue`(qzZIlWExE4=~E0uX0FxCvWwNbL9l2dwF@~}*h
zq9hh8y%IO(WCbXeh*ST@Go*fZ5z1m*r#3MO*vD6+Bfg5Dq?U<`;ndIDy{udqq={`h
ziQJn(B|CA?P4<=U2ODbp+Ugz4PjPH~do_*TcA+DlZc5R2-9r|gJ&+?#Hh!HSkT!1q
z$6}d~;>xt4)?%L$wa*TH{^zf&wVRtqMbRUM4ItKwmWC`SJh!RJSN*ZPqFhD&miGMj
zJ-zK)jkvSb`}2a&FHnC)4Mq#ZK-_549ie^a8zMd<zdatmBc@T!=x^gKx2q(3wp0!T
z%hI;BJnZs>c2#UsSZy3ugtwUxBzqrV&%^?vY`sWjBWRUL$WPiu5eaLL2;R|NpCfmw
z?mOFNG3Ri7W7|W{jQxSXtL3#+gI9vups?|g=s^V3C_T7;KKD&bN@u^Rzx3mX*}uwj
z<ok>>Uxaaxp9nUYo0mbe96Au8qHkYswHgYOa|c`OtF8PnD>W(meNFCqHZJA9>r>>(
zeNyOWwDBeAMFw|M2V_f<^c}hE-WkV3^RN+oJ@I<WU~}0@;0IlsX;;!*<jt@`WB*1H
zf3F-9Fi>xxjW_ah#_Ao;yPK6LAWbHRWUg=_X-0G^r+<avzd(wJw8RC~2K`GZrDeXE
zsZM8qU$vq%n>aGpLMAuQ&RV?MmMvp;r-Nxz-}|#9IFBpT?o<*HAw4y|(W|FigXrH@
zvd^ohJrq-0P<UUmrEh0gJNAJib2X!i&l%oG`Ca?{pYkBny7zi_`;6LY2<$SWVQgL6
zE%@B$<%n+y2D;A>IWu1adAcWxOHtAlc`wFOekXgk%h>)kqAR>P)Vn`}>*^e7m;xzq
zEas@YV&AVsdAZyRjA9x)skG0}X;%b1Do<XPZ!~^W#hdDS8~lpRv=k+_kClF7h8wB-
zvGyJa^9{O8)M?g&wyH9HzC-7ihsMh329-WOwDz>#14O|%1SWy@_jfgJ$*1!euH#Rv
zJ&IG;q(vk=<?I<Ldx}AHPN#3lW8ZWIi_nh0rfE0)`;jNqJ`Lj<t{+8o5mnZf4pJ=X
z>i&HyZN_x(vDVF}0gwKsoF_MyT>TehP*4xilcGaJ7Zh1F{X~URd|XA^XCzL#(hJeR
zzzDm78Y*7{ZE|On^)GE?x0R(cUb-4}a*B37K=@)*>(e6`gw7IX^+dtS!K-^BpT75Y
zLxa#qiovU%IklgGc0S+OJyQmC*redR#Q}|tW^WCn{l7M-7`<P7)(cflou89?WbVYz
z<aWTHV|nbaN88wC4^}|BuI-}tKYltc8(Mwkv^pYc<ADTIEK|yUB{P2iU^H4P<uuA7
z8$>}=k0R4k^03f780^5d|C3z2X8C{37XCl}`9IK((17b_zPDW$Nf^B0K=2+j$$rFr
zsBt4kX?rvtFLp9<+sPt)$C1P*$%mN9O0FG79fhT1^AAhuo4Lymxk7u6be}qMx1%*@
zah`HSX{D-YNi1;bkA85+C3;N=$R0ee8DGb~{+IvvPT=RspIc{1e{S|)eUe3brEux3
zd;9O4EVF*ep3CIC+vb-I`pXM%6b{-|F!irFTr)fhmKlFWyyOk<8rX3mC@|`|^&fgk
z;iE#u!d$y=wEvDyFi-xbQ0!_M*RP?D8f!`!wX2wi4xTtl&b>6(lbi2e@-B3;=W>8Y
z<2&KPxaj>_Z<KF$KiuK*Nts~QmC*sE^LK@mf1F}k*qka1w=9^CVsxJ@)T?vs-+6|h
zId{WKHFMpvcsiC!D}DF58{>UNxZtPJx}o$GEh`~Q(jUBZgcrJ{rU&li1~fN{yAqk9
zKXcar_(<>r9{(Gos{(+%q$#!6>+!mm-+ohi{rQPUgJaxhm%tHz{T6p+uFE0ycEXHX
zNprX|+duI^nM%s1oxIGSyqndx&+SckVxL=x1@|&Aw9x-<F3Da$<XD%td|YKcVy;H0
zu$6r4SzT-j;v(2V$s&5utKMP>Wwd!vJ&<0sQr9)Y_-~xsUxNwS@V-Y>tar1&%JKJG
z89in@kOiyUUuhFMn)H=5C9^RVzIA0Do8fE3_Me}o+D2l;_z|x-mRe``4WCySTmE~!
za3`?8eJ@V-z6@)_&5jmu{aIQ3Ou^3ftM?D4MDZ#E#AG-QKO&22pfC66Rrd3ZvaHT=
zl@&~do9#3Rf&m~d1u&&eY`zPTIrT-&XpL<FUmCi5ovBQW!CIl)^!FvW`tD0&X1HeT
zKZ*s^#*bcz-T-WNaH*(Am0nxO>gNq2C725zXFebD3DtH%>#|O<Ycw8nGmmo><+=7k
zTL?SE+H&NizSiUQgkMJ%M&BMB_wQ2M)&F;M*__h!VfD9g9JCnsQvbKh(w^<W(M$2A
z&_AbE6E{tRRlXzs`RV<rJ`EvhuJ?Roye4a)nwZqZ!F7yF&-0z`3#vVkBK>}XduOg`
zoR4+V&`B944)9qNVNQY!@Z2qYLHd)fWodgwj^J>-52&Lx?S<(7{&WUbFZT>R9~B+=
znuo~dk{3hK7CJ0WwJn=<WtjM`oZ4Q7){Kq(tFN_>35bDp%6idI;7=5jmzjFGgC|B~
zXfdqzKp)C6!iA%;xk|bJ;`6Y8077qIJpD(+&&Jf|MpP2_I2TcBHWJk&4_SujH}gcV
zfvvh%=2bH_QJReN>;g-1NH?JZUO4`q-NII4U-W1cn<c#Mjd=Ax336_iraxzL7xXdN
z+?WI0n9{@05DB}RLo;}a{x?@%bW=X>if3#_;WYL8z#ew!)rZEWOE4mZOki|PE=K@Z
zVwEvB#$xd|)?fCYTESX-n3nynVwQ{h1j?I-!fn7;bDo;nTKt1OZkmhK+L{?J;E7|2
zFZ}5FG}ZsGm}D+@qZ$*GwhAD$b4-j?3JhPMr{18F4n=$lQT~YzTStF-_$^pN<miQG
zWC$N-QmhTuGM10B8%^gMpEHkf@7e`z=b&Uo=2XxDN{~!k=(ek7hyKwl-$36NCT%zV
zreB2_5q`-kX@cLzQHhN5EbjAa+mTfKw-z6PVGv6$x0yzDh1KuG-r121JlyC8-n}1a
zfgRTo?x4P-U*(K(Ok}zDv6$T~$%e-()OW|p|2jf>Of@#m3I%;)1hvchCIR5Dag8=`
zE7FBXtU+57xK?KmLeC}xIYWj2wmeEU@n0H<m*@ppV)0XkbOb@4AvXhNzr>9^>kkkM
zDSzMsVrc872eomZ;f}-{F?)?x<RU+IytJ6v>4Ii(^~km^kXb+T?l{_<ohmGu0{=G8
zSTEH#=Jhq$y)56~ZhA$ELawF1Xpih=jr8(&jgGVAL*=yh|17@)YJaK?d=!&!DBWjb
z=T0H3Zw1rB39{Up8jX2=%hEV1W~DtU;S1l59K;#=+6cNF+JwQ^bVJat^dVNrO-^Jb
znH(;tG`<uHd}#*cCuZhOfPVXVd-7ic<0=*p#hn@a>}<~1T%`9K-Qi(${0}gnVan-W
zQ-^N%%?8)P9Va}>=)aR^$T;)|I`!DPJuCRJF`-eX(YBanwTOje)~%j3yy>0RNP^B^
z?je#d_Q1lkz1X@hLVzFKW4w!EydUFK2U_V5aHwXuwEO=ba_S28%%wg~x<MS#3$)UO
z<}b;To~j-jTy7KrKeNg^`E{7ROD2`H!)=juHIU-Iks*aUyv~1A<SsYaWW`Xn$x(d;
z*GZq@WuMRdy^N}>vl^XzT<bXc+)di$aMWCBFMo$!q~jD}CICEtnM`zQFz`KzE$bD0
z9G8kYSgz4U>l(~Pq$6FI6REwGdcz1ZJh6oW2tTt=p*W?1+oAg%Xi7YKq#$c2)hkII
zw1ob7d?pb;WfJH4PbuVX%x2`baSe84(XcIrr4Pc_F|bCXgjMv>b=%s9?1CT3lm5V-
zDB?|S7h5#~6_{G^iRPKkwaRuyf73vlbl&21@`#0kb3d`kx$EDAfihMLW79ifd`ynI
zT+WO{R*wUTWPbi;7AA@RB9GU4?_Y0jFLW#&5MO)>FRo#e*C4Yt;?qp@T4bgX<YWLW
z@o&;KUd=^TQauy-vCv-RdTQM<PL=0otT5m=L0;fL<rV*nT%tT0;GXwKw?(Gz=c-)U
zn!Cr)ogyrs2Fr96AQ1@6qd=9jJ~VzPXxu}=bn%Y^MdOf37e5v~!j15F){H*3&Xc6p
z^BC0)xv|i1vsH%6qz{9S=;U90230IfVPqzFkIiY+$A)CH8{AFWR&ZBXqsK()JlNR*
z$gUIU>%l@mCF^iK&59HJhf`EZeAqMe1dU<Y_HptVDp<lA=F0{o-%U)T5&ZY8NwfGB
zE{xD)Jn>3G6PI;*Fp|p|70kux3G37!{%Mx28xGn58Ko0{v!&%c(78ac>)+Bwaw_jW
zh5VIL5K8>C6FQ&{w66k#Xvn&eb$BrHIMB$@J=-jJf2ip6RTajFVJ0xu^Aj`9df0+4
zFDDVxV<G3qER+R{y+|RjB4Z&dbtrLFngv!kQ_wB(!~k--7dfGqYQYhWK=*Ka?ARKC
z9K+W{qA~DO-$@n->;Z&m{5HP`G=<xf$A}Dh(@k8ZkqAZnec+Fm#6E7mC;B;s8rjHl
z{lXXHrJgZ_j<gX^WO2$t;MsrB;2^I!jL79?b5M#0Kk-8tR_qs89XI_npuUSL3A7x@
zt=K|!6mvf_feK^jL;_QTSF`Tp+Cr=o7HC8u<_q5_7P?%vmHlNPa_-313v6U(O<jL8
zQE8N;dxM?C;3x9N1XZg%CUAWc{U;u>L;zXbayR`KKvV}>{~Ztd8->ernlQLKO>KP*
zl)WVWXe375XrMlWuM3d-ZC6`}TSGvW_CuH?xI|A})1cfh+bUQ?b+6fbK!@}H;?N?X
z%`J$lX9~XX*eu^%;O#vfzWz*hn%ay8y#J=>Fj&uD63bW`GZ1{4U(JgdOp7eeyO@Bs
ziUrErpfjf+zCWn=97;`Wz~n;7N2dw$e&At@4>CXgd#OIScnAu50A@BfQX`H7epHMm
zk5<h?^kv-rdv|dq^6n(LlrN=SHHVQu2|e2ev66Y^55A6~m)Pxxpo-fpAV1zj65}Of
zKf_Cknfo8Ww#pr_$%CDy^OaSnU)hb(XWkzYeO?I>M4@|Dt3WZJTK%$8WNE0>L6w)Z
z7ox2Pi$L%Ohd!K<B5uQeV3umx3}t>@XI1lBN+LCMk=1L2zS<wCmTqF^W8hgA)sMG$
zAor`d2=SlS({+uVFCnK_L{5VS0y^B^jUKtnBgWpk8Wu+YYP-+|Y@ZnMvcsNH5H@O!
zx|{&rO$S0a(7r~feM5GV+B}C6=lO(76l|%*jQNyGM}p{^?a=e3)l?6NrU}|t;dRl9
zvPN6ZY&PWV1no$K6!n2rVSs$834Ln{!{q}%i7uzW3Au$gPIqTCZb83$`3HHyam_%i
zA><iN^x0!Na3dtbbXjxV$4MNV$D$uefcGH#oKqYOS8V<g<4zavJ_D@b9d1U^p!wMT
zU9%dET+be2_Z;d0m;p6pJGTSmL@B6bo<pS0VTyPce@%8r&G`<!**r|b?1t3X*;IZY
zCm{yf>^e57nw-!sXLK!z=glL0h1%yZ<-AfS>usxE_${*W;!W_rIwU)T2|W0WGNToK
zZcK{fumWw5gG^3AC)0tRwNg8_t9Z5h0}OJtY82{x!&zFFL_<H-R&T8j-=4xuuLmna
z>rd0xHm%bkaw0%T*8_h%Obcf3spd)tTWUZr)FEd}3UtVixE_9p7qA1}Rb$CcZ3q6Y
zLz$~69k%``_di~ddYY{7;rhOkyY9TIglFhrI*|<)bfLp#4bIVQh3wHW&koj0%IpX0
ziaMd#F5;-5d(=^<3b@B%Y_?E8gF_T3h?CHb5^x*+YO_ithHO<0@KE0KsC_iFB^>C6
zGO4{K#^44IWeD;(=y7AJ2-xonUVOQ#aUmw!wP-S(SYJNR8G8fmm4XDmSd&Q%Iu<tw
z2g0Ux1#qz@P*I{GbnqUSAczlI(2oc3bJ;N_dz59sGlm>PYB8HNz`zal0+wLHOB^BP
zGYu{HExHkxY3N#B%$%*I2dV6^!HpM!>;s_#3Y8O#fx7gD%;OM2w&5m^?EXIlRt#7E
z#t9jj-2js#I70P;h&jX+x~?jizZab4xq<YAW9m>E+voB>;8-&7WB>4XKXVy&5{+2I
z8;fg$l-jAjpdrAZ2HY3#ct`^bGk_02(XZNIN5{6~Ac3N-4l%&{yuWnt8Yd|j`V<U2
z?dPCo*q&~tH9U(sAphvTUK9qwnxYAS{boE?rRm@RJ+gAzryp?xXL{?si0>#L)!&R3
zE^)H{=)s=#@k|rM@;=ji^g0*QE$CbBn?vmcbVPQ;>$BhDXl+f*ArHJAN0fnI*~3v~
zLFaL%;o_4OT;jVpAnTjNhJO92&sTkUL;-m(FD@9rwe{i)e3i$Mwfh1$3E(8OLRY;N
zLI$P}PSMB>h(~11AY#_Tw3)gtp>$ti;pK0(A%l;(2;M|QR&F+sH~^VRV*ugtGt}FH
zdQ7p9SH!o-yb#jiaB(FE<PAFXR2F=u52deBtyvLTpxjN#JTHX&cp2<b!{(r-fTned
z!2ntI8(f07r|F0>Y>CfMhe#Y1(*e%jIP@T1fI96mcspbgB+=tFnZRfVH9+K?fPnEl
zXIa}cGYI)5MDow+AgHbjeFU71_)1VOFwc)qy+I&uh{>pe-waJThR$(;rgMs#sZr2g
z1CbmudYy&tV3(U{hQX-&Lq0kDUR}5K$^QeJmNk_=3=|`k#*kmA@KT&<Kf>BipxWr`
zW1?)zaLHR{`UL#7`XU@Wx&;$Rs~-7;iH{}D8TY+-2AL<U2lw(d{K1E}#jArxmZxTH
z=`dT#kH|g;79J_m$tB4GWs;i5Y6Y?QJgDISHukKNi9%bUGQgZf#^3#(q~nm7DtLoy
z`frC~-va!7Fg(A$V831}aS^mqS%+Q@;6D(Wr;_7*$GkyLQ%|tjgcNW~vcnN^8)w>!
zaQAHbRtakcOx9EV6rpXZV5OjOM*&*4zcT6HYY=92kHphLYJ!&#vpyy8CzqD-yCr>n
zt0N@zbavJ2+e{O?qk(%2sB)g{K@wc^_F+C%_y=_yG0hnpz;`3|j)j?KBBih-?lnqI
zsKgVfw;mdMLcD>v;T2CY&KB;5zFn70TleW?7bI~>8xgP{nTLP@#7`r9F2Xw4+*#h2
zW1m2DLfFVcE~Y$<Y<7|{{`GU&yqG4$NpRwAh$IzB_Z${g>VmJ>*p*1fuFrP|szcj$
zK)<%HqTGLAis4n}D^K3SSAFg^5lZ3rc$>2Labd^$BGnZ5tj{X`9a2+3vwb*!L*CQ{
zc07!&Vvt7sCb2LkgaM|ADt;U;T8<-Ulov~EMeKK#$a7UH&+f(4ONJ0Tfa>pxg*4bk
z)X{Z`!Y-l#@0NJTO8XhRf?jXI(~e9*6}_0-2%}P}aM4<ND8BJ$W*XwZK>P`jWULZY
zeRPIk?MZs~TWfXIf^{Q?7$}|>Msn$f|L~BoGsfRl%fN_AEyLy!#`Tg$>pBt0?tY$V
z4jEdCJ5qf)XJ=0~Xd)m0@#ZkK@pSmG%%l<NQ7>X;ml3`RDKQY9myinAllWQwKwu;0
ziBXZZV9tu<5^MbweBI~76t(-Kh<gzg>A4O&g50R+gfHO^{%f?`k2rhvGFTG}yz4?M
zft{n{1qOtl0xzPsIiUz&ty;z3&^C?lAk5-><K^vKULmZtoMnR1D@P6f_aFlT5EaA0
z*s<-s<5~w=%xM~%_<E!r*3{}CG62DB(2S^&cTR!lwk#>^^66*6B5JfGrct)~_Ir=`
zUvMFBUjtn69q1lSk)r~)wAzz?b=xc|7Fn?YyC9b537@EuUwyLX<$e|IF5?fp_@?3(
zC&xj8>K0F#WEDRpCwsH#W<k^iofUhS#LQkRHA)w{PP>vJNg5_kqYb6kTTA}4t9tYX
z&*?ID7LZ4F!SB41?`wG{t=acSa#j9H8?0@c|9+V{%=8gZ>!-|3TWnNG$0`cCVHGCe
z)FUMFF5lD8J0MCM2~8W)_|Kv<b8~f!I*t(+0KjfnLf-r7Pf}e8_kF0A*9v=%lo0>G
zfa%3U2VX%3g&ht_x7&{?yYSkdPO_!jEM9%aTJE^uq@<;^@3szPffwkEv!D~4Z^|<?
z`(|J|v8R=VsN^<r{X(4+?*bV}!mI+!YMp$i=VgJ(kRBc|T32Bx|G08}*PEmdt^%;8
zCq0^9ZM#Su?j}r47<5sX)DYR2bAcT|oDg}bNHX$uKRTf@ZLHrD#q*=nO7>R#Jjfez
zHU4H}KV_h!xE(5TxZBp6D+mg4usZp?R=hP@%QOp_C@mAs1dfAa2FZeD5OriO>v-V}
zJB-vC+ctw%YF_-{_{|nDc?+*&;nAH4V>^TjF79jd5$oHEXD&y7mQ33e`l=JK7m7I!
zX-^uKzg%278NNG0{uX#W88`#)w77So#1ee6^TtE%$zXsWD+CBMH|>k?x)vU%@uylb
zuN~&?^K~lMogymRU^)^NwV{B2Qoq?dX*&<a3AI5RF86vpcWeqjqosx(^*P|7fBnCK
zojf1E3XyI~`CdUNs>(OE50auhjg{|fRYoQTw^V6ONEZKq)oQ#?l0vZSnz=h&ouztq
z2xR*TNbPNU<0fr0!AvsL1B;rhwmv*>=q$Ld_KOWMsLF#&ZgPh8#xt06VN!%iT1#0*
z)wKi2t~BV%lBi&R4sQGCgwgtH#4jqA4n#_(H+m+{;9L1^@DKrVh?8lp48lOh6(3}i
zb5dk$>zg8itxYD{rV4r&oZciZyhSqlIkep%)(#suCF9JyIynYqT`MAZHy7F%06EM1
zYD6z5?Ny7v5WX8_N{%~mK4!}sJi<$<-3ax+8qvYw3Hl4$^8@7f{&mJuC*j|j`lize
zPGsIekx7}36|Z{L8@J?+zvcbV>LaY7b>QKlWrf;lbw0atIx-L{LA_BgEDy};>%w|O
zt<k%C(vKTr;V%hyk~pU&8Yn80U7YROawE%!8q-4lKELXz<T+$sbMm^{x9i`xHeJx@
zI+<RS!{g0}Z4NOVKSF^ZMH8B8vtMYCNhE#W7u}G9yLP6Q#D1B$sHYZpqFeNq13mUd
zHUt0VQk6xoTd)O%ii%2tnnPA&{1s<|;-Dab+A*VL5J5qG?W7W$pE4www2@1pwuzze
zMH$m)0*HwR!Xa9Z?zG(AYWtyg$u!`YQ%Tah-Mr=9m)~8}RWFO62As&vLQ+;13l9{g
zlCmu}-cG}(ztY>tf%>;DKfi2$?L*#kgcPUd8vc;yGm7G!m=lRnPflSodOv3sxtlDD
zPs%7U;w;t*a(*qmSXKE+@t3~m(^l$t;+{`~oIkPIZ`Ir4=vjuXCFdebINR0q*?xpw
z-8r#IhUk17=~1NEYM>crQ1oZ6@La*`yw~QOhsqlJn3S_>wA{}(WX{ZgwZ7Tquyff+
zqb<T`@Y<i#$_5hGF11h_(l!z2&>a^zp9e=jU*U!+T#YKZV74z*^#JSlZ1<N;c1|3%
zD61#;p4eXJOE+a&{t2Dsqq$jTQ#*^gRs;~{lm90hazpw5G4b(#Jv}66;0*_OIagB+
z=ox?CciHaQ58<+Dkta$HI^VWEO00{%?db7JUA_2r(E#0;%76m)KDT~;>zl@lr%Jy(
zo}V}Wp4ms8A16nm$Lg$iSD+mUM`YqNS+%*`P|Eq<hklLBw}(D>mfBP~WBeN1Eg0bR
zwvVa4q>m`<*J{nl2U-sE57O*X3iAM+{;*4aMt;}*-pyG>-ag#-U94HraG~bENFg#c
zZ^Ytv>JzGqtiIa;>h@ll@K5X^X`NpH%_)AM$wVPPDlV5VRwQhNY*f@sLMY!Cv)aRX
z`*Yfwd!u?k`fKUk=sOM<`ZiO4pF0*AkrLymJL29b&Y+60KZZBuH9|ve>)r<GC9JC_
z$R<M)L_dXP&)id%0%tgcOnT8->HDVCzc$lLRH<fX2v~fRf8<cN@OXUDW0^a##aP;C
z%GCJcm?7;>$-dAf1sU~%y=Nu%(ZU!}%j_6m-A}64tv2=CbA9q_m3HQmer}PAOq0UL
zX*$ktJWUbNJl``~>9z2ZE+B!vR-_}@y{Q<WlUQpjv?AU$Dw%A2aFF8j7s?%x`V-@7
zyQ=dn(77TF^%;M_S0Lu1W)?S*MkLu>xK(%gL3qb4rS(yXRlc|h_pyMbMbZ$d*NZ6I
zioXd;Y93-2h>O|pD>#)smBm-e<?hZgl)G764kVv)UJ+@u0@WB=<K_IkcY5emDdKl@
z@*MFRLsg4m;b!^?MD*gQy3lLZ6`o|XKCCzT)EkGXTE6{kAZ9((`NLMst5IftH=a=p
z&DFYronkj-j9`i-QOjYk#v=OaF>krXZivgQsA}+HJ&pT_FUtB7_ljxz!1TJ`cU+wT
z(((>VHJpqQ0HxtXc(zQc&z{Z*OK8P7dfi36rcsZwPVL|NcRh@%;*U@b6p>;etR@g8
z)c~6r+2p~|$Y*WCMi{<?UgdJSLM5Fm$uOz|#I^4SGPfe7-onJK@9UT&(AEA7{y*77
zW}{OF88D|J6K5dNDJqjfe!`}fVt?@L_e+ZewQNEML)dr2p%0Cv%V9pcbS@uS1n*F^
zS*nY~?sbjGv!)M7*0f^Z9!Ry8I}qYSd`VzA!l5NT_0<+J6&uHs%mhacXiA=Pel{+t
zJoNQI8n4XeiEE_Vv9hS$d@Ez*zeo_(No4XN8L)>S><O~89O>zQiPpt|4Dxo>qoI~}
zD?aLtl`9xI3O_0tJg<Mw$~2Og|5kO9ngyqdLfqGjf&f%z#;s+LoDj&`n*pDt#avxn
z9^nwWx7n!QPm?&y!OT2rUlGl4=W6Lm#}Xy?hDan%qkodY8-E~^iHJhzD?vDYiR#4d
zMt9?%c~qmuJA0$MRyDLBMI>4$hHX{h60!#>19#voc<Jk0(+}XIL@L^I($23oVox^w
zljl)=^DwpPEU7^m!m{x&*eS@u*S1Yjm47jK;RPV;!b?bx!T+&_>HRcGWZt^`4qBQv
z50b>1^GiQ0fEQz{S`~HkqHg>U`Z(tyBw%XM(*mTOH+hUR^cFfa&886{UAx|nT2nAV
z^HuU)5zZUCr+3ZvowjqxS*&f)Gb2Y1UGaZtseEXI8waqBD))1c&qchq4Mfwav6Y&+
z)wi`x4JX4@K-O73D1js0+jaUhYnZsnrF?`gOb{0-tAvCDU-bf{<nOc(5vQmFQOiV}
z)TPHG=8+n&etqAsou@>s>s_r#Bbkc!*msFEdYe0zZKB#CGr5%r)20WFu02RLxsE64
zN)<VL{bjSex8HY`dYBXWL+RCfhjw?O4?R2-b~A}5SK&BA&kGf=rk%bcP8aO(?z$`>
zM&hGO0bdF(?QW><Am8*f#4)puGyhj5x1DJlX>|mB<yMXd6?Q5+#G$<-c51olOd62)
zRO_wT|897A_fUsO+n#13?1A~rz-<IIXD)}ccATR^m$cizy#e_WC<rD(KkT3muJeEV
zLI$)?i(h^~a|nVFpyhp;)nBgi&xIh4Wp~a7!okuxa90+v4$y>7i~f;U0&gL5%I#I5
zjQd=Bfv)?fvbdP{FSSIh`wmDpg{$sA<|1jj99j58t2+<JUp<(;>&)HvAN3|eFZw<2
z1(y$hQ;e8>ukc=IS#OwVCfD*8<7QMQcr95a{qwLbTm2S0Q2!Z^;VMOwnQlx!IMgr&
z?E0O$3+%t;lNbFgFuz`IJ5AXPyo|Uf((T3)<OQ{O$h5QXOLTdqh#~ctz|$MQw)OcJ
z|HbsxzvjAyj~%_)Vha9;aVnODftx>46+qU(SgiF|=u8*d&#7cj3-NbEA~1>{U*@NY
zJgBW&JK!60iMwfVc|7~Q%!@w6e>J%^KFq#aiE&JUdreBPuUZK4Mos(a;yCQ&?LRQz
z*I(mf_^O3~b0ZdosAHV<q)q`%>l%ZRnNR18&LXm~d*+SMD@xH2&t04!I(2RS?4-!X
zuKtpayN(SZ<ENm+BM%x{(C_Xm_}~%ZQaRv#q-PD#nyDSGF$ogw&Ow9ZuvY*FU=Fxe
z7q~T^eC(fMRuXZ_^q-EI8!XHg6Q|eubv|6%RcIy~9_Wh>(U_<zI4O6?H$RSo?>W-6
zjB<A<)iQIxlgi7N*j){en?qrP)L29j9s*O+FWWa%?*bpcY7H492O1Q&s6Jt=)6NZK
zx028@dePBPv8@z84ocz{^wPsX_2a1SW#Err9l};p-}<)i;ppJBL*g<=e7^E*-9de|
z<xD8HH-s4|Pchrm;!ozpug+<YquW=Bc?WqYQ~8Z?TAi%a@<bI<iuy{&f8KsX+;W|f
z4S%#Vf72ITomBGMabCIeTHRxhgS125&2r~5NyHcTtUb~Qm$4_H#MZFY2AhIK1Bugp
zg<pAuU8YFPm$%y?(L2Ljcw7*{rzB#9D9@8~GafOE8>Rgx!BU+^+aR=GF7}Q+c@dE7
z^T!}Dbz$nP!no;ktxs_AiU}CFtGLg+Yx6r`eX1{3*1g__W3>Qwo^U=2#RsoY7KzrO
z?Kj1*vH4bBrYw{%K=V|3kMNhcsl<wDb>Oxw*iloQeksrL&%!;OPu~BB9M$ZzP<4$q
zH8|&3%Jpemz17zt%f4Ygk58ckL`w&*@kf9@-a3W8jgWwHx%Krsg5>XuSb8~s<l3wv
zX$HQY%|`7<8lG`%7p<@8Xx4*KH{v(Q2UJW&YXL1$8LCdnRq>pCc<9v5B~$}(W(!DT
za3xqDx?z9or6--f07dWw`=_!$63?3=@+*G``=htJnYBGY@`(4XbuM^8BA=d4MjZBm
zoqMF|%R$~2jCESXK(cL5q41PbUx$93jkfPulB_0m#ub!2&X6gEP7;CeRG*1u<X>)r
zjD=Vw6<MRpGn8Ak=#4IKV68E!`(p)>lUyN^0Cm0;{bQcJgXI&HR()AVv)`JhdIEYH
zu9zm#TPV`VSvS};L4`=$K_w#Zx?wkIq5%!a(@GQei()rMz>6Lv_k;e3${%Njbw3(D
znzE@{xEDR#eOG)|^P!DLe*B>6ZtRN<bpI3g5oa>~L7Sd1PgkpPpEFBl5yH=#^bw2T
zOAY2>ZZhR425U4I)eq&h0J9>~JwkrybAzRG==U)>KOU^A55Ir0+f?*?oLAmo<!Yqq
zIj9h!kB9~G&XXM6kBF~I^p;qGLrwa=cdVh51}COQFW@3_?WBnOi8^>NQdmR#<*I}S
zhd-&aB*>4>!$?|f8w8ywi7OeG3_33fXk1*uN|D>U)J$Ds?&b7_=TDd(7O6BfAb~Bi
zd7cf9QyzKL9QZC-8BIRQMMiv9n8GQ|*?&afiZsv4#P!EdhX<ipKvGn{>HGt>3934#
zM*>QmeK7B&qcq$BW4stfyXExNM_~`=cMg%nC&93vx_u;Fw?Q0_SxJd;iwHf3cm#ce
z_u%%?q!2l@|E41`dZpi+JVO#)zj)6(gex>PnB9G^NJlc(Pfs4iivBe_EUs{_?}RB`
zWykrL?>&WZ8`Wh~CV{jnk@d(@Aln7}KWZuub)uIg^-~^fd4MOQZ$Z~PY7R-#kKL+h
z)Bb#*L?qtBtD8grFAIum@}Ff5CW7An+s4hGXq_KvmOXcuB~b`H)kNs2nNBM|@yg5m
z-D>E5E&QUaltRwkbBv7VvUZJHNzFue(lBGLHuxWU?<7@Abo%5Kx$PrJ^-cQw^6<}m
z9c0z<liN~pOH}w@XHie3W4-}>KQr>YTq?Z&98^1<FS2j~Caa5Q$?f_K|Gt&JOuRBp
zVLYv5tfM?dCIeH4{_YIcylo`_6(04WHCBZgOx}~xUof3oQ|F83#T>aSdjFevzCnhR
zc!ig_6f1H5sk;asm+U@yBL6hBPwXZ$Y4<E8l#(<<&@1gPn~x|yG)Qrxz1MWHki98u
z?3lAWcFFi86zTuRI9@kRM{7dvh)>j}B-ebb=Pc4;(#bg~3tk-+i`UkqbU&;;uiclZ
zS9kCa6kfWDQE$+2iN%HY>aP8hrZsx}$UAR`sMP85B2AbX^95c~1w*Sc?7Y9-0a1O)
zN9n#xb?`w0_y5Gb%G}J%wJ7!HwUrFq_u`^FNbA%VIE!gK=?DX*D>olW4Z{qU^G~7k
z4V)0FBk(&jIV5EXX&RUHyK;qD`mFyg_`6E~t6_%q?nhJ3ERkZ}-xA6Wc`>hh<b87|
zPh+!$bdQA~XMREOxi?>M2CP@X))gY3102ChiO-u7j^aYOn=@4E6gT1m9vC#W@cX!&
z>ypxqPw}1%btnv0SuB3(n5FydiuIL8_m5l_{fJv6XdRS2=Ox10_FE8xn9W>z=>rd>
zw!1~>TAh~a1pcE!MLf$jpHOK5s0N6-o13s_GrZ5vvEwy%Xsp?<)@}Cm_Y2DR5vJ=p
z4e(KT!N5l9Ij^@0)~8=Uz;O1ae&5mIb9X_>Lsa!}@oTa|)w*slR<RU=TQeW2)}l?D
ztH<nM^i%d!|6#(qvD0jo9K(PPfH}!jKYuI(upIOLiT1|7tw{&<d+KA9qi#Y1hWeoV
zFv8H7X-yF5dhN=WQ|PsE`}<X<4ff0K^v-tK_)6-~Sx@I%YL;HYy*p;zq^>HcGa~(8
zdK)Zyhlw7Gc`v~PRP$eHnenl^g<Af1ciCYv*0$^ns4Wr19Z$mrQLFmg*wUx)+ibW!
zw=fEQuywNBO&&!%uUF7rf=fQ&OpbP}TYZ$)=5?xif<e$B8RpS<sa5?mRHAFWi#pq^
zJPRp}3JTT;)eM2YiMtB|7Mm|3H+&_&+(vE&?6(dc?|CC&QQu6%2BvK_$(@@h(fU+l
zCGPtY_!H%x$9XW(;01$0vnRgx#Z9E2swgLV{<U9aaX%l80@L-8_2R9IrdQf;Zi%_$
zxlM@4{YJrFZWW>-hy%^G@<ON6c6-H~QcA0zfRSM_ug5Z=SbU&nu9CQK*7uty0$vGM
z2Q#=!e9CcKgc%lwqvc-o$V<_heP@y}b)}z1vL2F{uZY0uy-TRLlyXRRxQ(ZBDFor2
zGgzF3Yl?-*%BY;CzT^ECu|q!X08+JB+N@VPXAx7NIBmG$H=t(Wz%ZeTgcM&{a@{ky
zEyC74Z`%WdTa4f{8$M}`FY1t2q79w@Qkw>eMKRyK>)|G%=a6a+fvi7h*ZNzhCl~V8
z9&jV6qf7;QrAQjon05*sRLs96ZV`g2=%+=rO33ex*IKLV!;7c~GR(zg?tMS}lI1vR
zv1C(>h|rLt>leq49oy3f`*6Q#-U%7(D@s7<8uo_oGw_r3EoLbt-%69t;MNC|H5w3g
zan;6u_NAX)S|k^J9A#|b9h_%>jq$-d3P%`yeS9ATW0H90>%O;j>1m;T0ZEp0?t}#E
zE;t7Qh1GrJ#3i&XsrQbaT%iTdti!}-R}~Vy#5ldz4W}4=;OoSe;EN*K21xUrE`-%h
zSp2jMfK^PGWxmV+71WM=|16TVZ>B!R9aNR?Ey@9<!?}&VkER1KU7xPfjHEA<L(2>w
z|Gs2woNp66-qk?e^NTd$zrQe)bbd#71D=ai7H$QW75>SS)Q$)hM$M0<;o|41V=zC)
zddh;vhbY<jjlazSPR*gf^@FGI^(q1Q67PAcP!(ySV8UH6x!DZ*u0r@|)Ep|<v^PdR
z0s8UCcfvc2Mi&<3j*ZbYKGA?SCm^9~t_q$<a-b4J_dJwaEnUMV5btwRbit_Fo}qMD
zu#DOU?f(*YegXgM!9VF$iQ6)J)W*I@v}NybxQwXMqRNDc1pH-vht7znb3JUH8q7!v
z_2ByPt9LNw2(8<RVN}#qS%=vFVzfV-4Us2V?<DXW6EEfG)Z<3XcI%S*$CvG(9q^EJ
z@G^3ksKZp<cNUS<@G-bN%0n(ZmF9&fhdxcy*IaN&=yKK*YV743H}Rc^9?e(sO%N-X
zXBVk0H4au-jp1JRv0lFDHN>9i{YrjF4}6n>68!-4H7*~%@`l$<9)X20if*`5%Anh#
z-h_5~`A23LT6My#lP~{~;cXp5d+4j4@pE`K5htD>vv*!SK-n1-_SmnfoF}|Q5eimf
z!RM(Huxk1zmvgLU=-^q@J8I>YXWvTI#4>p2RGxYk)TdbDxF>Gvupn?wCF2$3Pb1$0
zVws4L{Yu>utzt@Cv+Um34GogOxQ*#mUG;+QP1s+~fG$*VoqqEVPUtI>fK_?lK&f)!
zV$qc)O}BjyU`FkqQr>i+n{$L8M!5@Ovbb{;HCFWbo$J)&)h_~-WY6&|Yiwn3+qS58
zJ~6oR5h9bJseA#NpQpHt!aY}=v>$Z2am)39n_4T*M|aM<es1g#%UCu%H?(z)Yjfs7
zLIU?srcy+y$HD*l3^QT{!GTWl0&FX4e_S163Xo%_^z-Z&!3b^yMhL2o`tp~ug(*Fi
zXwtUTFshcQ*cXW)V{2WdF)YIg_U%b=>kl$DFDcuY<cHV*=hI4t=kg>-{OZ-zH^JF%
zK*bhLfzoq|P|4s*QI=1iV2tEtHeM0`viD~MJS&577>yEJwUv`1u|5Rn@nOBfds~2Y
zr&SwG!%4LY%6}~Qc$1|$Nmj<EF@_`XRalX8(0nfHEu6ptZ^!!*Hun@qy~;(ZTEVOT
zR3?Pw(yLJ>E?;pCh+krXgp-%WXA1$Fu?d53@I2w(xGbPe&&M>ED_>67)+_DkX2KZN
zmD_H@iqn>Ah+$v^>(>)vY1}5H9!vFPAHL0I{!Oi}Z(v$|LVsJvODU*6>UAB7AJqEh
zbKHZrXxfU;zhIq7!3NY|Xm0s#ad!o0T4O`_^`md${^;Qb!oTYZMzCcJY2Hjiw&O$a
zD7{hwKN}IwFVe?5WO(hv<Dy2QhLKr7GzBUO;I@lr&>{uwhdy~^Tq-vzEEo~9nIrOa
z^%U`MiPVedqrN6pEgN4h_D26!n5;w8Ay018Mt*ch8g43HQGS4AeZB=-2F&~X(`{S~
z6SOByxY;0w9HIu@1G)9CAqf7zTedL<r9)=)1{O{Ww;C$3a<}i2$=<E-4%hCj*ZtZi
z1C2>D9<Xy%^2o8y4?-Qro8*Ue=_ilv!om8?Wqkvc=7A4fMJcGeR9bh^fIDDkJ|TLV
zR?D3%2f8IhE<sjeE+u}mI+4iPNMGIH#x4;uzG8<DJUHzdsG`*h6q#x}2c|iflLS=}
z;JAGbJlqW!F;nw*a;WI#&hsgFj$&#)yf`6#`_R~rUGXENyr_Z1BI)+z9gAOKsm4_M
zTiv@L1<7cv%>(K7GU-W?D`Im1l1+JAR{%y0a;pKq{k{Qd&f?WnmVg7@kIwv8YF1sJ
z&kBjf?&$(zP`|xnH*3RFZek}!eZs*xEZQ+=al9+M3SAguc+NEgl0O)yWjNe4?wKcY
zoBj+`*^|73w3%D#OB{BsuHp?XxHEI#@>3Fq0OJ<gMChH;_qbG6Mp}aly!YqtOyrcJ
zX%o*iM6@tQUO1S<ik`I@h5gODgy&;#)6EJc9Nhp;ZfQAuE+T3&yj0MkI9Du^=CjAM
z*QzN45i*$IKR2f2)^u?KS%?i&?AX34W;eV27dkBr_68XT*dtkQvG_W~H_5M@)Hm6Q
zBA<7SM>r!q_RmgrTltl+j&d7XDr-u)i4}y>L3zQB(U8Sf)2E>I#>q#q#t8q1$Unpy
zNf`w^lw*V(R<N#+qQ5;64V@_krW>>Ri{`;??YaDqxWM@#{!9dQZekoEaq3;}+6DUC
zSf}X<Q_Mx<6~t_CTi0E0#=OFcnei%LNp1MQyW*4jc!M<hJ`V}aj9^Q1Hm!Ql_%`=v
z9ZhV~h-D~vr4&@a;3>JpKW5_*d6pI21Uzjj#CX=Ybg?&K(uw4*<VxRh7I|z+@mW2p
z$sl7d*R%-Fmd;ovknwa2vD*2k#zcDpyq0@gp%e;xJj#;W>t~oa&~m&=D&V_sy@Wqn
z56K#+5nBV(Oc1_^hhhN&I)&*B!S_aCxb<xk_i%!PNjdh*8-MzYO=QxU29m9zF?)I-
z(%twjzpd@<EZQT<xTgH%gH5}3WG`Axal^k{*>h{4)tY8jjT#nZ!F@4DqZCi(oPLn%
zad0#52vW^jd6##e*%*KOFZjE-uZVdl!gp&=@1<?^qV?diw-LB0Q4K{MPX2;iy6*d@
zsU;_00!pr*or0w-eX%@T=(^79<b)nCm*k0V{ZjmR+ENsEq-O73-NWZY(oUXlQZUM6
z_3f<K9z^0Rt}Kp@IH74saxnQe!)MhZ0(Dltzr*#L)x@eNgQaL4dELyLos0I!G-6!5
zGZ!#n8@!-ked`oFr$GN$@(4UvxIX4aH$Plt<ungBSnG*N7YvVv*z72us>kS|e|DVw
zYOkBSEWTVJ)p7Va?^WACEPRa!8*;04CEk1TO)}>&TfXD`mgcXFiR8VWg3faM>AnlR
z7Cz2a?X+DHd5dCjmSDCLm@=drYyulPqMA6Z$OT$O?qKKtT~6Zd`|pdO>t{X3ZL0o_
zrwT<<b5x7FsMzF;Gf#;QJBOV$3bCW_xQ%qdxZlSKmL6A;;POlCVgtntVNq+c0_O(x
zkP;?@dT;oPqr~H#p!TD}pkH9(LU>1ik!6Fr|I?2KKk+fiL3{e)kD_=!FJX%f^3%{k
zsIgDP4e_&NS_r<iZ7KJPq_}2%Z$yVgihE%L?`TUMD$CfY`D;mfS0HSZ_2s?5)Fvx8
z%l;^R+<3!ZJyeK}Ty_>o6Zxq>3#5Lv)=oPpZ}*;sJ%`|4LwrLh)g<(PN7ggWf~fc6
z=b(nbUqlZW&O{!z7V+}X9S5Us-x98sRO;ye$J$#)wG}=6!q5Q0rL?%yBE_M&1Zjb`
z6sY1}q{ZC{F2$h~x6+ojxVua6Kq&-I(BKIW2%22}_r1@(U!E_|TJOqyI_KofoH={X
zp4n$+etsQ+*OVl*yR}3lW<*Lc<*MzPhIq6)oyvsQyB_8uwqK!AT!bnyBm&H1YUIv5
z!Kyx7CfWzK517-3V+j5dpo;Ea1LovQ!q$(3<^OQ4Tyefn>QL(Sf2(GJw~5}Pjk&jv
zVCvghZR2Z?cuIaAgDQ+^k)JjqeEZq~SPxXo-YsTtNZ~_Mzy56c+%?Q+a!}D>8GtRM
zitX#Bqw@SI7H%H~ucBDJme8E-ZqrPKtLoGA@j>J&LGjw(k(Qv+rz18!lpfl^$!;fK
zXZN2G=T&P62DGU_=85H_h_e_Bp*+3Em*DnWb?8k*vem23$`7C`J>O3F?uLY}TC8tT
z`ubBFN_{=yR~N5R#MiCz%Rvk>x#{sICSP81kw`!gA+%><Ng4<?0J~TB%B1eEFFI-z
zZzg-*M(h97=G>@e@bF^T*P;?<2<-WYqO)nF=h#x>8OHLWDVw+)+1Tvlb)gViy*WeJ
zq<22A<n{MV^<g$GrDCZKOMNZy4GH?DPC3S^f*v(PdF&72zO)Y)bhyPvBaqDVk_Q@V
z(-lSCJG2C0@zqJ2>_KON1W@APbp@jy3BL|C#Gc2}IPKQ(yA44!V=8xg3WTG7yGqZy
zip4O}QFi(ru-km{DGsYmJd>wESpis4CrPbDd>N(o2;U%dOnh_6>{PKhaVi~>Am8Z6
z7R<3`U~SG%<a1B+ZOoTa`7M|bF|O%+-%Y0X^5@B?l<_usH;EY|gS^8(z+I+|y$RzP
z^_WmE+JEZAi@cLMLxk3eVrR;HsrHXo=hm`#Mi-=!xziu>ythvjc>Gw7#xm^XMC~Xl
zjlNmLD4BL;ANDPQ0atTz+o22!pX8;}exH4!MsdFxuqVs`3aZQd;^ZE}lie;u)cHrr
z%=+{Mzm;2jZb?En1TnCY7F=<37mm8U9TP3Et1=KDGGqxNnF!arSNaT456}UiiYB&I
zC1%Vc6bSylBX!_@H{8ML$(zp!8zwJ!8@hO16mS?)FvTn>uKcE8y}9xs-c?r&TE!*y
z<Vg_TbK4&%juF0N2J}j_3*;omYsqVMcMvqGhD1NGQ_r7LP5i_iq0OnjchStcb1rx|
zH!$5Tl~-y6{AoaeX_bbE7SJGwLA+woQY?dUq;MOABM;h}6f=<-`n35>YB~9y*fWmk
zs^s1lG-*k?#PYL4W>Px}Z$#(U662vCW$Y8Twa0Hh3x~J8tCWE5QVNo5K;@OR+BrOw
zTdwm3MCJWoNwD;jO$_UkhvgMPAKvbR$5>#z92>hUXH4-&Yd;Dga>;kfqnQ7yGfw*7
zQwxxz#KuRGkU4+LCs6)S6s41Rm~;FXm}keHRUloUp#%+s#uSuBR>D>psdTl!lJfF9
z^E3-3W+p(x=yJo;)9F-S8dfbH^3phNzecSi*thwV__B9GzD&eqS85qjd?|5|Vf$G<
zPf#!0eOU$E&2k*A89nh-So&Sa$b?qV2}+{q;}?8!yNIP`en~fMcS2lB^|u&&llmjz
z(9YKT;!{%A$sH_>1vH{8yTE3$n?BE`oj7`h(5CgoN)wpJu8_n<9RWU8^hNj(1ldv5
zOAGsjowR!>e&A9rMC8>^!U)Y}_j8C8Qe*dS-sF0|a`R8P^d&A`0;>c(e_eEaZ*|+c
zKhh%EDxx@j;6<=r51U-&wtP~6U-(1}vjY8`Y1ZN06b5~o_PC%ke_-!Pyu#lTGWHmo
z{7gBq7vz0|j$dIK73@_O0nSHA6?0ooD4F@wLdfPjkj?fw)1|Ta9|8RYQB585)uKHB
z%#ln*G&D2Y{BJ1YnI<$Hk=t&8CPFeT2`N#pNF;}0wHl~}Arbud{?@Ff1y$TI%dsCF
zkwcCYw1V?Ix+|Ao3p#L^+vI5^SeYZUEP!qkM-3#GO`7+KN6JV__f`6H5N2c4N;IAb
zW=*k($7(y=subdkq|?e9HvorX3JnEIk$c^qR_F)FhpY;L+Tbej=<|gEux9mj!Cwib
zh4!DG6TsHWS4)2Rv@2+Io$XqnX~e6Zy*yYu)#x4&=0V3cc9up$x?!sA_q7TAxFiOB
zFY45C47hDcVyAAk{cy```#$R~gPrh^{QgONS4ak+3Q*D^dUU!+;z|@zwaBz;&y7Y2
z4Iky+ex`#;i#oP5Q(R=#GHMlWUuNyus07x^En=;YUV~PNwJ21fQf#xmI@t6C(O&oG
zR-{W*`RcT=(g!ctmaB#xmoD}axT;dwvS1xV6IUJ2e354Dwo+Sb2c*AH7HLIDcU`W4
zH?Sh{Psh=k9F<r51dqktzR4|u!+i-bY7`%2^!;d*p~ch%c7ac1Bm6_|_n1+8HZR`s
z%ZNov^cu2+A(cMUbq>H>cZXmJm1?0v118ifXPyP_@+s4F=T9T?E_L7CbNUJyLkf^9
zlKQY8T>jIatT|@)_Pbvo@ZUFH!^xha;(tXU{ehA8#bNdWQN^&96IiOAX3v^^o~+<V
zDVfV+kTdD@s(<=4VHT|EZOG2Xu|c>+RP6&fRNs+n!i~B+wK-J&i0-6u_DK&H4J^b5
zQQ9rUCdY{BqVx<U5W&PVm~W)a7eQlst36sH1rn~8`b@t3c$E|s;;Ja;3|96r1|NtD
zi>bk6wv8bdfn7dm>UAMTl!Bku4LfqZMZ5r;)x95KHX7c;bDqmCH!xK2T6u`t{0=HI
zolS|={EKd8%5?aKGr*6!#IB_DU6)Ol*Se8SW2pdak3D{Lbj@tQAPf88h4~5+0fqa5
z#gsZZ>zJDvZ1@5gqddc>2QGRF?;~rg09Ed~uN)|DA`Rrk>KLBTckpNHJyBa{WIOkT
zUYuu4%vE{w_+v~NZGyASnNYm9=`t(vJy$ZAnh^Stv}B4%_^tyy!dPC+FL{02Iza=3
zgk@Nu?#LOk$(!Y|2GMh|xA&M3kQ7CQ8N_?uxV`_LXRVB(GcOjVYs$M5ts-+F8!|55
zk<_scS-c30#V!{)YAEMn&qEzq#$0(2mw}ve=WYXr;Sigwmswe6Zi6eE?IjmkoY5rF
zArlHicgGmaWBk!9TBlq$1%VLFljpec+onZj$-0?8*Nb46CL^}PKm}cXwI`S(>8%pl
z5G2BerV{nk6p<WP7=MxTa6e~qkH2zr?rQt8j*lGS<`VlDErVw8Lce$SK_lJxxVq)(
z0<xkw?0_o*=HU#YYQW6P;lMPM2f}@#L$5`LNh6Q(dXVtUj>o(&re|mmotoq`9)6w`
zffJrgj)yOrZUo(s1ts*rm>}z~RWzZj%GfFu$!^&5Q`O3=zy4FijyCR*kI@tZh>s4T
zXqoOnrl2=_#o{DjgFDHJgB&JDgh&ncD!uC^Fyk2UIt!74M3p0L<sSR7uV}~!2w#8h
znjRksl55d9&+(k<fo18j%N4ZE5Ff7lR{I7Kog8Zn`u76A-0>0zsh{214m}q7*Y|Ou
z@78~^>oa8tARtF$DUbu}0-}XLFcOU9o7_ZoT<0E>LEQ86udSFZ!J^GtM>m)~D5Z+8
z$Hf}w`ib<a=Y%~;rEXCS1z#5sADJva2`F#)ags3L*%d{ya!bWyRH@l$p+SV0ypznB
z?eLWVr&OmTq5T)Vd$XQ&K_hM$_I<QeKaw0O4Le72uC8;=Yl#)cScc#WuT39ZGpqO>
zWv<zZfqekU5Zd6N!ycyyEFmnAtWw#TFfih@MO>YQs}2$DfKbi?MEVl`jacdbYfwny
zanZ|3TNu@M^<JfgUkrSVMuzTHRJP?-cQKwZfo-oTPGobztt~niifgX1T|7m}@pqOY
zDIS9+PtEU$y{IRi>f;}{o`VcrOwlm4N3~cH1Z2@GQ+B*)xt_5nJ)R@^hvcNfZEV6V
zczwTgF%>G(!WQH4#*WQwuXHj+3&Fl<8X%sY@H9A68rO5tmQ<D;AKt};*c%TL4lDUV
z5ebz8hX#i}M$;iR91sAa$ArqXq86c$W`bnzRj7i6r>Mt9*E1D-{6H76E(HAUiAOBk
zbvl|RqN^=#LyxxaNEQ$6!D9*V2NthXE6}j-;|cNL6H70(sxI^?zdSGbWKVC4UML*E
zDWz%UI+=5h#uHp};*mtnpMMe%kT?9lTt9=>|2BE=e|YdeoevUI7v(55-)OTu;}(37
z_&w!bn&xY5my}%JlS2yaVWTScML(&6ncbN0q<6m&q+DKjGDTbcU-Em_3C80VQIKR#
zJbN4eyQIL()%gqbYM*OYl||QWf$Bo1{f&YAAIiR~uy*L$8Tu=(rL8e(#+4sjsbxSF
zs1kBLH#_qjcZhcaUt)G(a94Qe&Gqr-DERW*8Dm4$vEU=IiA^SCs_u<x<pQ*3!Sy2D
zC*S0wvBgY{74-Q;K2A*G0(@JN1Cu$`!;7?hp-hpNv<4yH;)d>J!*EWF7{`f+peFi-
zJDaR$`x_RL$LCes8&HW2b#Mb^s%OJ}tkDMKcAKkHAzFoC#70?Se|{KWj?=sFMNw}1
zZ<vG2M7UrJjlwr$hQsgxg6*q}0`I;_&zBdxu(Wq(dn(!%x8~jd+$as=)5^UVcbc$i
z;oSd`693wJ4d+?UX*O<{=ge7ZVSgoPWV9kkbLqTg2J;urtSxt5;0lPA7~akKKf3gL
z$^?O@YgNNa@h7fgD5K-$OAgar$5xrsO25d_?SoH1=$5^9KKkofXWLioN&Yo-?kG7Z
z+P?tn-PteNk#UWW#+@*j;xKhq(7vBZFZphKcNY5?Hyn3(^Pwr14?q)U7l4;Y*Y63Y
z$u~ot!i!=D@P!|l$4#7TuNGBD;2|=co*=>CxC!AFE>@?`ePkYnV$R->i+A6L82JiW
zPm_LT-^9K^iB&JiW;j0FpU*ZQfqP?qt%p_O0p1QBh)!ovQgEPI!lY=(Q1GOW1s4BY
zanBFwI%fa;Y!72(mbOy22Ucy|0f_FG31+Xlx`Q>oUU#K~g5fHvkSoK&?g@C-F~7()
z!N&A3UZBGu6$CxI%&o0ykH-uwG9oWmc1SRBi)5_>txP_+!7p9Vy$!v*k!;LsTpsMA
z59hOQ@&aU{7W@Mwa^rh;u-Wj>d1@y0CRj;p$Y%3JWJZ@naBO?2_vVq`mQKLRPY#ZE
zxIS^p|B$|o$$sv&l1i7osq$GUvrQLgWAIVk%g%U;j{D%g>ppvX`{~_l#4Oo1H6#+b
z)Qu5QH{UyPw6F(_9T!z;g7Ql0%DC_u%P=4DprE;=N!B^AGNe?yOMmci(H`l!-?T0U
zap0UwJ$gn3@w!<GVT1*y9wRSZ%u9pA8d1hu=QOSUPYU)jPoVovC(_%oJ9wnwF!%@F
zzz2=_QL4wzDtN$iRbwMQ!8Yn+gh-zoprFJfjgL`5A&3@>hhk7-;Ry`b%`|j$xW#{>
z`u2?2<nWH7m|ryG?CON)>9xfUNs~MX&t9fKwp1m(+-7=U(y{Izv@(B(sSkPDpD36A
z54wC_Xiu<l)ipX&qZ7CR`rmpHr$jPYARMJSyOGf<^=a$pcM4e-mj)Tn12$O7hR^R<
zMC4yX7|*Xw-{96Mf|M!+@zz0M$pa%u+0O3&8q_YL4H;g0(QH=WyCA22thvT|2As~v
zv7&sggSQfKs|?$#ltV#$y_AxaISjbS+St<Th%W!&bYTPT;cNQzLMS(H(-q8Pdwi)R
zx;|21ncK5#E+7>!b__k+_9idv`RTMMB-3^Y_o+Y3RmzTr(e_ovP<1-Wn&HKqrfpbF
zEN}D{&Mqchf;(Wy<DAGc6h*4m2z)Q8h9n2pAK;Zj*KF<s-;fjpgA@&%2R%-u^RL-m
zH*C&0>k4tt0`@_domt7LdWYBmKWsu^3*Q`s`zp$|B|sGS`?~EW|9YD7-vG~4)zrav
zFovpSS>pQM?8!#Yk(NCSe-J42HyO~B28MAsUHK{z-Y5raEqClMXb%2LF~1b#Jk1Yl
z+z)uh?zW-Tm2pi`wf<tCpnNuOB5*wL7N^Q~2t{#~qm=<pm5|(lELK!O-K$EoKYBac
zF=nQ#bcuG&;Fg2SE64Ya^D09MtoHxwjaBEEc-MF33ZZ4s-}tzf_o4Q;W-HG8C!&z#
z>>Z$$@KkVeFjHgkW*&>-1=9CmpfPp9XAA`2tcL$aW1pb$VaXT9i|J?)cd@S^c-@h5
z*RB%X=vYW&LT(K24Sas90zbdirqVb)ujz<SSmbfXod;ACdQ^11+IA(AAME$|DSRG#
zSW)USZe*g!c2MfNr9dbMR;6hcv4oY1%*n=I3j|KE*1)TB-~2r&UA}Bnt?F&qgr63^
z*g(m1ZXaIH7Ih6br-M~!6bfB>q;~Ej$t_Iq7LUMt&R*3v&ij2DDjV+K@o#YTi8|@$
zv=CCIS`hNClFPT`sVWz4e6NZ!7P+i++%5wFD_*Qa$}pbS*lpa>|6Q&E5q_fy*`@>F
zmCwspyKd}$$VQMjEWVwQ7cX4^N1O+2L8g+kBW$|Ne%93%`%#-;?mERBw-;s5OhN?c
zVRjcac!dzo!k?ql^CjklPDLMS23(NREb>)#zZnY*AkdH|Q^k9WARnAdQszP188S`A
zcFf}keCnp_<el}K5TA=uT#0!L_8q03te(J5kfPAV8|aP}=K66EXwL3{vmQT*ZF~fR
zcqC<F4ML!jP}c!n+}HWSky~r<1}Be)g?vb|Dnn)=j}K>KmWJWZ_q5|>&i0+V4e=)U
z<>$Dlo0O0VPto<e3jqJTDSV1^#N+u!y`=-x{UCJ?e8q?Bf@AEt81j4`n1$=Hmc}g(
zcMbXDmd{b|{_m9kKQ{6Ivy9^HqW$^|nISYpoVZ7`6B4RgrfY>=8%r-n&?~fXj~9mc
ztOid>@LG)k{Kqk}HR<`@eiK8h+?S*TewDQ}L)Epf?bH*m%CBDCuvnKM7ThsO`NzI~
zaTNR3Si$tJy~AdUkrzQ3$*DdQj9e>A63C>Pm+dmNj-<AQwjz>{T-2rXMWKA=2|MV8
zdD7F??;&yW$IzNrsG3TnN|<Jg+}AxY9gQw%Gel8A`MFXiY7ver7TNbAIBM-r^Vd~d
zVTJ72x8XE!No1z0SGqTAh9fU_Gd%_Qcg$bl?M*}0kdnc<G6%u`>bja%QGzWTOD<-5
zZC4&E?B=Xv*t$<KA<;Zdhk=}KTP<rBlJdTxGGEwc#3n`A3IjxZ4iv70ZXo!QliHVo
zD+ZYyH_2DejasUVu@xh`4c{yuHYHFU{Qd4EpB&YG7nyNT(3s6OQw8PKgT4z8+U|LW
z2Pq!j&|i2aA(uYngmCyAA#|H^W1i$AWgm$eA<I~Wd=#erZ51)+a-+dNLw4}XK~YV$
zLVc+#QMx#Ja0z0$J2)dyeXxMni==8N3+8v2vOZJcnQS=R<ye*QrkwNrLGHEK46@Dp
zxw7jnrPfUsS}A^}O@|*}QhTCPeyl*V@x-iFX6aM8U4Hhly-wC1Q9!;}mC#&=Nnlbo
zoW0cpy{v(GS!o1w{~5BnZPLaTnCu_lUQ&5H?$1JILp&6q)>XQKX~U$hC{7>xq9V&`
zLLu^*<@5o=o&g43U#_$(^G^0p-p5$^N#O~K!Y{cHT@P<@fT(Bxp92eK{U7?O|Lx`f
zabU^3-p(*X)S?J#$<qj|${rKa(E+Rhzf>~>)J5L*|7nQgArzd9RehuTE04Fi2F8%_
z_5By}W;WR;68&Eua@|+I8_vokO-@SsjH}qG3J%AD$Hm9QQRBHEB!^{@*Saek%U8EQ
zyl<xL@RCW{Y)e%<Ro^e6e~hB{S;!M32i568kL_}z$EH<(X(@X9FL#G2E86}mb^h1X
z7}^Tk3M@UP$<)J*oC+^K!rr9;0Xp*P!!nROElJh^*PSi2BJ~W`QH|<~#T|?Xeqvug
z+C+X;r&w^9*!$YfI8-3kJ{tRTnUBNyZ48K2p}Z^q_)#xEc+=Rn=?Hz_R$5BurijMt
z_&&{Kb-}8R1~e!!3%H|5exz<t+~!u1!iB-`gke**9u>#a@|&u$viq~&_9`W=Mp8f}
zTWihP=QT4aGHZfG*RN!#cR7Exe?Hz(-#jboaOutHO*Z1g#20JE68E>=2WFy!ZDLr|
z@dBTc8#RKw>=smd39M4Y4Vic-PpwH&+~caNn%u9NlWZTN6;iPzB>$TONwqndQ!Xd-
zW^jWD+}B+nue$l}%dI9l%$i?n-*sWQqzE=zk;YJ|@C-_bJ}&kl2rl}eK_y{e+3qkA
z<S<<*e&e>Pvf9)uXBCcC6yrm4{JnnOoO_`^QSpHk>?=EMdaIst$-!!suA=1-OpVYY
zu{TPsFx_buufF;2`z@p36dw*+&FM3<UUtlmW-()M;~G~^t4T1&`m1j*scLi`dfW4V
zD+gUQh(F8X9%CrD%RoxeVb~q0ozAC2&qp!bE08t@e`%9LF@Hk*cl`dAIGB8qq_a2~
zW7|=zc)13Rf>)2Jo?;EmpI&WLEnQ$v$yQ&j>h*ek3d?GSh+l~e5+}f80b?{-$>Os%
z394P^*|%IZpTM=2io~d_J5k2<4(@Ua3YXet?<+np)PH=##>#5)x6#?e0>SZ@&3U=?
z6e}Sip-4i-f{C5ee%Krme6~N|9P>DAe_qhW5s$lpu-f4@0*myl_u04;?wjkpcu{A9
zce~gaURmvn0q5oAwVMoaD*8AL3RSfqzkmOJmg3>Vhfw{pd0|$Y{*Gw*0=UB$=nL4<
zjs6*pB8&An?fBeIpGs6#PcN6@BnQ|^7+`9j%%QxH_|jDu_WKQgqO(=v6!vvs{VSi>
zuU`v{mQfap64hUdOG|rybay{mXQLk!U!ems0bQ-Sx-9(Wen$J5e;RNn+KarBGjFcg
zdtq(O+;hsmKA1@1ohxi<CXxJ5ib0$MB&b@Y-vqWj{}6MB`3m>QJ+#H@o7IGOEIyUP
z2IrK^V!a&O;OU!mW5}<x`x?J~u-F7P(P`IlGZS}T>(_7U``p{B?r@_uc*VJmt?SKh
z_1=rFEgv}^CfnPZ$XA32Sa4c8V*~q4la-v`Xef!Twd@!t@fz}~Y`(_VCb274GSa)!
znKJ4YD9QiBfw??4?$+EeA#M5(c&wsEFJq?pOFxVig205qg@R*}W2fsU)EI~-=PZo|
zlLm!pBb@AG8I|>#S@>SCprBy!+0FFPIz)_dF+TCQ;Q$^fd&_`fxAxI!siohW**K`=
z*%SkmtZLyOEi~kS>cq%e0QonHdlH<UWvJ!WfiI5XAa8(Q*(Nr9#^H>*4bk7wDCfn-
zcagY^)i>gnm3?`baJ1g4Mei-*H~3==HM=>|u<4&T%+?_$*?v6dT{;GczIge2+K~w#
znZQy6G8Af*AQ48kC(E4I-kG(1XNIqP_3Bmj>r(8q_f%7$h@;LN&y_%RcXf63m(!e1
zVj?2r*2|4B+8bXdcqhB9<x3p(RtVvyRyeLyGlA&{&Mak@_gqJZ@+mtA-r|X*{tW|1
zwzGW07ub$|wGt^14dQb0Ho}fze*3{H&45-~a$I@Nv*p>Wfib1J`K>SZH!P{%YGCj#
zBVPn#vE%<j?1^^30%BrS`uq1R$9dSE>*@L0poZ!{POH!<#)d&Z#o}x`0<qCu^9fgC
z9v&V%Pw~MmEiHj_2E`^lJrTg=IiVbt;7eqEQ)%=z_~*yzpMmos=o^!d)7n#b$KQ^f
zbMrmt%Llc~$uz7oW1r$Xq@RG%6?6FK-G%|e($dneF%ZLyaetwEp0MA#`3k|@d@2KS
zh7lOh>-X=Yz5j8q_Gd}h@H_>V!)knbF1<^vDi_~@k!(;XR4h&Md8CxPyL;+4eCanl
zL*I3|=sl>Y^gDcdJfoOshUAF8YbI1bV`myNKR@s1(u4H&3t29|1P)}^*Vm`N9u&+x
zt#m5%iN~$TyaS7U`}Xb5&$#4dP+|>Sx8tGnT=46g0~;-|xYWAWgY(O6zCfMD+Z~na
ztM8YKkVE|3YD2*3Zs?BPL>{wG;al+4+|1tXu1n+Jg*dvIuH9`{d!44giEQ$cAI{$5
zhyHQ~!zYZ3%gcrLRes|pFij<8`JFW3n3tzCB{-k@272CeC+*W@rJD%Tx~l}b8PG8X
zIyOcbhx7aT_J&sv&js5@hPgla&lP5~{P!B)f@duL^H(+8@)ldULG!Jw`ForjaOb!9
zPv5#0?NymUA?K{cyPVmVzzg48JnwqU0iI6(rmO%bT1dEzOHSG4w7b34&c>g2Ww%iW
z|66?QUrzk=y4^~nz$o+<vm3~1aKrl?FZB(y<utbJQa~6#GBT1B8~5$e_bwdX)lxSe
zmsKa>Tj+g68H<WWC^p9YNx>Dy;?D8l{f8OcPOomD>)*GN<yk$kAC3BmZ2<7HLErMD
zo9df)+IXnX;nGF1)uB-+iop5J4HR%9^B*q!?<-B2o!WxNWiDFDce=DN@CkR$E1wT#
zIHXl@6NRU5oGH+z<io#TuL9W{U)XNrZFP02&vv@8I9g!aTl`5k2eS+E(DLl*FUp&)
z;qmdP_pyXQG(mbFKmOhYi^Ah^E=ZTM!quypo)AbU4+bxKJM2Rx;%->ZT9(?16n2x9
zUn?P9&bR+XBRjQQ-f4sT&r~i?6WzQ9KHqmat#Y{n?K9p~WJT-oS9ak7N@c3WC}-#W
zI>~(!-@Up51&jHVcTg-H0^YM@!95|Oh3<!%rZ%F!pZbh#+U9MFc`!{X>HV!|UWF>b
zOLKQ--0}STsdc%YvRmlj(M&O4V@6Eye&UW@a+LNQzDD5?AKhRgdL^>uHnypI%(*vH
z#&{Xhx_T%zt3BCud$u_<@PYFxTfus@3;gHz?}!s=NBCLdj3Is>0Hl7~bYr26)eg^3
zW}B_BoXC?u4`N4pm#-?(x)eSN-)+U$oyw+BmR!3&&^^Nr;JUZ5bNn2E5mi+`FH!m8
zR_fwcT?VHoAfGrKQy1(nL*<Krv;B{<_~>;=vW=@l9);xr4ae=B?(Bn;l+k#rtSiK#
zGQRjvA^wE!3gt-Tj-L7cEPa1VE1=A|{DiRg;LWdL(`L8TyJse$d+%g!%6E}q=eb5_
zRwpGr1nlhLi?4IT@t&`WZX;YiDDzf}^Xi%47X+m^#vt!1vJUNb2D8D;=;Ho_;NUs8
zb13hiHWrb`n-~9)Uu>eybVMcW<SkMH+#F0!Y=J?+@jGj4)_H+41L9B>Hpjd`l5gya
z?{2Q4jFM7PbWo|pV^5!hO1%>j8jfqClb@3Kb+t=|!xNC6%%^|bmQ60pjXH%YH_TJ)
zNZiXGTUe&8?72y5d(|X~u{zp=xl>;pD&29m+~4Kq@>I-uu6p%!JXf}Axz={{%B0EV
zl=zT)d^eF*R+O=W;r_jQeuWu|Rlk0b-s0pHc6i^pyei9;aSlI({wyuE5$?^7ULSg}
z5TtpiEcVfyA&`k?<$hrGMf9Axptt$g)YPNxsS+*Dvu##6F(y=@d6NtKp~X;>{>28-
zIlcg$H0N&;WNc(;=rC>;KUZr1%C~+%jOrx&pfuY5P^>x1J+$r{vcjy*=OY1C*~kUF
zxd4K@Af48=4C%*+_bI`yL}w4wKE-t=9$8gtDeEKeeT~1^*WKz$N2UA23td;`@QYXb
zsCmtcJe3UF#1fPPAkRBE-KyXsp@DTGsB1yIaNSjMKR+yKFFxu8l_zDlCn6)78Q?RK
z)HG4f33)s#KQ<J@)S5xLiaNAou{d)}#9a^sM{fC6bD&FUT6)B7f&>-`&ju&WC7n%j
zN|ti7OU1w|V{nriM3rbUdYEho%jH}|<radyiJpg`KA0@oWd*Z$BK+AM8-;$c7OXC?
zZW*Bs>MGLiIK(<;-}&@(SWgLPs%ydoto)Hk_|H;2m}7T<c10s6ELxuAA<L$Hnh7$E
z|9C|ebSA9$3`1vp$`CLKogg~$r*b|k>3Fc!k=O{VPOP+9s2qdM@Vde&&KVvR=7k0|
zO#{z8DfWCLwB{WsFeh@w5mrg_llzi{tgOW}5GFe|ubN`+%3}P}<GKQ?K~>L%fdH-o
z=h^&5gs-01T|X#GL~C`3!K9{^Ipk>NUC6$XPo-7P#2C<GfoV(XIpyAL5_U~_|CDRl
z2<ex6T~kjtihXYm?u;7Y=Qp%9R`EJPm(Y(ojruq|>o}UH8gv<3_4Z3_X^(nI<!%-)
z*01^{hN@rf@j-v>zK;_~ue^r;Hi$9oVyK_WgNZ0(PHd~#YxbW=SxC*6(tR9*H=tj2
zHL4t!xIX@o_%jG4ROM@?zM4)ijJGODG-w%DULb_-QT>{EsP4RON;Q>y@M&&>x9YSv
zfZ*tDp$|(yec*qETcM}FUhE$EhVA<={)S$EEo|^Vs0!)awQ87tdd+<QgIW0LFj;4i
zs=Rl*7d6J?Td`U0<)5Eo{h{9kKL7tpjpv*b_z1UNO2+#H1ne3AFTzcl`)@<m|KY*^
zM7X|rZ#P>YH^>EbYgy|a_3sg}cR&7Z^LnaQUG$CTmt_ybsPyi;Z-ZG&xw(Pq4%XaW
zYCEsL3-+FoIHA7Sj*d<2Q)<Gm0#~rg$oA6lylgYWlNMw$3aK-nFE7=F@-ILtO9_|T
zY018?D^YAtnmKGQm^s`X_~eUEn>lPD!`L*ym&SI_s7d+^mpfC-L-2<03ndGrzw6pj
z?Y!xEJewksPAq9(i|)DDzv{z}2b?-UbQKTkzqi=@O-zsu>CPDeLvlt|Y*LTmQtz>h
zGR<JfcMjLy?Sh0Tkag-GP_6bRxW?g>)BVOq=Hvcp0q1)5#X8fK6|tf_MvRTYc<irF
zXDY_gv@M9+243}n{}X7ddPwqXgYT=lm8K807ru2Ha8`X>u+aewV7M$Bk%HHMk3ArL
z-voZl#spz`%~u;;Z<2~XuvB2nnotLFI2K<Ao?Q3YqN?3?c;wKPLts^@ld7P?r;k?F
zO?=kgoNX*k-3p<HT87MlGHV6QEimHOWWh`A*O|%@kO)0#&T4uHNEP=Y<u+T%9|!R@
zgCwE8uXh0(l~X}%bHOF$guy9bun{0y?qBG$&wwUgyuT8j`T)FA9?SyTRoK9DQ)IRe
z_YT~)9)6lIgO55r^nUj`S=PzYKBQgi(TdtjqsV|`8Jnr(c}|9L%r(<qrgCC}%F_eC
z!af#|^;K7Z$PTPI;3!=Vmsv*-!4Uy*F*QF2gl(X?rAqh_Z-eR*6UE+$JvFgZd?o9c
zGJ$l_``?78?IcUq7(x$%$6Mo8@OSt)<shhPJ2HRHtsO;@{~-T>Qu1WiGkp##x6=2t
zs>W?~wEXu5)miX(;XHkELjAQ9qh9;@Kfkl!K6uKNiQA6%1ojkc7bx2`7EN=wJ4Bkn
zy=-W;&^A{JD)DJgR5_Lj!B+``PR*9!iEy!P(N|)MC*xmXb2LNO{Zk;f5Ymu-aw&N7
zRp1uV&9XY&<K}Cs^1L&BKwV%dJOS(e_-edtdajlRSrfr|!zyGLBqL-o&A{Ph<=y5=
zAUndr<=9<-`zlXBy&nM(+~_8{P5LSVE)BE9uUMCFftkxP5Xw^3<_6tI6%lZn%=SKY
zVlh&3izi~|A{?ly1nD44+Vjb%E-8j<*7LTaLyg(DM=DAIQPLTA-Tv&0oZUz$nXf}g
ziuE28J`&bEFdV}fop4uirhb+sdWWh-eL}Ond_hc?XiM_Sx`h|?ujDIRkFdvv*)X^k
z<v<-<>2|~3BYEQa<O_1^s<xl&nkIsemXP1=%9qI8U>a_OE48n=)|>cn!CBZtj+=_k
zsOWs_S6wPwcGa4!L8Z!h?!$Cksv;f@z0(T#z0v)cVJLe%T%M+8a7q3bnWxcum{!kv
z`4QsItlu^ng97HB;gI?TG}OMwbLhhVA&Sw^1~=WDZzv$($8d^xh>6IQe>qSUtMmf5
zod23EMAY>9j-R>Hi|ym8^WfB4m_DO<=8%&RE`olueRvg{xr!XMkS@9^V+ezXHqFO<
zT4P6xD!LnPHj!u%P4%IQlG}+Hm?xne7wmo(tJpM9R_q@(T*pSR+9|3#aIO?PM@D^H
z#BlZQNXC4W=O_nlT(VikjhAWx5Em9CDxm3|7P})n4R>}&=Ta+{5o(-c;G0Bw1na+0
z*O=-{*BDmn?+I_V!UP9(R7F2bSca9ib$1A71#`Ch5h#($a;)8`pG)Ye7bubF4P{)o
zeF)niTlY{AMS_NV^GXeyL;YBSd%50{c9Cx|M8@$f8PreCF@Pctq|n<&=)4X8X7%<W
zdclVw5B;Z#_8)RJZHh%05Z!nCzV=RSv&x15^(Pkg&~^Js*-)A6Ls#bRRPBtY=W8MW
zkBqCU=&^I}0?+z{b7{Z(*NT2oPhSC@FYxQ7<!;0Ba|NU1Ukjikv!)YCPCFB@JKll7
z+J%mYX22<&nwU2da5Xq&UL@BIi}tN1m?f^4q5QB+b!Zi~Z#0At5jP1yS7So`h%V?o
z?|)|m#zEwYHPC=a>Ymgf>#*!QWL-*}A;JW645GA|A5G4Qj$<^Q>ecYcO}*6-{?Z+v
zX5at5&Gfju_95bhAfd$w;yLp*qC4(5^8xx>FGUsUqgCC_GCM7~s>R>c>fX2FSH1k1
zXFEIfH?un)VIWN2RN_aqbq(E8E~15iSO1<EuGUh>KOGgd`{B4y(`Qnx*)6OJ>H0;e
zA>yrCPO*Ejkh0agXh`7A)xFo9uPHLbM^hi**7b|5rS2gO`7u$V|66_z2DU6-x%1c$
zq7-y7vHph}P>l+%Egrmon$~K!4Z>aALcg9sR2Z@7HbpO!^SQ8|!1p=g<*@py=1yIn
z_Pl2ra{SBU$F+7YHN8h*x;pZsyM9JRvy80GRXTu)jL!bO^i+u>C8#l4IKFpBASl@q
z{}lHbpoA|hCdrH67qg5S75gwCBx?6mNYv`7&|+^ynbB#cq0~3de<;=wmNO~?HYO2R
zv<Dx8S|<#E=88~KRUUexJj6p|SIcEe`i+mMg@P+4>tPscOZec@6|p}}=zKVU#^ZM7
z2nV>!eLoft;@S$;(&r>SFY$kM{;LS_YF!fhWb*A3rAl#5j_zFR9WAn>FGMK!-x*N~
zFIjB+rS8fR`4T}*{CX&7IOw5A3m-nT>mNV%BUc*!hL$7m;+=Et{n|APz_7Rx)^?Fn
zoA!P-FdKRP-Yn_he)-19S#N|gpcz1QRc|i-Tp=_h<+E|hBS$rbaiX_0Kgk&-OKbik
zwkXl6M#>dIElP5zC3iZ@Gg<9Wn5=zhI+PW-=B7`(rbIwv_X44h)5sHVN8B5uyUl_S
z6?%MMg(sV0i|jLp^pXJI7V8(`$vEvCCnfGw+?Tt0CWZoy5A}q+fhUw<L2&{BIXRy9
zTr^WX?_J#=z&%~{jFd_)*+Q8DChVX7BD|#mq1A5t%$hXlLQUF?LG9X%LCs&Wke)v>
zeG$uOOOVoBmAts|S%mviMD*ih%QP=5a^GTZn?J7^f0EV4?%c&BIH`#TMsop^HkKZS
zakii}cu^c`hss-fEWksxjuf_b#<4qCGJa}0n(Asz6$z7hV*$Su>HA6#{Fzc1J@1LB
zPAMr4B)6io`g=I@iTQe;p4{hv90^M%xpCY_=OM$j9(WoOT?zS-I(A91IjR=t0sN^9
zlUS;+dH&%G`@$NB@E#j9z*#Olv}*MhY$)Eb1t{J2TPz_I_>Tf{3;tb1cv=(|DBk17
z9W1%qQ;`=<8GcEWH}tl}EX5nR+xJ-Q%qt>KQ9e;#ia-=Ofp~mZ-lHr{HX`cSj+}al
z`0#5AeGVhGR@^-{>)|UgdTbA1Eu@YYt&UD9La)x<`)JbK7Y2Ijd4CKS8K<7uqVO_c
z(DQ-jAl>0336o<2$KDWGD}2QTq@3_?SlYXdZd*R!FX}I$e;e2hH@7(U=R}VQPD7EP
zoU2Jyj_+x}^#{FC1ioq!c?8UF?jf|dDz@)8-KzO_+Tv<hkpyd!Hj8Io7T-|*pnA(R
zvA=IPb2vci{^4CVF9fZ76VU`)TIB{bP4}0JHoAbGyG1&y!qOc=(eA1|B$b4aAfl<(
zR&fW#c*tZCAKJd0bG-%VD<-+;4isGrE^@dp@Algq`20R_WY%nyUBC<$_s(UvC+Tp1
z6v`<5Js})W(l2A1@Z^Zwxm}k)nkPk9=iUdMUZ1;g5-C0k$vDCAa+8A(^i9l*Y>B_a
zmwvkuPHf!ub>o-w@Atxxs}0e~dFlgJ*zyGFTq0MWGNCFmQb<<7iggdL&klW&DTBf7
z)K8D4?P^Gs?k$zrFGb!+G{;L@SrALHoxs=xzOz(6akp&AB2F|?^+b@%xjkfSW}&V}
zQh%x71uE3mp5_}7E6E+d<9y%dODf4Z{2P+T6{KpW{cw={FA?Fw&%Q*Ad!Fj0pV}Z1
z1wC<g`u>L=M)vcx3W=vFQpD6tkA^y^bfkx-o7o%GRzm|WsL6JRdv@-`vs&e6f98Mx
zuhqIj#Mt6*^P;L9%j(hq50H%4M9MdOr+g=L@Ntqdc!1CxFhqz6`+*!`aVvJS{IR8K
zB0*m{#n${eBNgEIR7Ai6PUp48@x^tQnkHlz%6+R)#6J|U{w4E6?fazs)h}5~RCT+l
zCPcy$+)1+0yl-5j9yGK4%O;?C%-2h3efM(kgBrEO02lT>J$ItgnoQL1q2NDm>!G_*
zYcW)rt=0o??dvO=b+R>6ev?44h{wP~)!C-p9u@ikoG!JnCY9lm=6|G{V-w&5PEve6
z^(Sbse5tt8E$vp+;Z+0>f)milbOZEjYJ?Qq4@i8j_i|qWqK~{alp9ZjC|*Z<@;CGT
z?$;g`*y$Je7c&*^Wt~7|{=s3%+H|6wc%I&7(<hHe_e0qJ(GWOXZQ=<<x=v)Qsdep#
z&uRtfoPSddw-1};9}Lq+z5??AiFl@eQ!nv#<L0vO@a3Lh5vo&?j6i+h^Y<2y*i}W&
z*%}ZOe${6z<`g6MX&xa|<*`K+O54LpsGXRl-?AwHjf>68d)5LjtRt?Bl#JiQ>*<ET
z4=B3hRTU7rlbf%|b2yvxBv?En2AMv9G>-^Bkg?PalGVl?k;p%yM}$7`{TB7i?wC)g
zK&MeL-TvusDVHosFY8*h%8>`dPx7zzbdwLd*Xmvd7sV>7)^%S5aR|uByTy5_o<%?x
zpV3rmvKJYL7F2W8g}tYd@mYK8v_JNz(R#f4im#k`l2rhI6G3RkA*{{JldAij`srD}
zxY~jnadT>Akp*YmJ|R3w{jlF5&Wp#9@|csNzJmOE=shz;odq49{z7N+<J(DBt(U8F
zOYT0uUfJ{T=f&(F^BHA16d@JE=w&3#C6l5&2Zyga<hfK>YvrnY_f^T{xg6DYaz;}@
zO6TA5%ZWF7q?aO&Pc0+LflGi@jm@tvhKgr{Ibjq0RN(-CQJ+e8Q^dJ5RSu(Igi)x}
zlOn}p|1f73i7~EVxnmd;chJyeW6rI)JNB4$pg=z`u@Q#s6Az^tR_!Ht_82=3<{t7x
zo`ih*s#$Qj#KSB!#K3d`akD%!I<DiEQ{4;uVF&y?9G1Zt9$zh5ui%z_s5Z}N&>JS?
zO)~;((!y=@G`JN3X35y)H`He`&E}@qI}DdsI(e>cWJ_eU&m1YABzFf3Ml@3d<ti9j
z#d5v0`NhSq`<+H=h^bOsJw=ezF6r?Qk7k?*R=qW~f~VbKV{z&;f1LPis^*9YmSNaX
zO;>5qkSTGAYU~2y5qivsjCu?_*U>)d_HcHN%??#`t8ugRHhKTNJe9)@*uCI`cr-?2
zc_ST8bs)ibeTSnv>SYeaty#{@pHF31Had_EMi2zgiW5*{WV@Y-E_$AElRTnKGk4II
z{i`pFh*X`8^P(E3=u0fp{5eQUk;9q8TaxD*AJ;IUUApnr9y!8va=QO&;sYJ;LteIH
zUpLhN%YOuN)%A&Pmf36FpHIAjpACLXg_RQ*rUNqTBnKiq{s9hqTFY(?*{?YQncj$(
zu-fOkz}23uq<!=mhoKZk^YEsbA{w}%3SN!D6stV-RCDKfr_q`sqa9F0$oN<K9;YOc
z+IdmBypDU%2q~l1y;^Go`lNV2^)TB*^HG9$x~;7oKT+O^Zxu1Im$|Bfelw;t8Zdev
z_<kg$`njL}F>^rgryatXJ97G$#9{NHusawcIeik~6A=Q|IE!u3w@!oQ-A(L9r;RR#
zQcIz?vk|jy;_)p6402}|{a1=%)csBipnUE6K1Jm_)swRb%D+I96x|P`;d#a3I*C>a
zgwH1L%ev-~y|Viml<&de!aq79o}co-2gg8Df8C%!K%U7&eb0#9@FDv-rLxs2l-N-;
zl{!>RG@ZJLp*#Y(hv*5|B|}^G&HN?AM7@j#C|INc&5Xti%!fLYF`>8ERm#6$l_$ni
z7HaNlWmJD-2(Sd-AE|tQQO{!K&YPtxO8Hfn^iGmtaGV~&<JabmPB8;~FKlo7IL2{v
zoERXenL3Kxp0N&u6~>(a{zh2)>6iZtPlF~H>$j44_8^^q05Xe7ip0v(Im)&4xMaGc
z5_=;)zx*39PITFuxe;3Q$%L_-IiU8kL)J<q<rJ~lSu0A)au&D1l_F~B7*S3`$UaV-
zsP<LCigrVE?gkw_>q#PHvie0DWFT2VDM42&krjn-Ub`sdJ^{U%h_L4c?+{+9(fO**
zhTq@l_U{dQ$7m_qPJ8i!MxWw2%#`Qc^k0PJI6$BJKa)%9)*Jpcr(F*-`?bNrB>L9%
zpg-xbo!T*aWvGp&bVa=n*GcZi{j{V@?U4k*e3^c5^|$6!(He6-&M%L>f}(9wcls^q
zxmo>|C?g_IWaZtyihx%MHml^SkAwOM3qYrU0m46kcFx;sW-HI4aK)mTJ8ggr6NO{v
z`~KA>z*$P!=QNjtK3vhjuqVoHyvh>%0%hk$iuQ+jZOT_0C!pnDgz*Q`Mfz2Q$RU*?
z1-IgXuorA>&Fx0iZ&QKw%g-GlhvjE`tSx{qO?3c?PipEM4|bxXG&c$zR^mRmh>N#e
zvwsR*?bfrwhGl%>Y##axTo_St`@I5m;1g$Se#8y@(>QtYgiBeDnx=lAzLK~ySgl#L
zY0vY!=3N(}SDe<i`ywZPE%E!g8Q^x6YGGYYw|wG-1v|nzfWc=YF2Y-n?5)fF#Xbvf
zNX{xrIxwk$OO1xbv__5&62DP<meZ0NC}1`zd7%gm-%>$rbU!UX_C&JHm5%{F10Kb6
z?GD0{yx}VwDP3AL_ZXdE%EbloOx-L}B>qFq%-KRMP1?B1;G)D8AUXS0=%R4ZcmxOA
zoMkq!eDBV9n3y6<gX(PPALqNKl`AWW?2Ez5O6RtBI1J6)+%>S=o%C<9>s#e?vZ(Wx
zCPUm$&pR%Agjl2QWx`VFus7_JY~rx+k@1?VFmzG8r6c2E&y?A%<yv{79E;Qpy7Li`
zHYTgPCE?+!)0I6fc_9?seIjg`$B1s^i#<?n(E=uf6wTh*5HtI=vXW7#@$2TNEGUx+
z8Wk`@aledL;ub7pWhL2L8p|*a+6w626Sg#`|5E;u88~*s^<G3gP`{x<NU`=rVLGQL
zE&#|D?B?~2Q<7;;vIR1<B<QPaOtFQsdg-7nM!Ohh_m6N<hqjEN*?;m~>tqJ|%2Z#y
z-DDjMDmZ*HCaTgbPxrmG(8)}%Xi#ypa;o;CkfV(z%P!25ZA&di!P2{1f&LRRc2-^<
zQNb{`rSU>}@>95>B8wx*Ud!f_Akyz=$@w3Z?v5{|1@bQO8b1b$)rIaU&Aq1X-vd%N
zbkEXT2NqQ^pO??YtRIFiQ+UK6GoEyk*&n7|JsXAZhE59?1#2Sa2^$!HIcl7my0?cQ
z-csZ#Puf2*vCy}RtoWwoQ1|V~fBL1nauM6zVHm=We%LPL?+NxX4c~fbKnPcAc6SGZ
z+!+z%qT8Ilf1ySSyGk0o@}!RElw^e2BtMu$?a$Tt*s=uvF)Dk{(n8l5w-hF&R{*Hl
z>NUJ6YT=M;Eh;eB7wyd5K&XYnf|0`X0|iZxG225Hep&pR?NAx|{(Ydivi?=L41a)|
zTF8Xzq<<ufzj<m>pa=Kw^69Y$8M38aJ-<ibwj5r{vb24K)ABr-(1fK=YS}g-)s^>c
z{4C4J-ZPe~kZ#v$9ke~GpLN7YmhDi|KKSCEcm3ud4|BcDW|-#3?&i^L0@DAaCdu)u
zqh~k#FK=E@5tQ?j54QhEY!{rOf^;um9%h>!wso0A{o6pTsah!k%aO75Pznr`$_5KS
z8e7U+AhyG18JA3}XNm1Ow-rBPIwArv1edxjw@6jFnx2+j3=_t`k0hiRTO9Lhuj&5)
zT#%oYCy0_f)ufcFXNV3IhZz{}@oxwInaQGPSjsPI;Y#`DpYj2eoZF!6NnIeD<?;Ch
zt3YE2r4Kj+>Nap`>Y))~vW#OY4vujPRe?wA9f!HB#DOA1i@#KeUSGCC9mzJ?K%N<I
z3cxQE5I;Ek56E;m9w5RaXfVO2IcBizy5v!+Yb3xcB-UOQ@%{GEn!Jzox(Q1g)RsaJ
zh7@V1-K{=e9zEzV7f@{nMeV)pfWG6p^JRi@Qe#(iPQ#sXMwjfedSl!(McPtMKVx1w
z{}k<)`ex!=#FrtTqyxapw&&h#))(F2pNw#sBH1}*w_%fqnp-TTLAdzM&_(%gexAh@
zoQB9dCH#KJ-9Px6XNk#fav%NyK(MCJPpRCeA&ISa`fXZH{;qRyt&K$g?yt9_L;m+#
zh)2cDN%D?rLbj>}?_dr^UZDDQC*mLLcZ%6ou3}Feh>zDg#7RV6Re<_2>xG~9HN1Z5
zffNcQ;yNm!*RgqixOa|g3^;2AlnNm~cVHq68@SB%VmHGwb1cuoa3UwE&iP~Wg^vR~
z5>_nPpyov@4{%+wNBGW$fJ2vt?s39{wkNuK5l;fbbAh`587^ypW%Gs|miJoo`-g&a
zB_3&PDbT0T#n2VRHsK}lRO~2W^EdVrY4VB!U-cUVe}SW#4h@@`8=5Pcg3VM^OYGNx
zah2}y1JS^{-J|QP7d!TQ7p=v~+d!{M|1P&s&J#sJROQcoaQ($1TdWl8pds&r!!Slh
z%B0|#49;{frx}H`P{x895REea$V~Yhk2FYLQEdU9!TUSpwo()iWT$rJWy?7Ge-VGZ
zm^`0oVWe4WQ*B~j@eqM(1j3#7W!!vtCH_G?gH-mlDlkO>o`LVfq(@RK*xmMTBUB&{
zR#c&)z*X|5`;CZO0TtVxTDa&WT<Z6d+mEzh;icM7)DzHwWbY5gNiC6QHMgiiiSZYw
zpPK*7gn?VGe*g+Tg-LP-V7n{$=B@zC2GOm0c>70S_{TCcmxups3Y0`<q5h`jsJc1S
z3qxtU8Jk7L6C=dj8Mydb8FuG|p7tD|jiQyNR)%eHg-nXk7Ct@%&tRAcw5OT2d2{x~
zUViJ9O`>N_NA%R5p+6yhQO653Q&nPE*BDSBVoG5e7``5asNvJz?odpyNwDuK$sL0~
zSK>4#SpbG7*^j^)J2N`T!9PY99E-c4Q83f+MPubrhB9Ioj{Y4A(c9&y=;~4RtYT+b
zo2i!f4D1c$9b3NkEFBVFwY(ew;|dwGfsYv}Zj8U$%{^l|$}v{%dw<ql?^55&)dhar
zoQDJbix+8u^cvc;>y6M*T%yRu{>YdQoQQlAX0L19sIZ&x1nGejD+<gg6lfh~ssm|D
zWgSF?>G8S;+l@&>1oPH*vq$?HPFhQ^K2d&BKG0yEEoemb8`YWn7s#XTBL#H}9ot1G
zcUx|@mGsTdnY#|qKg(>-^2aaQri8`D*pQA`7rCvfd%+F^Esg67U%eCBiH{q?-b44k
z_#^%?TDPf!;kDKWKX?L-rGEQ)b2IC01b?}is$lm+34ZY&qI)wdiDJqwtE!pt{3JqS
z>KN`8hO1;BhkTK$sB&1jyEx&)sH6KBPBDdQs~tH#@GFSgZPC#)YD12%;!88p|BJUb
z4~Me<`-hXH60Yi!2vaF4p(qMtDp#pgDj``$Nyswx!5C+fv{<qw#3YrnjeQ?OLe>#6
z_OXv;7L3`<%z6G?_wV^W_wgLZ{m=c+Gk?q+$NTd+&h!1A&-z|pgnkzPRM5woqURMZ
z#fR$5O)Kp>{NOJFbX;op_<Sg;-t|T+bUd4?*{rw`=X2VVpXN7E8rY8c9$(quP3ZUI
z$JgA7QAmdg-dA_}zn+d5k}~(YGJi^g?C{m)+hg~RK~3t};Jh=kTUNq^&}^UXS?}{j
zyw3BWy32DjYx}?8_1R{B2b;zpng8~5Y2>uQbj@Y;tJLS;hQ=db{t4|r^O7BU09dhU
zeD=lj*)RQ<N`&I)u1=HcJ6t4>x$jvsdP6dZIW;%^|LbQ@IwXF*qcKtPWJ20T2r~J<
z9kdDI|4U!&|JeOs2W?Ee(y^V9$}ji$9FY0>>g->VQ~xOaa#&b2CfB-LJx?Nm_wY}i
zxYWA+wR>AnL}p4#qvOeaJGVys=Vt;TTxT!MQb{-<bjpC6OP%ujlE2BB$8CgeFYsT3
zx}h7vp0j;9=Ks~K6UtwmrI57hu>y`Evq;~0%i0D6=N~~?W^$%Eo48r;H(1C)vTv^k
zCVx#N(<%H6`@5_8Gh6Hx2@YR63SH=R7-R1z)Vz=-8f?~`5>Q=+!l<Ol2213E_+V#t
ztY9P75#(a#HG4(8m1!;^=ta#pZK;0@2r3J@|A3ilt1s)S<(=~!uE>+w!HXZ=RJF}b
zmDfUqF{E6mmGW=eUl?PDfxz#E^Y2yF8%ZQx8=NYft0So0c{_7tTY4*{C-v(qX0)xH
zAN-PH2J37Y_LA{uQ~6)t7w$56AEky%+Lg(7VYx1vxD_lJT<?_27Xns~2cQKYqb|G?
zTtt1dutL+6LaeaAwWcDpdyKz@g&SG<2G*Q6Za;|N%2{TJ_-{!a6a|QRO1`QYoHtrm
zWiw{JfXO1P@>D~`ARN4GdtgJvuowAjw};<cWM$n<)mpC9(xAszJ2(T6{!j_R>nIId
zyJk&Jj<S0#b4L0R#~O%YcC8=9j)~$H^%y>5qQmv7rHGIoy>|;2`@VO%$^T+<S`n+)
zg|iA>`h^Flb!q0v+t?nQSOA;3h*F>^rKzMY*sjAfzsUN#fBE%Bz`M}^#kVjIs$;dq
z*)36zYI=Ni_-)8mll3TwS@KP2kn(+Pp+@;X)5PE`H)RayT17R&tDb9|)nuK<HjiZf
zaGL%3<sRvgPdMhGCAc^wha2#DANS*Xcj0{1lqd6GUQ0!u%kl2qRhXq+NTtx0UuJq?
zb^yM|AxFbEU`2>2U%GBJ+b;ySQ!T$BjzLz&1-*%^7v0C!RV1iEZzDrLh}$gW_lCE~
z)w-y-0c+DA?3#n!Mjf2yEZ6&1QImbg;k&Y>8cKh#4zk+8pY*$XlS}(-zKEA2d>^m6
zfKa<i+8%-RhZMhoc7m1CkQe_C`#8>O2*$X_SDpPE4s~AsH7Wvf#YuSyK#GQKmSWz-
z<+ZTpV2LYvcGpwF{)^#m^WLcJXjj>_Gnr%M#8`z)&RxkI4pD9qHKCY<#)n!wmM;{H
zHum9?01e#0!m-lIZbJq6GFs<lvDv+)WrhYp(t@SIwpQ-cG3}S;7Tr?8B}<m;iT`r1
zGv4hi-hRx=Syvrhm_&_tF?C6eg)^^#g%1Dk+T8K8{w4pzvw#Ec?=4T{n(gr$Q3k0h
zgYjpcda2mV6`tU3r9Qle#_R+nreOI*^nlvfIy+UMpj&vGoD|NxnZJD7Igkw79%nE4
z4c}5(!@;{aUBRz+WKtCvZiEc(YYZc0t@FZH-OPoxs{rfli&odtRrG56qt4y{GP>i}
zT2+Ll(pBz^HPt66V)<+Vqn}K@nr$qZp^=ika(<SUB<r2JrgXa*H_7^C>IB|T$rWa-
z+O_0%F0RMc{oHTavL~NL$*A07-`UI8<{8kT`F4`-BdTFro`K87bHDCn#Qa)ghMHQ3
zC8sdRXoWnTyd}fc$00@+TI1p=uU%*~<N`qT*u`JZ>5vcZ5yg^v!OZ5V(p5M<qx89D
zYCl^!SQnmoU_X>>yQYVv3#x}9U3*IrvV%cmDy`*0?&J0leh#pQ$R+>x+2c>3r?HyG
zFD7yRp$j~X6t^>A-l!M!G;=bq8|PAUzsn}nh<)m2aC3g>C4JC+{r^hBEkdvh2cQsb
z_w-S+@r+R#34+IT$765ex@K$lg?C5tGsP5y#<Kyl*1-jj^4s$Z@x6~il>W1Lj?a(U
zA*|0fm}Vz*kYO+}($r>~T7gPuzriz6FX<mSmpqPBy^GNssZNfLMoHCE#!$ZqQeUzz
zmn+sIWcqjw;xL(#zViXpL+C1_-n;H=-CC(CYjz)fKPT6F&F(j!5HS0VEopJ@_k=k`
zsoq6ouV@r5Wd?JYn7Z5iA1C5YIAr)-R@T1PEC@VXqc!l|WzQs87fb(DHMQCN79XI+
z?F{oeGVN+9Ic|DdCWp#9iWEfpd4*bI@4Bc3IXEq?C2m9^h$1Um`Eo@~zYD5lR`b@%
zHKt!n88x;;sv21BX-o9ki`{ozk)(F!%<s@zUEPDSp6`6^^exXtdoEIrVw!(vcxR;6
z2Re_=ZbtVk*b4I3gYZi-rr((5o$JI{iakUrMz|QVPf)!my#(=L6#8t9ePL-YIT^&0
zC=UTH^ouc#SZ_7ATBw8btL_gWS=}v=vLIWP80aiSOwI>Zci%)>w?G-Cv2wX=qVKrP
z^`}e2D8tyl_(izxQ>88?=e(^Pl1<CQj8koq8~v+pnpawIl@)Muy8k%~>wpI8{4n!&
zRYGOG?X(KBo;=vFJ<u62xBJ`O;q0b?q_}V5fkt9W_aBDT;7dj%?Zjdy|5h}od4Gw;
zOFZ0L?{zqA?=m*OUZxc_ACj!QZ{_!UX-~6hsrkI=K+BeH(B7fet?Xuv?D%Yzwekf{
zt!f!JD0V3+Z((!%+!IQ>|4%keNy<kmrqn<(n@G9@?~x8UG7w4e;U6<`TIshDE5Fpd
zk?ZM`T2|N|e8*IznWYi1^Qf$uovgq$|5o2F^)I8x9R{;Xhz@ufe#K`P?ztNt9BRf|
z<pnn9ISXHn*oofz{(7BpTC5z`3WLX{uVAbftJcR(^$G-AG7RBkm9wisO2bQgyak4+
z^TzAxNx$ANY`tsuCsw4PVp5#A6@HwV+Ljmko@ly!P1(H4q<vQ0+MfSTE-3btR%t4;
z{G;qS<Rx5B@%Jfdg2+nb(+Y~Qf%fTRfxSnm-=;sd$74w}qOIxdmKDA-`r8K_qqSzU
z^2pjWMsq3xwMI}HwNQSvQTL-Gli5v<Xa{*ZvwxHg-R`R4(8np===G?|)I@`=f40`-
zw6?mC3n~&VK=Z6=Nxx3o_>;J?#Tw?3`!Tj*$PWCYg@MA#&nqPw86O5#)~Y+v_Ct{V
z|1-!N;WkVA>`9~UDns<E(E)GkHJo#<+vn|Ob`LybilHtz#8HbFhH%q6S*3^%n-&6T
z_~N|-hDO}IYr)Sd51{tO9lpspwVzt{5^g~Ga*JU#Ehm;4fXg;2Ydv73mM0`G>YC*R
z?=ksawNG1~fBIdvl5guVH+VUCTsyQo*5R#v$L;EyJsGEbVoj^J#e6{<H&tV1lP9H&
zF6XjYteEabD-h>fqC@Z+8uvg!gE50St)a!96{z%z&FWJ=4`~JTPknDX7`_>f&H!!E
z`EfBM7dNlB<D}#_kB|hM--U5xNnPH)B8F^lkbdz1tup&K!x9_AD7aI2gr%TCs3Mva
z?#UMtN08L*mBYBOt9GK>9CBE(J0RQ|{GFtVtc0GRB*Xp*ww#r1D0IkO&vZNkWoLRv
z*GSIp6}VS3qEO`sgU?F+9fhJclUAAXY6g0!O%f;+;yRX$5q)NXq*;Rt6(9?H$)NIA
zH!{ylNdxm641k*8G(3yAIO_YcG@a|yl~~_dYHw@|bv*iw-KRm|2}!exq1Wpo*;5nj
zlN_k_lo0&)ipq%e5LLc*ep<O!$d7d~h%6-6tW><t$aJ(Ob5H1Ym`@|}2=QX}DfJR|
z=LEOHd1OXdmu*qDyE(K?V9j@<nL5dZwH%w?ru<XWC{qUgwN6VEr|Z*+O_rU`1k?j_
z1D4T;t%?)!y%ZCF2TraYE%7a>`2kfGo{44Zd?&oLFto0E(+;^uBS~hzNlC1Nvlq1`
z#~L$S`SX`Z{U2c|>Tk|1q7i_ZVQ-MZYS8}W8VRF!w9xYvvLjBVKS_i_xu_hz)%ry6
zREwNst<%OPc7Ehttc41?>+D5IrQ}tg!VRDqdqK85JZ?cpa?K;|VxO~~71o)<t$vwV
zkhgkAbhL6KBdTL^e7hA$ac<|_zpIqB^!)Ug0LFX(Jo;4NtRp#|Aky$!vEh*B0NzD*
zN=@8m++CmNg%(1942jo*GL)nRqx2>w4=AuzBf*8hV<^@(^aaF~Wpn+CLKwHG#}rKG
zXXl3+bFrBF<j8K_ywX%J{pKg2ytlxbMeIgDp7rN2t1b~Y>iAG>TM1;t2ySFC<miF;
z&#w8OL+)RFVwpR+3-fzOuomAKwg<}R5K3EnDQ>$QQP(-a&y536-<v3Ic0R9YW~@*w
zw2ir`S5}~(*=7+KaIVXgOJDg;!xNo;)T?ra7K|%Tke;BH3s%t>cK?fYmA|3}_K%fB
zT^DTO!}C0?=NDIh=>#A@b7JzLvxE3|vkSN<s7yrd@ct5aV7%j>bsmLMoW%h=>t2XN
zXvsguYfUmV)xwY1?;{<1$?VEY2KhAekt}mVBPbv8%Mb>qq%*SqVAI8h)9c+|VZ+v=
z8O8mf<u?*uFxAg0nB0{I3$yZ38sBWKGuNMWa-s>m7nZ3r-<izYscez2I)_VO*inH3
zb!vkDhr2daB=p+r&JeACSYP1ZMy*5p+QNL`$gKKJ!MXJ}YDZV7LqlR+t=xC`nbtUW
zP|QDv`OpHcGbafB(pyXlF1tC5#G2z7&SZYlTs&bUjx!-z_P;*y!9ow`<7fiWJD(^$
z7M1C~Q|(w7E!t25`cFwg3-l{=JZR5NOQgOMvF#4gtvfo$?Ppo}?V!-yuwF;~d_8UY
ze3-xK{%)Z`O`b59ikTe?2yX`$^~F3-y^)(wy=_mqWt{=DqVCrbSQN(2hS-z7gy4Gp
zU>TzLwh^*!PqTd>b#FW)cXTWW-J7{KW-y#$JRJObW7M-&WM3YB@@WZ8*t=<yy#0Tx
zxT(?qqt5!jB>t=728Z75Cp2hGJbU@<^|STA<Zk<zoWHw6tJzAb>h3b%Tf2X6%HEVu
z+n#NEV1D5Fk%I!7uFaKA-=C7mIrs|pbSuFe*>N+W%^>we;-hzmiCjaXpUYf#*CXe-
zOv48tFDzl_%-pB|-ogBU+3R|)c*oi87jC}DI=%gV`<({+iZdrNOFSov(XE_Y?X?e0
zaXuq9X}cTkYh-NVC{c~|3Zrs7^zl2{ov{a2Pfx*j2e#;CKHiKg#!7uXrO<eK$`74V
zf*ERHetD!>veiaf^GIjcaRrTY>f`EmrYRSiCeiU<e(ipR{y5oVcnfenwW<d~XLJT$
zMD|){ZXa(yv21y)Z7cFjPPysKmhTIi)$kt{$>J*4HO^^`eA|R`3f=)?^n&t2!n%I_
zRfk5(wQ4l>HhJ%P&)I=U-K}uN;44-=;RTNstV4nwc~AbGm7Z+*{klHx-garzag%pu
zj$vmH%|@A+^+3DVYK^bF`XH_J4r*UPrQ~EoJ3Ct!be5^@nUm6QhvIX)e=QyyI=Jle
zid424mk57F6<1IZrmr!Rhm^zgjo>#-(F;;x1B_qcft|7wof6;r#7`@3O1r+vOX{iQ
z$|2(y$pBMN{B$}G%Imw3t{PR?#=54)&kAi^Ot{mzXM?w5T21WL;EDMi!kUZkVE4v~
zb}FlXn)W^{u@t&S8GrbUt)UX-<7E0I%@9T0`XuhlMNGj)?0CMp$<Vd&k@U{xh59cS
zdxJ#f*%KAEh0nL^$GtFAvls7UC`sxv6l8UQFsaU9R~;kP&%5_dLN3b#KT_W=Gn57^
zCf$l26She1pgh-x>Ndw>OLr~aq%~lE3i}?;@Wn;v39*a!GGRrX!+kvzD{p0_Y)q@w
zh3VhiUGHGyDfhyIolBoRdZj6J9}2A5TXkc3koIah5hwhoZ>NH<-d)&u`0O=OM{rfu
zW>Lh*4uJ=1hdBzlFSOgnyV+U($LH-%6U9+#jZ@^u;SZ>&$vpEOl$#<+D!=|)SC&Cd
zbNwE>BO?s~hqkjC6+6XGXdmt&JMPShrdO=sId_kn=4YlEIOi{8n6bO$Wh2v6=ZRLU
zl!Gc})1d8%BBEpb^4jIJyh0^0*4+5^YD0&od!Q>n^_tEXaT+6on=DI|6_3I4ztl}s
zFz5f4t#MhotWjZIm~eWZ?)t+?F4+(o9C-L)_r)VEnwCR<x*6Iu0G!>aY_M<6$Vh>6
zZ|}p<%nSUdWQq5(H@zZTreEDL8rmfrf0O^BUsb(+Y#UX^*p>%pcLRECn~QB<L9slD
zD_?)ai16gp%7DfN9OqC=jLkmvnTIM@E(P8q-h37H>qm{w8v&}PkqXn?6DA910rfMK
z^QY=7b*#6(N%dRsbG-mb<ir7pK4#5cFvR;<bBt!-KwM{I(5aW|=e4j$ph4!*KZN5(
zbOJZ7J8$Ri^-VdYKud#|OT(gb+ynRo_eA!c*cq50JKc)3CzCPGC!2vcCJNp;8EGy-
z-bUxRGB!5W>*Pt&Py(QP!fx4ba)F)<3_y9kS*X&@d%`l`fSx%7V5{F~mdwY{kFx71
zy&u`7WfU%Rlz4<e)P-gzg-eGao70a+9MW$%Nq_1a_Jh99Qx5P=xo3PkeV~J2X%*@V
z@SH6%{6r^<!3(Sk{d02Cjb08tQRdJ66h%eE$f4IrSw2V?EhW$w{sB91V56A;5I)h&
zz`Wl?NfGn4z$MmjN&GEzSv<6csXVzrBLGhFV8A>=Rz%b^B79nypuhx)K(y$7r(@f_
z14bDH7H*&%h?Z)I0i~wF)amoElRex%$=0yCz4vNE3zz*l-GDwD7zIqBa9_L>UGRVc
zG_ttc&0X+9So11U9uHv{2-~?#fs0?<paABo^Fy0Sv<Wg0%u$}!7wkCU1MPskR`If3
z;`7Av-2$NB#+${{c%D;g4!ZXcn{p2@kg+Lqa*rcdaN8$-iM7TTtlhNqH&9(d#(}Kg
z(8I|EIcW{b>8%I+m}~)_+!zNS#Wxc9{>toj;3F#b^m#3g0@CZ;Ptc)aaOizI-$b(M
zoD1&Q3F(Q>N$mMl45#oz*M28Y2-w>?u<z;#K7d-mf9%cNZb}^6LJ>JjD;*7$s>lbn
zs${@ljsMP_qBK)q{I0!|3g}Tj5V&1K;4|+WQ9rO+lOy(V-^a$6DxI4&^_~ud9+s`=
zofa$HZ+Lti98+`i0(yw8Ok)35jgxV&1BcPj<uy+W$Np&O=kws1bo`~ai8pXCftRjb
z=xu)|?d7!ZI8(puUvQlMC6Hhh1h%>T$hbUC6k`@_61`C--@G2Q8!q1y`nR5T<Xq0(
zm0+`cNaEI(ymy_i3W3Vh;gB@>tJqBuNs^~PKZQvbH08jk8&i~%4b*8%8u#5#-%Nxq
zO7ls{!t@Bgq<hLek6z%7ZRG4B=@2)2>xvl%^Jy^;ILtQv%jNxlyFoH-M0t^*C-|tx
z&X)qj$77IMDzK+t%tOYEokjQlz0CUb?#0*MjGp_6NA`<PI03iPrR1RN;ge0>a*Q%2
zcw^qsJp9Hg^5bm?Ja^5Y{+--XxX<kq5vIhUZMRZ1V#MafSz#X=pJEfXoSohouzqkC
zm(mjnNUn3;h7f@1Y~EZzL((g2Q%z~wnDJ?azrkWUO0Ov@xVX|dRtRIh0yO_X=2_51
z2Jtcj7k4`(&u^1r*X@`_|L@v&5ER7LE3D~M059FE2iG2Tp$MHQUmBlfcFh~&rq>cp
zeZ+WyZcRXXClM%A?6N+yW$rsE#PQ;v-Xf3Rs=N~(Gq`ogvUd=#ExVVfXIn+8=clv-
z1+PM*J{7Oy_;`+>atrmi)m0F#7vw+i`P-N90y*bdiV5iVr5Nyl;8ySTsW&H$1#`&b
zQ2Aq<YVs34<;$}6tQf%O&NU|bF<;+T91Pp;f!L08G>^8sufCnVbvWgG8OfNB2nX%$
zdH73E!94Mv1{m;{f$7lhDd>A^&Sg#M6r`S#;?2*!4u`>&R=sG$G`fN}`~9vT*vw3@
zpx2<>PBAC;E2K+O+5)XVgeRlx_a5CZP%=2mr#w1~ddVSUxcecQQ=kG&o_Iq9A7EIF
zFm-;&L+pC*6f8Gx#a-3dBu64NKw`)0F|FqVht`Py<U{>CAQ|@_ASNDfajOxyyc)X5
zj0XldPPpt}Rkx}-`hEbI+eKi?xH*f1TxGtghbrZ}KY<Pn2(W9W+*}GBeC=O>(4Qc^
zHKMI5=)Tst%FMY8hdfB9tmaZNA<X9hi=6`dn^8gn%|AhEU!aSxHVa3i!8gC5JJS5h
z=PUSW5Q5i$>SG35hciN!EvZ{LmoYn8{~RuadpB%~J|&q;gDn-Q?GLGFuWL;J_Qyzi
zLlU>Sz`{#b<5i5>;!)Z28Dd^UO&z9^)&*F8-1izl-nhwN4T=1A-VjU-p+PyfZ%C$7
z1k==|laF;SrkhAgtM|D9jaTdh@AjKE-O&G58FrvTh`0d~#2~$cpn`|?KUPc=!mik`
z#J`K#<Fcd=TlyGvH?h!--~`Bi$Bq?0DP!)Kw`1I1p}ZedzSkJhhLgO@A*!yNGv=1H
z?ZnaJMjcS7I-NZ#b&yqnD8+GSU$jqH#ti|O^?ok9DEMmqeZJA908wcWPKfbv?1<gy
zer!K3Lh|v;Wt{vXu`zBGsvTZoz7>+&h2(EkXcBpIh7H(?c?3a-LxE$2;5Uvf1pV(!
zTtCFtOGQg>r1EqiNaDev89M1X<M+S4hHSlr`mW7`vcpo7vy{%4g8@Xohw`ZI4Z^vf
z7od9(nVQ8sBvE@*>nR65CBkoK;<_wW&;DTpzayy^q(}6XJD_~zKacC%BupetX5&x4
z4~Yt4##aNhCOtWS22n*6u~$Tu5MnGh6wTl;h+O^$r1<vDzqIokJ)@dc-o3>_4VhgC
z-KikD7YQq0G)##xppTD5BX_wZM?qHQIktD(OO~oHLGU#$XU#me=iVK25jg?cy!Iik
z{*9cY@>3tN28T86K=Y+#@W?#2s0QrQgA|oaEgsx@K7fJN7ck#G+(`3IZU_BNf!aSy
z;=uiH<XpQTuPz=yjfwgu{qhgnN-N_Hw^ExyR0oG@cjSx`BLyf65+dE@o=1r0LvmjZ
z@jB@ogJ=kCEUl+dXd5wnF6!GXux?Mj@S`?<WQm?AOUV}OnIcB=wa-&(<^G<g2EQ+q
zotPM(9bQJde-2&<Ket%+_`)hBHxtThwc8xm8Qp12iP$RryoutO#WAyl_SFu&D;9|Y
zn?FLrTaZkE2<O+GdN=HlPeFW&OvF~hH%+Xc>u9>w$0HEwIhUEV2t?Jefy1dUyR(Ka
zGap3qQAq-dYH|A1=DjIT`iMlCm!42ucV;KI!mVuE8uBPhi$yught5lZ+~+Ge7z;!K
z&T7qr9x9B_+mFXW(N&v7r6SSm4Hph)a1B?o5?rItKz6Ohfpj<<au@}Y20d*L_d|-y
zlrrlUmDf_6z|RMvua2W+b0HuaG8lHc@C_hqMGEe%AphzEwa-F!M-d;cF(-cVTB&o#
z7^6fpRgfGwOHnKZhr%wEwUqqc%LNqIU08%c@QEjTuMbG@g<!u1-{72Tx4=F@at}ov
zOHhpwg8<{<AR6Ix;qUwEeBupfXx10H!1e6f*(rOUjU*$4>L-9_jJ)#h6FmqH%kENw
z-#H)XcS{mG_bzyeub%)7eBHOoq#;0vI<aFIdDRf)wmEt>YK1ZclM}|p7`wv-!S4j<
z7c0!@8<5A?;BTgdt4NxU-u@E0C}h5oh7PiUunkMIa8)X+*hqeOfBp}lp&^U%<K+?P
z@_m@18JKfj#mpUbKjU#K5I}*ph~zu!Bp>iEZoG)20mVU!GWB2OoD?^5t`eV9ATiH%
z*%>hnn2GxvP3VmS=ICF2PX++LQK0O~4KU_(&LvgQb_(pDX|jKFtSfB=rVZs@aNmZY
zY?{#O`3PZGIEZ$;vqGb9k6lnj1a#R9;;^i10YnC-pcOLm!V-35v1<pk0s>S}U=HEv
ztfo03Z#nFr^*GTAaZm{ojR^(ftf8UtkzoZ0IxhePjXbC(nRR;O4+6D^XBOmFMLQz}
zToQCNp-a$H1%)kdRIPyFKl#ZGZPcLHDFQ9WpzJq%L>dnIvdnyS5Q>n*NvMYnUUmoX
z4Y`0eO26hdnb$6b1c;(CpzjwT_u5QI-VAJwRuSt;HgH8vFx;4)ZP42H`^zv}7C_K}
z%r6H+(3J*GkQ2sh{ln#JwfCU)LQ!kOx~$a;BN=!B6zb2R{}2VX8g*uej^gmndzi$R
zi~K<jiS#{#-Qf9eQv$;H5Z(FljEhHcseX~xUyH=1b@OUKvj93Ba1Uy8Ix0@b`B296
zcW_6P6>O0|F~@qVpg|jG>46LRx8iZB>#3i|!Bp&Q_L$kX!YPo|yza6DjTP5|6f<y$
zR~--1*L_m9b4TRFw|)>qgvS38geeQD-)YEd$V?tQQv<yrw{ZZy+sJmC1c~g+GJ6<R
z9WLD*J0};fl6qXz;d12t6_~965&WJ02#oQ6F;onE?y0x?q>Pp@oX#PP?<~0KG|_2!
zgO|dDzFvhk3Sp>1<q3efv5-EU{I7r>G~99p<-B08YcH^q22rJXl~<h*mmIGEm#Qd?
zAa<qfbB2>98+kO$as-GHA)T5K8G?%p*20OCt8oxVATf-HDxE&3c7GMNUjX?+z4Z`J
zX=JgpXWTV%sZjaIC<;x}slO>mTJ_8K7$rvV^?TlXj{;l}{ye-T)Mk`4ZvMq9O?%YS
zoHs?>Q$*-m`boSI4OBpWgbI;59(dQ1HDptMV!(+G?~v_!FT48`YG-uMxt+gf>JC!R
zx5bUPet4Z>a{@uUoCH98?fP}2g-&_TOsF`7zvq^HL6j=RGn|CDy2+0Lo8W50w>jbK
z?^cmBeY-4EE}A9F4Bw-WvUWi<Va{VZ4*J>NUG%_YB7a*(`SEA9KuE-CMh?{p(W`Lz
zhX@@2rd;j@mSl|#Q3izUIhT9sXQq(Bpwq?*3qq|Zg)rsMHrWM+rINDbpkX7Z6!A8F
z-OymVNfpubVHnuMr5xTs`|Ac`1qe$}*m!4AP+XcD;WF6z8Yrd$tp>Dc9Dq&UGEwpo
zfFmk1iymMbgiFmlL~1kn<3mGIsK!Cav)S7bJqmq4xtbu=5Ue9UycE5m$FWdHAcs{{
z|9gNYf40j9+T}P^;PjD@nc}C%$vy_%=vDet^W)YOsM?}hswNg>GlI)I1*K*P@NWch
zf8z;j2U>1NCIE?lLwwPG2|?xRIad&VX7fy6h^?NrsDW|h2kya%#1c@!0H=2gbng^9
zs`Bmks}<l-;(&6<m+GUT_Q_jV$Ty(vZ@_1Rx1tf@Rzl_K-7H{k#5^XPD>=w-9@Kwt
z%r+1MzXZIznI3KA7JF$FILuq@+j5*TL);;N`Y*T|oqrGQ97aDMIB2N3BY#7PiDxXy
z2oY4Vxu_n-C&`3YZvQ0O|D}8$#D8rddDORuPobY9TJJfJ4#v;JC5M&n4WzQ#cmA+Q
zeT7i%9~xdZXC7gfy*qwogC(8m%#}@D_izu-sZ-_LcmX+E!jQ?$(}+1pD_I01O|0M|
zVy@0!nc4bE5VfIOL@K9e(pgLru;>-~!idAV`!yS)|B`t{Pf`!!>q{-;x~Zz??(IzY
zav+_VH)jG}%E3XU-AdcEjnT470^?mGoAuBPOEc;OQ{Efu3=ZB=l57!wG5CkACf~Xi
zlwHAT#UG4s<Myw}Wm<vw8s9A^)|;;LCRt}TIq05;v)=utE-9e3Hq@Qua}$@TjMYcp
zBb85EGowAe20j()r#)Cqa<{A{O*XyQA|e-sxurc@dkTqa_eybhYicPOZ@~%f1+n@i
zC(lp98Z8rEa6KekdlJ!~LM8hpFKlJMCx-I0%Aw5m6W%GFmlJv*wI5FApMst)0&4(e
zZG}+7`R+aRZ88DS;3CVVhqP3a<IFxm5rdG0hJ8-^+k-9^IQO5IOeSA<yl5)7M()_-
zr6SF}St~YYbXvh7_`%kEdp-FR3Q`kHm}XQv@YPX)>_hERbHz9A`*ui*;Ap#%zv-tC
za@<}vsy?aoV8i65{U-d^@!M_$=9}gPDNkvQ94Nm15Dvz84(iPxH9o;VC?Nw7Ua?{l
zx*ixO8bcL+vCrQsP4KvboVyE+)P^vJGZoxq+qZu!*d9ZcggIhu^c*y*CH*trZL%MG
zvs#LuJsc^CW9@pCNI|_QfmNJmXL;TNjO1`AwM}y<*M@s7o{cs8-~taJMn23cFT!eX
zg!RdyF^$&uqp_r~{;EOO2DUymAZ#|-7oGKIe`9M)(D8t~)LOHFr?vi?C9xJSC?ooJ
zh;L`bzUe<O(7d2{#13Q@WK5Uz*4m}1Quf$|Qtb3p=#3@J?>|Ews?R0m`Wqv{W5kQy
zqnd}$+<1-KuQ}!@)BEw~pIw~J#h458L)lX520yC4N_SXlyhFe^jfCW(1eeQSXZA0h
zGn@-JK^^avo)3{uc$zQM=xN?uVSDxdCiABmJMIKy<H-J{OG)iEO~i*86W<%^u1#<5
z9en!XIKt+^FPSg*Z^mw(-GJ45;DB#1x)*;`_1Lxw@q<K*?7FWBb|8=h;jew_qa^mu
z5#z2w<P#Dl)H{vDt#(@*1|@KCjgA@l5!6Q6bA7AAj)dO<yR|9zwzhHGwkYI1%3ota
zFO6^qPjxYhHf>T+_-{jd`mX<_P5poDzF}w|Gez5(1~#nN8034PXyy}<3UBv_&0T+e
z`_Qr1pM3Y4_FhR-xKR>Ej6`!*24D^DueS6^f7~B)c<XIlDVQKA>oQvWvz0MDBtT;n
zvox?oR`|P$2iub1=f9k1iys&7VU>muz|y?}!KyxncuSO%o12T+Gv-->>G|r8L+SO7
zf<dgS;-9a@>fgi9H0dSoeevx0)i+n0>W_?GA>5geRzKqok{8{BY}v3}rX_+iR_X@9
zNgp6Jtvu)utJS>$ACu1jvg-q({B=sbpnO#2TO?n0)QO`orct%k59WvR(CQ*&@<S~3
z6Ox+2Gwcm}6DSC1sJDS`c8QEwvUBi<+MP9(s2|v4-gxUbHK)16TgX~~Nwk-w8=iqq
zK`-2&b^(u+JvP8c^&n}yAYOR?<q(vyKCHP$DQx0+@|j(8I3MYs1C3v`Mij3QWKPKN
z-+A~e-<@SqD7pBt%MiWBkr*M;zdA&l22Bdqp`lU)cNbTgMjSc@MP~KXpM!uUB2u1?
zeoxHX14>2E*O7E0GKCH#fDzx`iJ%hUb;`7J8MB#BOab|gH~Bz6(4;b0IYId&0wP?_
z&+<QC1`CEOH*Oe<OFW7QO$*0X;-JsG6vNp#W6D9u>|YWFChF+o<&1GlN$6h?QotNi
z6{rYOf(zEEvm}>Q<ZE7&ZUgZ+!h#esT*k}-9APY~3xbi4AfJKU7nklY5#8fxm>KQi
z$|<Jt1`W849D$aIo*TpoSUIW%!LuU^Q`S)<CIYSX6wj`~H(ancZT>lt2XBBDh$vZl
z?F^FfoG4NVA74k%i9T_o8w9Ub+{g$uLSf*%D<uJvDPm+{B}L*~E4=7uy)smWfq0~+
zYn0k3j`FHJJ)J{k)Q%A)Im*l0=%0wkln2mzQ7nrYk4*<>D@1%!gfW--9lIZ#4HRwQ
zm*524-}Mpz(hwA63Yc32Y9e6%IYg`PK?F$CwnD~3`je1}x^Rq|zD(lnTpof$+~3C8
zkSu*&5(hAds<iz)sKLD%DXg1-*@|ll^n;~O5~V8gXcj_oft4_kC~v>thWqq16uws!
z^`W|380J`hvIaA1A-p7bETU7E;sJl@pTYveb^GNf#Ur3$Lv1+l$2bqcHzcsl;gWBe
zf?H_&|7xfS2Xe7{0WV54V8#U!ptnZQObCfq%OHw>*w521L(OOpZAeUa+0?^EAhw+2
zHiit4p=G>Sk2Iw0_~DKk_cC}&P$<BldDY%w3^Oi%@4e>YF=t94k!M%cM+W$c6Laul
z<i69={rd?r|5)m8PSqvmWrHHFrz<5N;UF#y$iB~^ka-H5R>^p&#QpvlLI7&b;6VQo
zyD3CF&2WJNP_A<d*4+i&8|W)C&quqKHv6F4h-p<LMYUbkg1}^~2H0JH0|=CUX)if(
z=HS-0r3=nj=}k<;=1bIZz4}^OI-3jtviHRmgz1Cg%QkQ8)jl|C^Q?afRnacmNEVjP
z<5DW!1doMd6y`@_mE_L(&*aTXfV}_vnx(iKgYab-E9*>A9<7uo2@nlIa2>YZG-S>T
zhC+X08#DRU2tWCO2J<%Ke)t<;jY3-(Rd9TVgG8h)tqzi%6oK>(T|{TMNZ3if3W578
zopE!xb(Fz0@mwuz$Hv%6I|{COy&UBVT*6AQC6ILyHVw$K`M0MhjD_N%($-wu8>&ey
z4E{HudDsk;N`PL*30eo98VH;b5Svn^%1Hp`0CF19HSPydeu5!53@mg;jai6H#kn+4
zq|bN{=b{!V+reY*PlfgGCt{I%TrH~Aomu8I(@<&h`6*+SDi;=$=wnJR2KpeabRrM)
zDWTS@VMr1zb`QG@weKEh+OBMLXr<da=ne_9iK<$IBLZU@o_t=#MHH2F0&K}#XPr6u
z;5DC1snXyeA@M&F&<Tk6m)D%nfU)Fl=)wJAm5*Zt0u6D2me-L0KryvX;+CF5b7MR4
zQcW(IDJI~#2aM)3j&##_Y$wu2UhP1w<O%f~M!KMKCF(K!F9`kz^b*I=VLoVX=m3tw
z`ohyz6g0r_;XY0zykHOe{+$U5K&>CrAC`R!sa2Y@<>!opl(!X;<`5U@5ob0r^cGR0
zmmtuP+i%>#ZWt=XBJIQlZl?snC|IG`oXQ7&5wx#dn@3ap<^z3VGl_l_x-KuJKt%Wj
zsSgrf64+D*gQm`-zIcQw*_sdSOXA75Y)02fJ_kkU@m-t&<UL8RxWk+T;shg8aETUq
z=DhF-CX~~}X`$fwiumv_L*_ogFXQvwrf51La}C;F3==Ya&@}BTGt7Ap3OeTE8(N+A
z80(TqJwQ}3?!)<otD@XkY2|Pa3qD^s+Jy`rnQ@1Xmg|F-4&Ia~k$UI3TYPQi2}t<?
z{>@+_6^;?v#Ep<js>x?oVQQeibo6sZZm#A%@%b>y#%z73UM9M^N;$@v{W-qwhpF58
zc)y(>f(nmVRrD6rZxr@rWxt)#T8|SH1wVLZ4C%Umq%aJ7Hw>>7`C@HJY8VX`Jl-T~
ztHuf9_5lWhZ~_Y^n>d;Q=s*@EdPKF$E&=v%2q})Xe_d9B8yY4Z&Fe|U*Zfs%?E)=+
zdKAaOK13`ml|r$Vt>A!2YgIOryYn>1oR$s#F<wSaFr2ag@3~4$!=rmJ5pFwXDUMzM
zq!DJAh5Q7l$nAM2g2)tyPG!Bv_8vaWmlsS!7Dd{v9Bdoz13RimaUW+^G|t2eMuDAE
zU@?{=R~VjFf|rvusx%ic%R%%)J(Ii%7*OX)azh4yeeEj(cpI!62q*e8s4ii1vJ4CC
z3S~QIN&)FhbjJy{b81A=LJI~vn==ZW{>3-r3e6JWJo1iTScP0IZZ|YkM$s5#a=TW~
zJf$^L;jWyndSu&n%5*bVL`#K@alZ(RpCC&)5?+Y@QWizm9mZqIPP=a!Nj%d)!nPPM
zn{tAr-%=6r7~>gPhFTfZI>oKF-Iz|aENkV4J%g`mw`Q`tq$N1Zq76RjqFsX5<#9an
zJ2eT+GJg02faXFj1I9sf^2KO(#{<IZo9-l`MgrjVd`OB6a0%l+8#a$I^#<c96qB0|
z+!!Li%fXH4mk`r+1Owxjg{g#Goj6fc9YNEVfj<$=XR=slCTK&MaX+|%wp3&aFQrBo
zLL3vE20T#0jvu4eL2h6pm*09vsP6BGGGWZrGdGp%!z)v0+_vV?4Co^_Zipm>0sJ|w
zt}U<;>sJJ%_&{c+NOlXnlH`hdGenZP*rs@riv1lH_MUG>3cc(MNrZ9I57+=xDx+NH
zc_?gr`iv7S19*fIUP6BeVo*t#w}KyM%YHG~-o2u*Q=trtr5}70c~BDkGd|AD878@e
zjz^5i|IkI)izoh7OWo54_9?TNmDsOXq6O!X^fpd7(Gk`Oew#<m`oY+^XkHXau%J+c
zh3wq`)TRX53Mt4Nnh!vPj++w6x+*O$#J{XeDAYaftaV~7U2=*uBZL*JlS1pTXAoN@
zC`--YTH3}@I0BGINHRa^d_UMiSllT060P~n$%8yIcgJh$3{f+u8zRvonXCsJ>+4tP
z_^#i?w8zYWr;gf%;olFGY+b9y2S8Mg_$oK}2xb)yYjm*eig|{ppOc)#5Z@C$nB<My
z2#t?YH=?-BQ#3aAnL`{JweBCmhw?Z+OAyvb$q-3kW@RQ_p7yad`%2FUFg1@*8KuLr
z>y>^@rv;^OhD+V$393N`^sO*$f5fWOk8u=}um15;l0V6Oicxmo=W>4?AmCtz3x#Zl
zD<NHnmfj2N-gzNHX^Jn=s9Ubv(jxUDUSw78q8NE4y^A<r?PGuD80iqB-S<>Npz&Y$
zlLh%M*FjO*(t)zjA~OyqOXj`FD}nKR`jugoekn}l))#$rYJq4UkDAqhK!iUwo<m49
zkNSigBE_Q&9uf8PC8qCr4voM!%83$H6}ab)R#mV>;|%_}T2kC=H$fdX6T#TcU7rVy
zFZGdm;%_tk?w63r?_3TYVc{5*vf#&pssJ@i*kjf4AM9F9%#sHM_KL|L4mt6Q!4F|%
zh@P7<*`XVT`QEy8lH7^|qUmv_0QNa~3e>DNrI%t7X(Xf{>n_Iq1byl2(7~(+t};8Q
zp@+vDm@CsQqr`$4wU5H*6NrTb{0heap~;I9gf2SZACFnCI&)DIH)*z|b8Q?~q&j&e
z{BOa*z#Fi~G3Whmxxy7M7c!l``fRH;#3YsCp6$&#P8urRp?eyWxyPcrxv8h*`;5B>
zu5pZC>7_Zs@>5FlbHP3AnjSB_Ju4GA;Pimylyxpw2Un%hC;Z&=_anNz%u&p@#k3cj
z1wRFjx4OO-$lms&)rbDHHppE{#eRh{ur)0%ybFrOuy<0OrqTR;r%CqCwu2*{doCo8
z9)1+VsuS<o{F-OpS0uXg6TX~6D}ZiqTqed<>h9}P)VG4gn2MjFzo>Jy@w%j1^!RvL
z!`z49+9cE^!R6lZsL)5<L!HwK>PO~UUB5VK4l~UvQH?@LXD}iFQqx-+g!@WY3eMdu
zqI{Uf=%Jp?6;cx~F5WM&V5PfgUQd7gvH41yhqkhg@*2#E`{RLqf%p6Fpp|W1oqGP!
z0ZRA74W3Xwohyg!<Z^3&jNk11&Q;T@3e#X#Z+BJsG?4h<mx96h+_|q=b#DU^+caAY
zwrDC3^uLStncU`PmK*fv5{KjbsH=~o?mt$J@{p09e<_03SH|@u3BG#Wb=}!JP<Qp;
z=d=fcPlMg_>FoRZ%6Gd+=cS5!X|lib5q>H3<lxEozWSQev3WQDR9?%M8}VUr9_;~l
z31q4w;Ka=7U`f`E{lze+7?Z%3v3&J$i(4|eTYgu2b%P&!4sILjjqOQtv$Rp`PYBlY
zl(6AW-&&~BCugiB-+V$XvZ(pGafwnBR8`9$$&oC9S|G92CG^8wohWn<w?PGC-uIFn
zz3?$F*Kj80Y83X3+C%K!!ncDka{qPPhw2C1?WZzU3s!F?8n{sE43h@^{L)WVuAHOl
zGB})L17b%dB?hV$6nzViIysyH_W8D$&hSFrsZPF;)5gxWst5fZ;{12~m^<QJd#Ef#
z?dTY?2o*Bt#0X?W1*CIbo%NYcKiEn(tg33S`_-gB`eRC@qSVXvQdLbC`Xg1}x`zGv
z`;>#h&5Qu`GW|j;qwn3QI*ai%|E<TftUO$9Q<L<W9#K?d=Q(k75v8t!I%c;NvBVEL
zOD#Nhg~U6(bRa3}p{_J%RIu%PjrZTzLL*`Y?>^r=ADLMHu=m)Vu8L5DE2J}RX}HMU
zRLc)K4(Cu)9BRN=F=|F87=tM+<$CjmwdgAre@3fGAEPmhx4u!|Dp$K;E4!Am*{^k)
zcOj{vhw>W?k0ss1$>dJ{xo+Z;m4M3g*YcQE(lthf9Q2@XE(^;l1|(_KwWY+RvG1ZS
zF}2fJ`zS(uZ3QAfZ?LF%X&=kD`%dheXXQ78wKAUVxRcZL+GWww`E{N|hQ$E2A6`RO
z3PD!CyY5wwe0rGKFZK0JA^lAMdh=d9Uz%p|ScPVKAdsXrGQDJ=Tj|5Hdl<H5M}CZc
z?0cN!F-o&V&u-@FyJr&#Q#M!l!%r?yvwEf47151bU{T#I=o^a(izU9+1p>wHFjG3+
zCJvtSn=@^I(yk^wuZDv~$9xdEqlvK#-?%<~+qc<7)epW;3Cis6e|?AgBVU6fUGubN
z8_N36NU~X3|K~TZUz3@O&aOASFrsU>`~+co7`M5`uz#vj%tqDyyRm@@bXLN-K*ZD-
ztNpBYFf0mgM`B%lkZCU2|E~Ld<mb|SzgsJHJyH)k<geK3IoDa1pl)bXb@VLG{!P*N
znzu;8zc2jJnoAEDf1~xgGMm#h2aXT-`4K8NaCWM-=<D6wpmN+uQ6=HA6SKL`_#w8%
zf56h#K=M_~KvK$}!%hwRL!Jxy=vC0!H4@$IFJ~10y;S;6p}Y2lrCjM_Y_wkCLdnoZ
z#s7K_I+O2*9f~BaB^c_El*3>0PZ4u8+<m(OAGi?FCW{7kXPLPR&lLvjpB|<<&(?;f
zT7B}AP(XU*V}3sxM@|zykH~C1mz)*Q_kyk%0=)<uS*KM#VUhxe$1|3a1s`;9r5K|M
zE+hHF$))xCg`@H}k6~VRS7t2v{k0Mn{>3BlwS43jq)X_L>#0)}&m5jxAs^aj7o-k&
zlCO3Li(W9=F6XCUDHnfMC2|;;fW^$4CDx3dB%c(TtRaS)iD>DJ*_d4R_$rZ7I*x0@
zde|$`cPT}0pgX0<XrywKm%DJO&#Yc<*sA9YB`v@z(_O#Q_D>C8xZc6mbk3n{9lOiX
zk73^Uqfw8#&CO~sUc?1Z;sz<x9E)FJTMv;ux1c?NXQ}ELw;5MGy4m@CLtn>Ju9Q#7
zxhkOhA4OYx4{#Md*)@w>-OSKq{w7{@wc~4J_W_CFhOuY$1WStEE*H^C-UIGIPbHB<
zSZS#_SNO9&nocc(k1*THRcdW9XIT6_kK)Lz`XH--n?oj#TJ|u1T-sAnzw;P!|0%V}
z)W_W#?ZLBKJ=xCvH+rp3Oo?`sg-H#+ECY|1J*X|{Y-%_cEKu~vaci_h)U(6oFcVJV
z&-&}Mg@DZNx&TDmDCsF75P`)1#gghYDy}#(Zf&W(SjJy9yR~qCpDu^BUh*nkEw)mz
zG8Q8jV4JSTIMct>em?+4I6}yZ4n`&1&k*?wZa0R#9=1g*4_fcWFftLZXffw}fD4hG
zZPCc^GOp#5n7YbqPt3KB^~sHQ+&Wp?9>m`3Nw@6tBnS4m_k>()x3sMpKe-a;gSyD|
zL|u>_&qW+0DE}!Qa^X0m$0JzqL35M!@$M?pNUA(-g&*6AN$^G1e57d)Xk4nVRA;o;
zy3UQzCtq;``lQFK7Mz)NzJ2P7Bi~r>bjI$Rhm4gguWc<0$MpTARDMNmClp|Okp$@w
z$q>hoZSdf?Yg+V!f0BR}mV+yB;E60_z`+}Nwri9WA=n+rEp^avkmLA!6c&;+{<<gC
zo!5?L>y@=wbi3pG1}KL;@wOe6cS6!R1%C|3?{W*Ziu?(Z@Sg;2?!37D`Z$GFFUwh0
zf@MyxR%=&!04iO=BYvL5&d$WbRAF=k_eP(F+HBqG@p1>bP_hr|<Wl9IGiMh($($C$
zd~_Bnn3*~rP}qZsA1A|u53s|T4|=2OO^c0l5taduIq`&akvB<}<~NvGJH4S@v4;sy
zIYGWJOy<10Q3lMk)xv^{wt4sqq3_5^D~h^-1kVu3JN35hvLUjmY6NC&oZ#aH*bbP9
zxypN$u*2MKfUtO-7t1tcub|QX=C?h&jEA_bZL}1Z-}QL9%Le{?b-5`y#MeBIv9Ook
znT@I1TxH#+62}Qt)<NE=@eH|9S^gv=hOJnk*Rr2kG=7GVGcsP(0xL2;jomWViuKTe
zAsVns3$cv0CPT;EQW&HZ@>6h_R{PmJ`J84VFffD`7^h*Kh^fuv8O#Dg2LH|Dxn#_#
zY82*b6v=0}?h#e^F5S<wa85DABsr*cP9eIH1BV8>bbJSNuiE3P81IvsBd*B$;L3V(
z+SwMKi-Hrq+WzLSzRYjEqWxXQ<=P`yAn<jU=)x0R6Uloh3^s_o$eD1^EXLM9=b#C#
zl*^337K%>rlBatC2DY2ZSH1fXE6Mho5x!cmm_NwS9ajAt(jC+OXnF@{GR*etlYbD`
zPlW+l`JTv{%uw9gB`FL7!Meo3K4LQH0j-1QNW(Z)4#@A4d$ghCQF{2;@fwO+sSE})
zMRy6s*azcPmtthEhZ4bkvl&_Df5JYjeV55c=pKucTyzKDawQi1nB4g{SCEUDVJTUy
z#M7eJ9PGQW3^yie!?v`cf>UMrf{nBqkDn@Z?l4$W+{&n|!}#7nUMgj90?KnVSfA+u
z(>EU9R{$`QQ{6S@8SDz3ajTY^sH>wqv-x;>$^*(|m>m84AU><L3ZJCP)IT;!WR0kC
z!9kiTMqe<xF$FAWT?)jS!}d$bBHxE3(rP6^a2##kNwF4$CZw)NFU54xulu<?DtLT9
zCpt@eX(Sx)i&=Abb`j)WAnsZ(U@T03-oPiK5>wVQm~mH$cxc}Yv=a&rqND^e+)oQ6
ze+9=Gkx0dux{2sGI&ur(J4R%1Mak)2oJnID-0aY(A)v?PZa`?ZiTtJ*F)^h}8c>*{
zz?L$7S#;C%ggsIJDo#K+B{C5$;J}X*@aWp#>}ec$n{9p-l6C*Iyj(6{eQ(pIEy4fo
zB>ANGKYBg>m&6SxN!lyVrIUY2*#Cb(YhNPtcuqzhSB8oo@zNxB`oEnj?l3u2PSN<1
zbm!W!f4@XzE|&`~ExBKdq(qdcq+A#5UO-w{;>Nu6XIU&iXnlR%FRUXVN5yhL)3P22
z4n8-yk$Z_RSc3!tf%dh4{18Mo<R&0ZK&i&Y#$m&o{KP!73pJtx8v`Qy<9Tx31e>&g
z&a<Uu>!krV6?ze$t#28(pStGv`L(s16+{)ri;w|k%a>r#D^k@<ZQl=_{OlhukyB|d
zX35s}<l@ZKu1hv5t4<#6tP{3had$u1?7vcbw|?VwG8O53h)$<3$G*e{edV@ZSH)gP
zKTy{@^?_q5%(PPO(6390ci>`;6Ww|<Yf3b!O1yax=Qan?46&Klukng#_vmV^ciV4q
zbwzi_Rq<<-Z^|xE;qo;YY=G|}b3ny;Mx@S!zaHQrBD>zB7N~r$r}W|noF!WN8X;7F
zUyK~zI8B6TeY<A)?nsAW9&>lxrvhP|_KGK`Jo6@?JwJUlwtR^Dn00#~LKuSiwPMuw
zsB=MY;Zdp=-*6s3+zx#yxm_>*&QpNFaQ_c$ZypY1^#2c&q7aF)Q&GqklHE{AMJN@O
zEt0Zi-<hFJWlP95jFK%P$sS{ub+V3q>|>q5;GQwddFJ!|{+{P|UC;IV<M}<$b6@u#
zbKUoubI$!f=e*DRob!6W-s&6D^hWBz=lyv>O9H?}w_SSsOs`$TS-n+S>3_9UmSxH~
zzSocFtV|h6T051(L<flWHn{CWL-rMW!4<G#t)cWz_ui1$0TAvXg(ex}XeMr_i!*je
zZnE84tH!i>l{$3N^m52&tiq4m=yL@3-skt1&>!f(CaF^PBf8tjy=JwVzAO)E$WF`U
zXiBKoxu1dM=l)D+cc-nr_mAIwI+DI$d*=<LJyLHf?_t%kBmB|IE=MB2HRLRGBc+~@
ztBT4Yq_}(LFfb2AE1$v|+%E1}{h91gT$U;({Wt=<ue_e(WCyBFL#L003j6m|bp5`U
zH%1Z}s{;In$bNbPFm~2A4rjpHfG}c@@iAZySuY$%Rs)^Z{BYO$ePl}L0dmOQ{2jw8
zb34p-*Ux@^eyAMgn;9J66;@KYiNG1lp)B%rl~qPm;Hsbdv#@Vbca)ojG|2HRqJs^`
zd_5q9q_0o>b}FDC&X8nxnc?foRZ&Gw3`~u<&JO&OzvAweT=*;WyKddedt{Y57oo}p
zx;FlCyVsl}{<nH)CRq)VjZc1saAR_`%6N2Tjr|2Vswk4uU-+XDxo~4#U#(+&zqQe%
zX(=8{AEP#<do_PvO@2J|s^%M21Hav`A%ubO*MfI1j2qyhO-1=(q^#r`sKVsg^Lhil
zh8^_(7?A%pDs6A+@csXOCFm?jMy15Lj_Jf9Sy)uR|4$|Kt^9ve-~2C=2O)<u1H<m#
z8}GnP`A2`K{d^Bw>E>y0__=}h5!>{ar#{&Jlp5_H_QwwU7u7EMTN&IIFg(OXAk^JD
z<@_+joaJ`Hr%QTAq#J_~`BR~UOr5kFPv6c+;xN)J&QE>N7!neXeDt<D`r60SxtiQ(
zrQSuUd<wdGYxI-fdeZGy6Ys;wP7w+0v!kBEXWv+w-8eh_;qI0%=hi)oJMT;HSR}U>
zE|H4{-=nWtWIX;Br}~z!ASLE(B)i#9x#-_x=2D@R5?A(4BI3*9?>HJhKRD#k=Js4b
z&V6zHi#wk?Lt>0xID^KCK7@=~TK2J%DX+IAn(B?ciT-Kz`Ci1B>(hfRp@I!9TD%wI
z1`l6y;WOWL_S*62b&_v;8mV_+wQ|-m$NK$IU)^__+1zP?H>APrc7r6yddJvBVz(Ox
z7SIp3N^8ukdIHg_(=+y8o-0I024Hz-@TrkH@y9r39UCnK?}A5fyTE1>wW=881aUXk
z8#``G9%plQRFIVvui3MD&}w*H{O5PN@ViJA?O3M^>`#u{Z^#BwcxYi7Vf9ttj|})~
zV(&M!Xj*kdoc5PA23gN&Zh8i1=q}5$gtJXc?gO{sJd<Mls|`#~PN&8po!4t@J$j8<
zQUqKx-GLSr!`i=Zep0#*og116SX?8}r;ifi=G7a!LVv5{M$C`CTR@N!b|l-wHNovY
z#)}U-rELbwWOIaBG4MOdcAH|Xe9XP`G@*-YX64Z}s4RxMB;ugXt2xJ@Zb{@I0yTjv
z$j_!oKA}U8_67B34QLwucE?Dpx|LdHuj`0RWU9qsezyBghNCJo@Al)H>8I#xgcD$C
z-1Ebm>iyDN8a44Fz<L&;uyR>)=0}a#blm&(>GSOD^N6*CZOL3J>wci6DMPUX;j7R(
z41<^4ik?Bl=yTk~vZhO?xhws9<DH$uO){JaoZSi1)HmyRewkY3RHpdTdw+xVLsT<h
zndEIm>xh6Ih?Qb4904v8^jp#6;6X9#8#11di$H{{jv0qHBQktbXUG;}+6wm%Qr$bY
z>^^zBll3>>ktMUKvh+u_8q}(_HN#qPJ8xQfOcj!%nxA0gq*1%ov%O!MR$~MhYDXEb
zdM~nXUw3H>R$w1T%5a)3M?EfJZQ)k=_pocMBb&t8H~q<pkF7(336b=Ye1izQ%Sb%F
zjz(1=mf#94%eEzSPS6v6^2kM(GACB<eP0u$EXh|Qh~s~<8tHWc$S=RO&7MrCt~XvX
zb+TW%Ai;6=r6Z@~mt{>d>@$BLc;$zSN0=Q1l?h$iUO^y)+d7`kMl!FDpS=$4ij{*?
zwonu`d(}R~YI*##xvqN3&p$7G;gZw6$ZzvQkG&lPM>VPLGG5Qbk`8}}`obd?FJfS*
z;Lg0YKCe%Qt^cd#oky#QfOTk@W}Hrih%j-+?4O5Ae^w=S?UwYe422s0REcgP!Aw5J
zihS%HVO6(ad#O0~t$y8a76@>eEomO&;jHvWwmz%ia+hg>Wm;u5%FV9ls#sh!ZEf7z
zj29XWoBUT>wTTmsMQj9efjDb=zISRMYcEBF0AqnfhQg;bdHn5B&o?C<2#1F1UTa~V
zt{HqP@tA*K_2Qr?M=@Pq^TaWZ9|Agy)J<hsBQcAcW>3}?uW<_;smx*lEkajTrx(y_
zOkb?ZggmFn1X@%EXbLqF>G%k3kCyK6;*PI0=*Ix2$H*{?Cce?X{IW@$9Q!lnAf@L~
z*Z}NvZRajbi6tLfqnsC#6XbiXADHMYV9nm4dMV3ZQxK5i`CxXH*Xw{w_mM03ssAd7
zVj~Ucap2}ri^<UBKakapf#BZKYni3OIYZ$qb{3uisJ6$!y!hA;p|60=te(}Y5f=L^
zBBYz~{DWKBQ~uiE!M|(X-;wQ(5s!@(E|^6kEJ9TCuUgQ;$@$3FA`#Z3K6C$SIWnI@
z_Zbzq7IIaTFxwM`jxmXJx$6ldqYtQ%qoDID(^k=cTo@gk?WftV_F<iUh%pLN^~qpK
zjLYXJ!<pOfzU|eA&*M@uo4jwE@fyXlbv$3EeZBua_NBF~-U$^iP7=Dx+Tvj48J<ey
z#cvFqoUKx3BN2Hz5mRWZYN+~?A_i=i)hfFpoV8n)QA-SKai%8339ub~(5`|`h=<t0
zt!5oxl#GQV)5NSb8?p;Hr+Gz}(?&p;_dtH5w{|;=1Bf(eHCTDR$$Qg`$w{hj^gc2%
zN{L7%DdQtflbV)kq2is$3wS{3dQJ6FSt`<#^tV@kTgSUq+7U4gIj6fKE<w|&Ezec8
zG(L7bR~e}I!9r@1e-Kxdl9{`CC2lQB;pDh%pIU{`Gk@7pKCVk$yfc8~47-8KT=rpa
zi`SZ?AzjSgjYsLhbblc%xq}Kr1C%UMeOV$)6_PE#?1U*mYN?^UXf5{NlbE5I8dUBO
z!(F^c=-ICitVNoCT`Ioj&}?RD?t1JtE=_eWNl^z{Bd_ES!`V4he#YP7ym^eHEopE@
zQsw9F&=-HtMf*^AHe;8>GVIIpn*_I!<IaKx5Zd{wdKa1bOLmlP5=*-W=xPO2F?LZ%
z)K^m#WNiCkv!e}dqnKZIjMVhx^ZbT0{+-aj=e}~9Mq-+N4>tr_BV%7`Ixw_CqYe9c
z8=qfUn_kCl>h8jHzaq>F2O#-+7?C0XjaMtL>KD;h@`SPp9S+kR#9^nL9hBmq7LDEM
zk2#^9J&RhU34+Jjt_=wZVU`Q3KWe>D*0J*9>Bt?t=AG5+jf~w#eBR`Be>tFbo2@Sb
z!2+n(+(>vU?NO~2JxW)y9sQ1VseUMgfcKB#nTI9w&7ne42x$mDHuWW;2+~8%Qnkgf
zdOPR}P2>D@#<%&ygh!gc*~y}+Gbte;VMbRI#*4}9$a2xd^m|`F#ud=|azmkscUCe9
z*(s<0O65s#={_&XlW@U_1*Y`OUm|sxDs~Pkx!YUck9bR%?n0;sN)~i33_bOM{9U=9
zh_1n}lo0Aw;&Uh%y!8lei@o;Y!FQ~I!;5Fw;l=$+`w0gen}U@(z6`p;IFP+%kptyd
zxg+M(gs*&t_~y?Hp@(-*`K;#?)WJ9sW&|;3MUnQ1oqv;9SF|6@tUGkc_Ku9xsSHb0
zPK)CytPtDVNewJHtox5f>t&PE87AKe>NcEat&y$X674LUv8;R?mxS4}7Zgsig+?PZ
ze-o<gI^Q>=cWWGS{0a~r(=wB83?uHO{3G8Gnx#THHIUVNq?Hfq%9M-ux4>RRWmEVm
zX+YZj<H_0UEA=?Uur(rI`$(`^y?vv1re}8S9IY&(y)k}9auhQ^dW_>ztKo{47vho+
z%IE&&N#l(v)6{*UoW%24*bQ%Gdi?!%hJbiF$@%#Qz&+<ocs|_{x+vt_m)~x1&KFZ~
zZxycSuT0Xn$~XKJh3p-qXE>d56ibzB6jM$4ono&*vzK+8IS^#!b*B%n3GZ>t?lp5_
z8SQggm)8?y7;b4Pu*Zo&pBX<&*h~k?<Ov%&X^8k$oDI}H@PJ?m)dFsUeKc#uo(a^9
z_|ZcZ6Ym!K>vT}6&0sJn^RUXlVb#bWx4~5Wo~j6k;E`ioN<5IV7ZUTV0;D=o`SN#k
z*`ry#uDZbygj*_&a|7#W>CRC|e!igxL9qA6{F_?yZQwpiKok?~NT^&(lTMyk))kbt
zMn}YIgZH2Qz(f^v-NB~Aes^tnu!8S}j{U>Q^>GP%2y`5Zxlg5NGX+=pF7mht8(v&I
z*mGNXN)&U*WPol0RW4WnFJgZ?<KnDDz5bhhtfNDWS?=G^UPN92{Xw44-}|nJl0(5a
z3HXfd>1Oo%lBYK!&oUXFcXV?u3GTw>st-=f395}F1%<sWm|dIM8!STf+an}f=<!0v
ziJ{QQtk@MC>`o!L(e91fuPUrTouNtLGYhXl>3}<KQ)$a8cK$jvK3d9L;JCsdi6jf6
z$Qk^-bj~yX%-x@<+%eFWH$KCzG^;W*@C8$$rM#r6bdnKCzkID|;{ZE^(K_$iuxi)v
zqD+_JB0)7y+lPlw;u)6jjfcWfCf@dW%mEy~cpco!qoqv$phXMKFi4t1p;sEp2%@ft
zZ$itclBc!jl5UM#Gwkzn+F0kB-)nl=oN;6sRldxy1TP1;ti_)xYJufW>KQJmtr@-A
zFmw@LE>945C3Fj2XWChSvklQ^J%u>yuk368gE|&6<#HXW5qn5lyA_N(<;%4A?xNm&
zxoDE)4U=X+cP!AKb33!WaZxN~L3IJ-9uVeegsV(S@!XVCWmnSF-kzq77w^RnZR39I
z(gv78JX`AZjD(@&adVRWeoBSCI5dkhujhgq&gbq)u^&)B-WKjrGMDwK^vk0DE-rh7
zaUk;}vhrQzJ5vMCG;{5)lGjxE8JqlGf2|I=xyI9m+@nfQe~$+#tX~lmyGK~8w*Q8O
zYlP?rzMr!fC#8jEQSG9TL`(2IZ`!<&zBpNrsx4)@v25UF2}QGY535T%`3TO__!U`3
ztoMa!0ggkDWfu;YezEQ_up(WZD->?AsVdiGDUsi`_8Jg6<~>s#(-A>G9>bP$7zma>
z^xCPG{8EQ)XGmubNIPL%ESIS1;BF2=1rH@VKVz~dHiw_>kfD{2Gr{m~FsY0VgV;~d
zLqZYb?03X!eMC-3oNASD?@NZ`fzpZFS<*CKeDzf^&0y}|I>*>v-lsax+~H%(-kHl8
z3@_`{Bf&mYs{>tPeL&fC2QA+RnWu%`vCpd(BwKSo@%gTL0`ghM-1dQC?@~`d{uJc?
zRw{Mnb$f&s&gnn)itRbb_R`CF5vA(mHResJGu6!kY>feO^96p&3Dfu7J8MJB5yN`z
z;p`;X+u72O%M~bmcwQNFWm;`!`WTykEpW)(d{F|9&fiLPY((O3^EcR4%Uk~A;^vSt
z;*M#ozF<QZ)GS%>b}$xeL}VOQu?QLaSEtN5uiuJLjgpujq~&`JmD$Hpe-nM3Wh)Te
zS4zPE`VO_|mnq{pGdLQ(BDg2i;s;QGI!Kc#jT$)#cm5tuS7|d$G4NMHt+_57R9tpV
zzfRlLO69!+9T4=2J9dgktmO?tn62^cLx<=-z=sL-rRPsqc*Dk)r(aA#&k08v9jfF}
zn3>_0vQzdoa_MbX7+aqkX*b=|$5V)^&Z#){`~r^u&=O8Y#4W-|!C&1>n(*C^k1faG
zcswiiupW<?RFK$c=~X)VrBmJ9a-h9x*UWUzFc7fBrL=Eq4JI=ipcDg1!~NHJg~4-1
zbgkzms)LX4YWMuS|2N;{_RDP*+33{6P2=@pw%U7sNoQZ()@8T3A%?$xEhBN?%4w$S
z9b>+%fxn!3#}eZy`r`*Bm8#!c|C%D1GYr>bZ&%rwOnH}O|7suOOrD*xcPpkEzTJ7z
zM`6<G8ZNjV{#>K~vpReJb$*Cgd-y6$m|(=MnK+ZBdWW2%(l~!=mT=UqwuOUW_lP9y
zF$=rbV(e>Q0=*Wd-h;Sv0arFb?JhJt8D34py225`xC@c|bG>JHn%iF=pau@QJtrFB
zUlmTGPqH;?edvgE`o^N-Alc-2>JrpKZL1&ldm-#^BONXy{G~xXs1)CyJ#DG!_XCsF
z^_Vc5K6m#|UHdaV7zTJ>JO67YYu=?%qnG|&EN-Xo@BCkHbbziT>FnQ7o;`H?!Ku2m
z>(in#a#zPkgt&h5Za)6w8kpON;6(Vvz76G^Yl_tL`losQGCe;=ECTQ$;hFLOygW~o
zoE`E`*wO2Hx7jOc?DWUV$2KHivw(E*=<4)Zj{!QRsa_|SSjIUNd=6>yjp}}%D4d*%
zzqjZ^9R2`391d-mt2eOtTbOV9&w;Tvn-{729@czI5V>l~bNc%8JJ0T>2}n>R4rsL3
z#QnN^t<FSM#EHer;!^^(a}##{C6(`5bM>W`Nm19j9&g{Q?%=6shMbe~4IYJJinv0^
z?=3drOAT}eSFcp9#9uId_Qa$XNvW`FwoyyQ90Q3YtW|s`HDBnHd3ht1bVQIHb}oK%
zye;)i>IZy}DPI9Bszpe0r-DxYv$2er=pLjvJrIeGykzykp#8Fs5G18A3BCH2X@@fw
z@wFY<{nW2LRCVqAWd3TOhnL4Ix4N6`lPwia3X}5H?R=iXVkxow7WfATG)rg5@;*T&
zYBP1m^=wZR^a;jvgz=WhRUCSGnTMor7^%ZM(oM@T9qC7!^O+7mw##pekTVmB^<xb@
zp?6J}ozQV$KQ&UwK%shhcTu2~ZK*e5Dt;YX^|<bgzfJyhqE<R2Aikt?>!aD^Q_OLW
zK&4qihc`aPbSdj}W{NWxQR|eT7OFek>M-$MFz38hb3HnbMPPN>P3G`9pVimCM>R`2
z;GPy~UC)tEU)R)qaH4pwqNAENJqIuI5jQ$q)*6Gbk>l7%<g$U-1q*qd3ucu$_O8Za
zR|8<bB#BZdmdW~E%k|c^CmT!#{hqy0t9B`E7mB;0YSw}`u+urD6cpg>i~Y21!`!e;
zaQUW+vP@PE6O<KD#dBBhwUpbF6p`3lCrjQCLlW2CJ{RlF_Zq%9bzqUQbol4S)7O#c
z`vy)T`3b6-L44hrf1}J|!gtDYOj{3Z=ZVid>&%@x579MKs-vHjbnj<%Z6@6rIJhFV
zI(g}>ld--@flJ-y(GH}-TTEJSxZA?BhH&R3T^jfKnJPzJsm8zf_~e4cd&d`iE3&U0
zz^DtgxSdU8rHvd}%|GAhsV8=+phna7!WE9VS|K&53(<Z|hU92$Bh~*ou@`2-f+{Wl
zG;MiKkl}lq%SUL#c_-v}$fT^=AL>`eqzN%cs5UT*D6v|9)jz9=uR&|x0#mqVYWSJQ
zH|wa6kiHD9Ryuf6os`MjE?TJY<REvW=Ly<<lXDNMmm1hAEQOo2R~~$Mti7UY08vc2
zo~~RxlI~*MIsD4uQhfBe6B1VXtuJ(ZvU5J!nS3xFYUUB)`kS`*3n@GGD3tg_9e%Nh
zly&^ZcR5VQss(<}^y49DthGgKEvtTW-T(IX#V{pJrJNW+fyVNKCGzK3wyZn1TGTmS
zsLxb|aOS^NO}3Ob>@7Buz#e%9$4bSox_F5^vwJ}HK=wXYm72228B~=y;G2`5rfO{1
zfElLMpN+LjG;0mz*yGB1nRO-3S~h=y_h0>(&Jw;&6-OcW`PvQzEQ$AdeDJ-dG>P}Y
znFR$b`Pw9HEiZBkHT=4Ks9F(qD$Md6=eZ`zlcTU)S55=oyS+!AnGWju=5alBh(Ud+
zD?xQWPd)UsrLIdorsUiyNnu~jo;I5|Zi=ZKQIAJ2!L%PP$pg%)UQn3dD;7qa*?OYc
zf`N?BmvSGb=~Wg%D_h@fsfmItpg(<e5BlPnMn{Rukphw!VJg8`SE;R{D53=^`0@z1
zz3}8`^MS(D;tNLkKIgaY-EtzUm%M~lmhpZE)SG=bOsh}+M9n7MDfLnRV^ToW;`G(J
zb-b;xJ?zce^*}CXR+o$}{{tja(wgGR*WVT=U(aewh_2V~ViwwPtL+W5rbgVSBtFWs
zc`MhisiWeFFBbk})Cf}gT*E2};#velcqnpyHfvw$3E5_#+)poUR@s<w|N6(46J8JH
zA}6|5kmos(b1D{kS6K4Z)n>y^hLFcZSYwN3kv>P4h+ajkRYy5{NEO<*NPg*Zq$d9Z
z>(J9-R-20JE*h!ULS1{;C!h6Jit_P(dh=3LQ}K-1s7sv$DU#9}8{U$}nwN9Zl*<yL
z`l6bt`^W1?)S$Q8r3>1~(98j?-i`N@!NOrFrxD%HC8VNeQ|C&nKS+sN^*O}R@2MRs
z*BmYvW<9F;qvkF3NBvt%4Pz;54J%gtHN#h&$V+#Y)-E=A3?9gQ+O;b^dbx%bX89BI
zNLIs#+w6s-pJ&}z`o{f9$0t!~kNhif-#x1~PX596zRVDK@dQ;WO5W6{bt`%5cONtN
zx|hC5m^xdqd$w6mbbfX!ZzMQoB`&OudiBElpa*bE{H84rZqceq%b9&u^AKF#WhJht
zV7}-O_VCiFDmmZT&^QMr4m&n~(hOVz@ofmhfi3GeB_U^fH0coXnj!xo%7EW1vvTI~
zScd1q@5$EX7d2@)ly6j2NcBCJy5_iv;;|B+x<i3p0j~#M;X?x)L{ID?Ud$8a^b@7u
zVKxe{h!@nK>@gDSQ7PFFnKS`Up}*8mD(RsbzNPWRNuA!{Hl6KB`Aies3AHp&$uR19
z7bn`=`R{HqSw(nN3?tpGN!M&6%u8G<`;PGg77R9{nw8G$eA;#9@`C!I7h9vlhf%lA
zdrgF?<zx)Jr003d=hPjd24pM*h?`hND_9qkg!5%y>riKv`gmC3gmw})rFKH_$fj`R
z6f2Z8aYIl)2mQXi(ZBUpQv<i{1Xm(8oyvWwukJQ6E%Z`fH&Spqlb?C|W;x3^KldY5
zY1;<c$Z05=so{((s-9<I(|VM%{|5W8y7*N`?VFhF_E@hN{Ho`|#mR#hjV<HYj9J=A
zv&4h?8FE$L=>}YfBn(s7_iVZL7AuXq;I_qHGA>cC_dia0H0kkGWB>M1iz%j#<8Tf;
zr;aJ)Mp9j?iw!qOGHQ*c&o(t|pVBDy#f{fLDxwV47eE2&q~JUxs%JYhSQ%ybXCU~j
z&!0QwlD9<H1z&X=s}}EAuO91yU>*y!zn68O8g07yu<PUp??*nO2={L@`1Xm})B}i^
ziMIJOWV_c7tKS80xIQTEpL$9Ey<xV%33rJf4V+kjCrwK2l}+Nms||P4k(r;)ev@&P
zu(19BM48Tk**~xTf$4YfynlBzjKRV%Q{GFdn*A(G%RdY`!eu7w)S+L~@>=oQL`75_
z64m*KD({@dh5zdb3!}P~aQs~v7gf~bn9;C6h*pbhg>gvzjX<}zeYtp>aJQPP<R0Oz
zOQ+su_hYlFDYqwelb#UQ=jE_-d{mv*4%Hw_GIHA?Nw<~qa{B@kE)MHL!4T5i+c?Q)
zDoob;EykE{t8AWiOJmYJY*cc(-6(HdQVF7n#bfEFiPQtrShw9Q{GZa7OEmR@t*@0*
z82WIt*7k&bD1XG7c2Eh|u2^=Q=uA1QgedVqpl>>X6%u*=dD11j$jAq#Yw~&7J%xJp
zG4)PN(Sh~4ER4j^m|XDn&7VhxJ950v1)Ji1n*?WrY#wXjP1MZ2m^-<xfI{=8>jmg{
z!T9GN@1@mk1zIfQ2yG4P8j#7I-?hZ2{<wwd(^KY4^Sb`#z6~!&_O-l9L(r@Zp3?Az
zYLr$KQF{o#e<4eBwkBc!P4~GVx4gy&B(}|(p}&_5yH!~k4IbSB%Q%zZcjd1J>*pWG
zmS8leP<k)_G%@zqZ?jItbGqa7OUeLg5viPYVS6$09Ne*)M!Y|vm%Di}C*MBnGPec%
zK$gHe%_f)icKk*=o_NjXA6hLUoPQ2y0XY)}we(jTGDqlQZS_KH-FYfdgU<>>WNA9T
z#<&4HDeaTFq_v$997SSZv8l~dFl6Yne0D_Iu}dtNJ(~G(D68_e#1S?|<;Hn(XyA-D
zhXgECmwQwLr58y*Lq@&~7i>1y9POZ<JqM$rX@%-_S!FXWbX{fgD*jr4UQxf(4E`nR
zG*1>QOntj;Q1{X%Wd(N)w3lRd-wH=-L4!tw?%L8lil3r!b%%9njd4BHR8e}r*6twT
zsvcT2Tq`8Loq8i;-{a-P(`pa+F}Ngq{Kr$1hpP10qvFf()tHWh@ls5rwI4Akos_na
ze=Xo<mdZ8-a~%>KHBQ24?bKvp$`Ja9hr0SLl*@!G?$DyI2vx3579&qP{-Wnk_NII!
zIUM40%)pz@)J&o8kSm%!{q|&GqLqArm)QRbyV~TkHf98alvx}Xx=~x`H7;RZ^Qa-l
zKX3abcx5WCsDAoxB9$UtJ&%7KWOuf;^w-Oe5l?at6~*{9d>Eyja8;l4<Csg<2DHG}
zL1iAB{$_&9_aBJg<Xt?vnU_FMQWpcR>p3ahaq0o|w_=1=yw+x=gW~D>*`TOFm$uqt
zJazU_MS>>+(J~wpTXx}Vid96(E&bCdR#Z_uoS+_Nf3OVbmVUhAp`6IJqqn8;wv_(d
z(H2p>%v`Jv*4jUtWw>6d;>yl~2o$bh$CbN2El6%DU$x69u9j{LMuvlrZAu4aq(2x&
zJF7LiXSKb){U9h)rQcW4xwyT#<^V+}XoYfU;Um$1p_@2E#Rk!0+YO25=g3&g(HCMa
z)0#f7nlH1Zv3+$Ul`kA~BsOLV3P(-KcM4W0^N{`TzGzU2p7eH7s=-$t__dxTx6m2&
z^rK)_&F4Sc)Qrx$Qk=;p&k*EAZN!a99bo#7ck3TDAa&d`fx4$l&aZ*pUL*NKzQVs<
zbbb?0_{>sZrvM2C)Sl&;2}N3mAJqc#tt_jfDQKg~66qXchHM_TV(x<prr8aBEd*j5
z(IWqv;pMJHz^UuBYP5{=e1?g&p4Lwyf;H?Q?o{$4)UCgvg<_;1o}bh(n(v=HwK4rX
zt=Bn&)5KYY!V{?O;ZgAv{YxE*;fydlYJZ*hYtS!+iC$5&(~(b(Wz(~##g$3!rd3;g
z{L^DH-vBd%Z5*5%+fvmUVw`Uuwz|lZ9oJvmz)h9qleko0&OMLO`BvqB=+!~&$MjG5
z;{=@P@~CGCJ#RAbeXaSFl?6olhj%jLeTIX}@-Xl>fIWj%of5z7!a!aCIC8tjb=xhq
zWySl6bgCU&(Z_hP0n*8ukdT`%X|Tg!?S|)sDEDz|6W0$>|Ip4G#)OF>W(RF3ruxOn
z-~)HaqXXo9uzF(CYG@99t_S8@wg=xy^lYN6AxE&Y3;vgszQyXpRERY^6vT|}zy#wG
zvzkXe7*Sj1*pQ%&d_5pgnOBO}U)8t87S6f_`^8DXK&c9m8dz#j$_cVsDTtMtHETpE
z1y_|WwrYLrfITA(={wSSWILV)+mx6rJ+wUhYAyaac;1DCnG7)w^h38!%s2d{e?O!^
zq^GV?*IqcesYU(6v`#T|+ebrtTg|{N9WfW2xL&-kP9X%MGc`f|CXN+|Z{pOG7|z7f
zIQl$!6f+XaxifB#-L3Fuo<H^W)T0SJ15?|}*a47qxksaR`X9Y&K>cmi^%bC(_LKf%
zCq6htgO;KXd&1bAHW#K=G}w>8`=Ne{-LU^n0?v@OyED4tWN!lect){%=Vu7Kzs~@u
z^z53|RI&pgN*-eL)}vRMn;s8_;Nzr0W}$Gp<MD$0j`kFpzo6*IE}$pINm0}7q@$*a
zW*J8bXh?i}&86g4Hi&wS=wa2tHi5b;MEeqon-h8mc4LKF`Z!z?mL=m0l#V~k?&v|0
zAMWSh^=K0NhrR~kAf6hIaKp<?XW9sApwJ{xjz0rdM`7<vp)#j^-=uMbJZ*4ZXUBSj
zz6lZ!?%StX0LfC@100@epx%$xF^L#&{nkpc2S57}ytD~N2n^+vcOn>+ix222I-Or1
zwI_N?l775nRJa9<F%P(?o$4_Ebr)Z_Ha;g@Qm-;lk_Kyx8x(Vmtpdr$tw7SQI%}9_
zM<%+p&`L?_pDyZFqNTLKS`eKWtgOV)C<%@@wteI@2d-n@_{n?*l;{a3i9W8v(TNG)
z{C$saHKL7oJuxTokm)hJU!fCvo$BdQfm+hn%bKX4&)*_Req-3{IWkhF)Dwoo(m|9q
zPMVO3IOBqZJHDiQO4k$oxe4G+sy+0>irBu^fEh{R2`GN-)>GIyi~ZTKkN5huW-ry5
z1>(+#8`08v#;Cr4zpbo6_@ooMn7$Llc(rbAOQWITSLqn!ryf`RUGp%{gp*k1P8B{^
zmmUW13+ZseF82UQzc98OOb3vOp(zTfI+xb>>VP_OIHHuH{vm7^CGgABgni}P8az&H
znp*ROSh@%2Y8yqT?uIwzaj@tGuDWcAg7}{x!73~IkG;|U_$=B@aqQR1*6Vsf5-3>(
zwhMB!iTPO|(FP`7i-qupyFU&BK*Dg3CUYb{=kE(LScR=TftI|#{gBSsh3`C6tcQ$=
zR*IqMiBTjfmcDM2Wjq7$_7gPNZqnm#;Z(^B9AW(7jUn4h^3nW6aEul@eTdiu*va7D
zI>TzRr5#>bVnCOKqjddnLq8{o6kZ68hjsRVEb#OKsMBystT&>3o6=7R>RR>>F270)
z%uI!foWQG1b4+sBp4%Eqf}>ateli^YsQ9TqwN?Gd9ym-xWK&@*ATEf4J2jGA4HkL(
z4%gAAK^)7CQ^JH<Fv$Ijg8{wx!stlUGvfc-Fi`ldMJqZ;e(4u!kF6N0d-a2|Sq4s%
z5eg~{NK6>?bdKi=Md2Rf9j4$;FC1}e^-{A4MVjX_4&-kFIT}J+<Qo&StTqD|8>>=}
zT7Lz?+xXK!eXiP#A}5{|S#(@bQ4Mf-wN78AWW*<CWumVagW}eG#0q7NLJ6GmOgutq
z;a>;0t0^CMf^lrAMdG7}0Xl99+?+*~$X}gB;PRzr!v+4dfTfE&2&e07Bc@B}6O2bD
z=RS=LJpVS_POs7y{5bL}<GS&pc<Un*3_d&}mTs%qJqv0|D?n(!OhU`wjy>$(6DgUQ
zK_C_7#t5BXmJB~+Z}k#!d<^K5aWZtzMaH9OpBs<7SxLC38P~?w%dcnn@;vudJhic?
zk+Anc=Q@*5`&)U)k*y(^DecygO|<d8eN!L()^8y9Mk+yP?MBBQEWy%T;-oi~2*(R;
z|L_b~y9S7_j&1@&8yebUGrlaw{WVB?T;(~X`{_3)pkB)N1x63PRcOXjJmoVFLRJZv
zIur&7=_cdd1t8-ABk;Z@;QiP)9(BWIYQD8<L#tYQpN6`y9?FL`luH@(HLROH7aJ$I
zZ>yGQ<I~UExz&NtU<Z<k`jRy2+zPOpMcD2zG~&<vcr=6Hf^h7>V-xrA0I2NjLDV^2
zTJf@t-U1;5x4qR?iXUE`9EOes_Oi9wLWf-&edgSP<1&j<bk9gItkxJ%`w~h5mx~X+
zHcn{ek3UO=N~AM16ysO@^UZ~O<HUKo05!g;_uNtgriw4TBkJ1-C4)?c#WW0v*MmyM
zNv1ZeI=RA_cnt!meglNpKy?+c4?OVZ^0ox(oAz}{K?UILq5^=3aZ%tR7bs%ifC-Y`
z@>wDIhb!%eXj#QZM`NE(oDIBnPA1FC-JL<-B$|Ij)Rvg)zw1Iwd=?m}Os0A&zXiw1
z%FOm^MsSc|yfwTadDFg^9=bY%&Z0tt`v`%3luEEb+9mHNcPCELH*RcCxCmaIjEmz%
zhr&;55k`sJ17Ln#^Qb=#4_LxGjK7WSPK?q^KQXZ7LJkZ`P9oaF<oI*kd75BuOZ^<~
zzJd1oCNTmb2;t=DScPi@TsJLr7B^9iJXk5y48Et*F_~%C2z634k%F!+*t8+jiqV=2
zkjV@Zdw>qPaX0+U36%Q`gSVM(2%<KDN}(_P#a~*L=dH-VS+sQY@06v}Pz~TA6BROU
zX?NEkG-#nSxE7kS)HFNUGvp-|$6p6()KnOu_EJsC;AU;W4{5!+6`JQJ1K(@rf(<q*
zSFC@ry1G)xS5>!3Q~SGw)@vI%%7I5{P10h;xEm3tvKYjcHF$xfRWpW8@x-a#Bun<v
zQyLL+5bm+EbY@)Vx4}YcU<*3`CPchPCkE-4Fxc#ACyGz8wl>Y~osbTkoS|XT5^i3q
zCmGzGh{M#pls5Lxd@zz~a^oI1er=fnJ1F_Z1$T9|3oZ!*ezv#Ch%+xi+>g#vMx_5(
z43)ao8vB7(LiEr=5V8tx&$)$sq7Y+K5R$#Xs_EuXgjMHk+==3N(Vn&5RMihd$=3os
z#5eF!j~j%Z*B$azyJ0ds6CVC6VISx~lBU1Gg9N%@?#H+}?t$B?BQv-Sr0HbIN~<E>
zYTDq>aMdy8UHc9?Rb&=Txp6$E2w-b~?Idi)TI(4|<hu58&jTU0V>9((aWr3N&7Mj0
zd9)ceDaj3wQK9nRv0^wPX$a{!dG1EoHah3}KIU9zWmpc_9&5GK;;r%AA-Y+An6#o1
z-vA*jA{7Cn5sx#_FF{`uRk^4Blb8?MOsRHer<#J;<&FNkmo7x}3_I|G!50WTf4l})
zBtNlbt~F@~#5|_FnxzN2=i=$oiG5#(j)4~)2g%G+XM2xLVo@?9MFqgg#QpW<9TX~D
ztIoWcjO?CY0+N!thO15mSjmQH8f-^nV4lK~4RkhD3jJEALH*Why}P2yXl0xs58OcF
z(MdSXiXM1!&=+x;&g@WZV7iOa41Nev1^Mmi^{u$-81e}a&ZOSh#WllzP=5pFphWzN
zoNTxN@9-1tf*XI|cx7KoU)CeTXn_Os^kWV*;?zM<77si*fvv~djX>%QEpYtT`w26;
z;9D|uUr}10i^-Tm#vK9F<9Ni+4v-7*sB}?5s3i_aY^^^u>Wbz-G@K;GZ=pLNJ_9<D
zevsi*3VIMY4ll_%M11)%LLm_DK^t!oi|Y9qM5+h{5npV7VygCqw9=m$Zl7wZ?cAg)
zdCgxx1QidcbV*h+k^1_s*8B|OP&;@f`H2v>tH6Yf{X>Iq?-=#YYVWEpV-TgC2m><l
zDZ>fI$fra>gwodic%5rN2#5o?wi3g)yu_h{u#RFl{wv*5ApyjK!Hm<-BYli(-)*-@
zOt}e+1HEa?J`9+pEYqpq6aJ&hTsDA%F0FhGX6F(t$qRlHrSv=#ZFkWw9WR}xIy992
zJ|;MBqi1i-BKXM%gZBgP*zVowL3tcq11fob4Mk*L99YA`TM#!i!Qd6z;2NBE6iQ&(
z#c_jP<5F;A-uNdqbRAdslBw_0REYB^B@|mAh{<5;di2wG^l~?ZJ8c|&3>1TTEX@~0
zGvSZ#|005jW8yNPNqs$?aU|HP{ZG6VSbT9Qpkoe^xQP>e4NlP{8OAnrVicffhS6Ov
zFc#?zU$=2r>9B+gfa;_R0`IWyf^&5Zo?E>XW~t_YA)ph09~CPCi1sPiBcR0%2tu@p
z@?>g#CskjFPE7@&{0p$7467FUhiWR=#Vob$j5W|VJQF^D7aS9DTW1W@jOogfoZv=R
zz5T!m5WKI4{_Tcn)3bMVW?~miW&s~_@;*qmKLq5-TR20iWZV|GHNy5HXucs;f8v2l
zmgyh+`u3&aIL#sE{7vL%xXmvE7Mf(EP$La0XLzpQMo?gF*9|sC0krW_hwkkhEFjP7
z0=nCE%94xYzI$P(;K!Co|AMokjwJNfky+q$lFd)Cx>CVx4Wt9ZC3FBW{_^59FY~J!
z3K%OCu)QBoMuPAknjmT)$AtXuFqA?w?9Rw!YF7L1$(HBmm#<w~+0}M7dS%b#sW@%O
zu##geZs8)y8k5wvFe2{?${x&nT9g3SV1*Z0CXtckyF^3?LMEj2UPu<?0Gs$l%_}QI
z*_A$cph_messZ9-L!+?*`|%afpF3L${t>|$BK=&$VDAjd4(N0T5qU%CF0WShF_~L6
z4CUJl-ro)FBdq5H<AioxsDqxK@8d!2Q-Esi2T|Vy8KG~$SvnFkT63=kh>}WiqzBYE
z-e7vwl6?)MF2N4e$LzA`t?v_qCV^H1Lxw5Cqv%q5VlBKhD!x@_r`C1{UWSLk7=wc3
zk8LcFBy|fm3|PGx_#a%%Jt|xn;0Pwlc!Q{rU=WqNPnVpVv4;VTECz0f#-t?j!@qoZ
z+jie%DQg`k<N`A0@RC}%wcr#KKzOmGX(`eu#eDQf)4=&lDx1PcXk%VfE8Tu>9ma!2
z+yM1E4KJqLI*mPz1C<JdLtZ$n#aUg@2v%k=Cr`9I55sDeg80V<Y}ZDSxd?FM8putX
zB?4(8^<mL8+YI~?`#<9J`%*x*0mc|GNOR(`w%zgFj`ifG`76~eC=B53uJ8n7I|>;<
zk-<pbH?N(R%CePb`ZZkuj}mu65h+h)aZq9@0+U$Do;^t||BPdMy~`47G8U-X)`4(s
zrbz~%4aXnaos&TTGAr8bF~bXvc~P|^*-bMiLkJc!1yp!z0z}I=$5^fE+U?GB+w120
z=%_*h>_H7gwVtPgQ7EiaYsO|?FXFd~!x3sp9+_fQ3hem?cY!ON=w^S5xu81Jq8v{M
zF-V^9>~g_c{_<A855{-xn-P*U9VZA3$}BPIeMrLF_>>niR<|U?y{Vgwa&8Fh6hING
zJ1o)Uz&iCw<Ju}%+9^;r#boDICTvVHH!EBR8i^Z}$w87ethc0S9KeyZ=rypf{}POr
z0#RA<N5(&WBPNYH9}~bM16{Ka-h#qBiJ6cMiH|fYoiGalx;Aq`7Nt<*+Z@4EfakqE
z^ol^)zIo3snhk>4K$wt-osU`(9C{HQVSqV-&)nY<kKd{LC27ti^T8w1!F1qX<^uL{
zW2xmfCVy6Z9%gIR;dIYijMhQf2Tc^o&OX*bJd5zSSj1k01cx54gQ{!eJx6h`IJo4$
z^hcu*c|a%|S1%F>ysJS*KM-AKn4yXvX=h`zacsZJ_l)-R6u05^$tq#?mR|`#B?xRw
z{Coa3lx0vNf)dyzVIa=75T5PntkT#37T0EQRd0tr({#Uw{<+?#M-@1JRZ)9ATqce&
zj{qoUWxkNjkFK!lx^rjiMX0jh-6FCiKreZ+1=n8$%fgww!a%PMH22ZRF%7_JlR<&%
z{>U}p4*53r4S#c@b}#(4WHK3E`;-vEcRI4_hCO~_oWAe0htOm|K%^M(^dfFv=WXnW
zIO5LzuK`L7?|^iF@E*?Q@zYR_E(Nv-$91rZaN352Z_AP&Ug22}PwvOoeq!$I+_-jh
zhC#sg^)(9JLYJ-`WnZx~J|FG4gIH!8STof7vgWtWh&P@DfUP{hbJXZd0}l%etL6Xn
zTFHI$zjQ<S-+ah)og0~b{>nk%?QITkJ`V3=g5Jme^P}ufBv)15Y4G6)KJrS2Id$!p
zd!OYe=^OvITe9C<4BtRB$7^+01VZRqbI8*c>nqmbo1WK!Sr?)!KtuKZC%b>?+I3vu
zMl?kE5q~))hnGJN_|Z$%t3V8Epc%0Q`^vS@`&w(L#gNQ!YWo<v)^3pA?b|@KXL6(m
z!^dIRbr9}65kkNcCo1XtJrEvEN0;tPV4)(&pkz3ZiZmKViDTXVud3}s?EjCYKsJDJ
zQi`kZjGuqX!g84Pe~LGcDenL8*MI06AGzN(5@Yu|U!7}_{pPY_eZ;@7*XtwSo_S{-
zn=&^QQk54wdUB{><!o0-Vf(X+k%D`#_#{m?jf?~&FP_#<{aJr_`AWTU?C<{Zla4l`
zQn$r`{nO7UM+<YEryk>_C~X0lGg|nJ%n2<WwOexP_e}W;9e-WtV^y(>iQbi@i3(^0
zpRrnaH+IId=h&d>gX<?d+J1>?dL0d`ysB_AKiX(c+WV^<mn=tt;JNl+-2ZtIP-CsA
zbh6E{MvhlT-d3aMqVuUVip;mV)`w-+lHR-BdSHFdOvRaHmu38A>JEokHM@SOnXXjU
zG2RcMS8E1#4|K7tvi?4_%zH2u(M-|P(9l38CXyZn1ne=V$p@ioYHCzfRaFai3*Kx+
z0|bk|0Vl~^^YBHHA+~2+|0E4i$L3$&=z-0SZW}*}V6g1J77~0Srw41K@kc8nb~p&f
zQa-4)i6!*W_~R6T%i_POx%0Wy$Y*T7OobbRyjG8q#pl_6^9sqB9lIf=D9>ZVh8o*O
zJm49ApV9W_CWc?pqE@2O<}u;e*xbQfJm<izwA70OJ~NH#n`w#wyYR2&W)&;;9oN-_
zf%!<aq!;`rGeB0}jty&9hfQurv~Zr7L(hfvKap(MMwdz&?&B5n=!f2Q)rK~5H<Jkb
z#g{gzMy_6jW4j0Uqe(*!p(^W>3I8K$@|99qJ>{K<YO7-Yd+pje4_9hhCR%srRJ|KC
zO;&28e)AS^T?j}$_T#_W;AuCI(%Q)zm3?@%d|<c9@@pq8PXEjj4@%(ZSG$#(g-1yx
zqW9lfCKJ@Ib~QfUjri|E0{@;)iu~cI#?8&W&Wr(<DzCd!&qS|TBCIQ1hof%Yx)r-U
zO&xYX&+6t<ITBPODrd4o-~YFWz%`!mbK^#c8#mr>g!lU;B`2%PPU$D5t44&dc(CrX
zE%QqV3I3;dJni0;(b{oBAV9)`z5wB@?$iIXy|c~Vc{fEi{8wE@XnX5pYUArtAt#k=
z$nOUJI}1mfhyWH6+YKHhn%JJcg&6zueDK^3$3L0YY8Iy}J-pdGLc~T2RcVBh<N0oj
zwf&Hgs3Y8=WPlzK!G<T+S>t&{n`bYyzfF~F6ysYuqGE-dV<snnzaY~>b|`!2zaPd>
z4<(9gB5+4jCSWToD+oaWff#ilzJ@I|7L5qLtZVt2!ZN9L(|?8*?$YgUph|tf<0Li;
zYcFdQOioTtRcp0TAqdW7@exj&dH+|M;{5V*Qm7U!bCenVIX5@<5-9y%RHR}RmOl_~
zFsae%+h+Nn^2SuD_j$sVv=1#E01D^NGYRa2`h+m)^EGp4dvL;)Js;_2Y|ajz;o<aW
zB=B$>VYh;RUt$p=@holtlhi@|Pi@vMwT8}LJ^+KT!^0||)v0@(Uu&l<5=C`&b4x?a
zl{t^(?a=KTG3cS8p$q?c&%MjIB-KjeKS_c_|M%|<MKyhZ=XmrqjDh8??1@_<0U9H|
z&OJ@Mmk2hg2*NMtfMp0YpEz+Mrx$d8|9-xh|DC$Q860SP=s;t5(#BMEcBp8?A|4-q
z30RnzaE-(k3#7=_V};Ow$o}Z*6*R!<^iZ!ky1BWn&2OU%EJ75wF6Uw+=l3PIG;Cf_
zZ<?5Z!|bolYiR{5P|)?v_~lUxsM>og6{pI0W)z(F5>LC3U=Qu?hKPF?ANo))y`4O5
zets;H1lbmsckFvUqxM5BA|IEj7{2FS#jH1vWP+0IZJ!CdJ2pU#;g|^<iHTpA`hB;@
zDuj-GzmJj-Br^Z{yXWT!#(D%E_pX@!0WP^0-y||)4RMhgY{1yjexEg0clTB#D4Cv~
z?tdCz&+-r7k&bu<VeY>|(b?Q!YM?qg-PYWkcMmakW}2TZI5=3UHym;5SWrb#5!aX#
zN@$2`O;TSSC-%SymHvKyVgIJ58SqDsCO%up?tG3Jw*Q3nM)N@KKcQ~+q(fSjb=B3Q
zpF`~*Kc1~zT6+2$Dj;D~*D?2MF@pX?jwQ*!YUNiIT4fl^SBKBn6LH};iIlrJI}G#q
z^YsvKS6jQ6W{j42P<>q;^^h+&J#1(9-*ir&{fl3cwNW&~Q1k_6b=Ozk=uQ9BSKABy
zV`k&wz`LB0lvQWz)^$)HXPBtb#oom^63`~Ziwce*{hK!41DE%1LvTlicN9i$j}*N=
zTXhGSvtxbqa~k%7XMAX&EdMCivBiU-+LHy$71h}#;yf`M72hBM0blzkF!LvG{m05M
z%1bVIVN$Q#o2B}{fp6sa<$EzEy=6vAbfI}~fN{E#)yWn5%i$J$GQ}`lZz}b%U53YW
zwCZh#u7j)4p(S;(Zq*v6uy~9v1pMuv0K#2=vc&Xqk$CBjOVEc-|Ju%=&nkm|K0lpS
z^Zr(&hR}%FH2>78?$Fg@5h>M+r11a0aUbqiS+>9Wd#me!cd-o%i`CixX$+JY{(m<G
zSXhFPF`53zd&A=$h9A=jn!fv{{iYX8ITHN}9U=;nmWM+X4+?w~xWOh+Z^Z<XJDxgq
zhpnaO!qn7M+l5nIMEVbZp2rPPJia^tRgZ-_E3jF@2Oa)_QxGq=n42fjCc6Fe*faMS
zBa@bkeizSLIQIAJzS^+)t<iHeG@N^UHK*TB<Aev|)DMlboIfhu-M)I>G}*m-Q{e1L
zkBahOnvH}S{!&`(=kL}R@4fv}c>Z{7^5rwt5|Qm64qv=0G$A@*!T0v9kCGL}z^T{$
z;@#e_J+$N@D>LKFj~|3JavR)5R&VBgt&|q2XshbJJ2d(6Yr!MM+_Goq9&+Mu2VpLM
zSo&&oy~*aHV2`2itj8H_u(ForZPbbP`|deI&S}zvqa#nRO`e;bJQ@2|cRFxd?Bc?&
z-o^S~KQL*JZ3L-4F>BXUl(0{?f`7f~S@>m*dXvtr<kGiqKgu;0>OT{;HZ>*R->-i}
zPm#5HQXsY{wtu-aBgSMm(W~b1@qJO5In`YB`+*C5DI({tCrp^0sZ#&2F(2ElTU81_
zv%h-oYt@SbA!~B}?YsJAnBS9+x;NDM|5?gE`I=-WH{Y*ViBp<OP2XBBl!#ZV+(^1~
z({0x}F@_rJ^(guFLs^ex{(raig)<i|Wm%soXT-$eu7{?<#!YVBY2J0r@e??don!HP
z(oVLu`bVNiC||CDJ1OypZrnio;w5kQ$Svf^0hpY}<!ySwm4;`|6X|Qwmre)U^zkcJ
z;f1TBE8qM`66@dpySRkBac;}dw~{ygS9pKY*;W)apQgNyXQ7|4+SfO-nmFTEW!h+9
zDFDyYvr|O9DY~~+ZB}PI&wW$jpTYB(Cjb3Nd*)|rOpU`c6-8us%WUEoRoeGLY>Ran
zA&QtHH4i}#TW)wBf?RWs9a~Mrk89c#y>}Mv_P<$CnB1-Vwu63Gym#>#t#z?`avDS0
zub#FNms-5jd-Q$#*%GFg)%|tU2S(R38>RR*gJ7`|3#AnAnxu}Tt%vTON`tT+OQq!2
z>?#GGsl}~jrMlDJ*W_ZaSluX2>brAO<xRJEo64``ZjILy+Bahc;7)1q{3&ViHyYCS
zZ@Qd@EXTR3Q67&Xt>aV_jZL2SqRvkQ)R>{u1&s;1Xa^DP=%+Mg8g?e6ACppRXY9s?
z%I{AJ!04A`#knPKkqbX#>?Ph1zFO~`Fz?9yD!i<Se5-%cbJkBT#dPg*NbBCqgFClF
zNHs{9_|%^iiPMy-Z#v3`H_c1DYveseYJKMWJ1cbpwI2~MHw;{qpxfHMbGd%b9XYQ&
zj89_QRCALXO5b!$EA@_Dp1bwKyGE{CR`EgEv)(7}T~AQX7nI&Fa<`u=L-y%APaykt
ze;58bG4PjOk)~46`!1)Ic`j|~`0XwUiMx3M^CO|#@^+@TQ^rTx8e%}<AE7s0zYh=k
zTs&{-WRWc5@i6Y1(fO);ix<C>J|xNBl{TJoO_aKwrF&vg(9z;^tM4ZxQ4)F+0j)CQ
zrLTk>_c;1x`Tt<(D*T#kqo|00(t^^VAR!1N1SBV*ARQ{*9g?GK(nupnrwGVLH%Lrs
z)adTm2<g$W?fv)*w%_wU_jm6(_dMs6lJq=XKsEgF!Av}nW!(!6d7)v25g?!M<9fcg
z^T3KZ>+8PdT@lE2u=ZU<#42tjxivkSg|Q}(K39aZAzN$ue_6SjN{kMz;m+9Mq6e*R
zPb#63viv2`A<yJr(C!OI;>AysC-0SaXDyT1-oHvnb~;%X{BZ3M>&BR%mb&PJ(6zMA
z3h;T&AF%n&>ba@kM#vKi%iABc+8t+d6KU|kT5QN7__C)3)RuVF(v(uC-R7zuJ+~EJ
zm*+d+WUDsxO{~d3v(b0xz*gqGCu3BK`w)Kdks=${lhcKiIT!twmLi<@5Z&`H!6oX>
zMaONIoj6S+$Syx>CT1g!`pBPt6LtvgU5>1!*l~6~TIy^A-Om~Bh#g3Zy++4gj?>wZ
z%gX<IWAtV{7*Qwt8$tg5eK9MK5NBbwcIvuDM(o6EXf#M!J23X({qvw&K6Mc#;ya}d
zCka}RnSG?iDbFYuv3c(a2T{IeW3r2~>~}En%j<1y;%MWu0g7-Vek%77+V|gNxTS1U
z*p%^RNk-J<I`xyYGr}-+)Gwug{U<0c!ehEdTl_>W6v7IXi&h-52S-=cNN+|ODdQ3_
zSZb+@$btSs)ZmOM|8|}LU(5x(_KC@9<ja0(XDO4$lx#~JC%SdshP5bQbX8(k#oW<g
z<a&ijT=CtWEtRQN%sKmp4{J22w!`u^2W(B8hVSD{d6(_Hj)IXv;XxuhB4|kG=8$n-
zVM`EwCS%CXZApIB^9ReGvWVp_?j&xdysuNN(&PPau0CFVIrA1MZq&^Dhs@gGjxCMF
zwwKfT?b*~?mwzEvbg_JqRwA@Y?L_Ct&^<Oy8BEUaJ)?N6$I@q%pVlC!;q(G}_nL+^
zeB{p?j~TcyET*?;Q148D&~sYVGKjX1Ap~8`LueekZCp<Ar+H}r^DCy$GflCuEP&?^
zPwWI=>X1uAPOQ1Hb&58QFyX+A)P+~$t^a!_!x^5~O{ZQXP`o7z*9PPX2bEJ%fn7m8
zEt(<LDJL9l6rZ|xP)=bm-$<h_(u<_q@-!NadHu-pk);6%^0Ny6UM}WgJ*xL-26~>Z
za7)lME3`QK&1-k)M;6QYH(XrG@Ddt$$;>O0ga$bXjbq||sWyX<KD)KLt!id`v4=@y
zP0`EHMY^oaiEE~Ee1?z6q?_rv#f;ngv$&zKy@!=sud!3-94$_Yw!`xKry2VIo&9Yf
z*~hVuN1<hBMAARp`iq#Y^60yrw601MYC1x31$@M5hDd=isu<pi^)KxIK*fcyAo9DU
z(y4@(Qwe4>lu6BU(#=auVI4ZVx+-rRV_eVB3u)<!CR;8hoR>qMOFFd5SQ14=%Lji-
zz3~x~L6uRRKZv&sL{MFsb)o3rhF2vEu11xQ;2c{47N%m3@LZl<Ee_wTUy8)1snLcL
zpVg(QyZ+QBds-?Ii0i!1Z<nCUdAg3+^lAUjefHhoMQ~Q7Yutn-D%E}KUr{gu-8HTc
z_kTt5bHvr(o-$KKZcZWp$94m8;L8p1_nlL5X7xY9w7}^f1wf83iQj#!0q=EbzJ6NU
zVNE`p?Jbt+zFjB?CMFBVD;Es)kvgwTyi6Us9kj_!qgKjEzZ0ZBW1)?{3DKMhE)T3L
z70y7|rYFBXO)vhkOW|4gaxY?pBdVFVxfKW~I_?fuDLJ{30L!}cfybwtNh+Q~<sGk?
zYPoC|9CgWVe;dqM(CsxpG%Apu)vlUIAQTXL)cGToz8?GLOn}qK;ZgBnqCURiOn4e2
zb^v@o5V3fhE`R3Mo`6GX8>gcd`y9OlDUP0%@eA!ehUpI<?BC*eOd==L{-N9s$Z$Jf
zj(hJp-j>x5;LYsj{HQPcV^;tA>R6)9tJmtR#%r%K>KzeQUh=Xj$<wj7swnQ(aL`2w
zq91lQLU1|#`FQ-2Ocr8)?^?BhIdZI{JH4tFbn;K*os2PEq|MgWaERJ-$6Bi0`{Io$
z5AsX7pTyN)76&YN<gO`Y4O7^vvwQ$kMnz!ErbrskNXW{`vZD1$o*uV_3|J{Go3AXW
zs;jkhqV{6^+$#sjWdn7%Gll<w?I!u?>u<~7&Rhro9wvL|SqXM<$7l`TRPs;W>F;Fi
z`iT_-BN87fKWS3?AyabY@%JH($zs)6?mNXHys6J1d3MLrEry$d%nuhL=N-MY#%js-
z07_Fblu7?m1W|OPY^F;5?&8r+$9MhaiPw=qQkim<YIrG{%K4?Z%YqTr0#M2FROA5N
zgzl(BOg3nmurLkxr^JuN$wMYk)H6R*usDL$(I@C>LRIue2X{+*#Y&pn-$)T6r|gE$
z(9+}c`WlIuP79-a2jV@ET|4`fBq+hbyALhZ(*g_D&2AZ93vi^tODE39LlbFf+N}Rt
zTmK;?!wyozKNvcL3gRZ6TkQ&sr|`6(4K%7H&WIQO=6wbAqm;^&2E=gzS#~iM1cJ2}
z4#U=@`OA%)5fjexnpxBlnNlYu{32sG!)Qzqa}(<0>O14SZis!W6Ka*;;~7`VmjUs}
z*<nsVb-}~$!R|!YfP|b$m4w*CY4K#OVFMVyKU)(~dh=<z$MYnq-0d28X*-=rG!B=C
zzno`kU+Eiy!QMrM<x^a|(;a$O-*lx3;uJ7RwzXv9pB}Z0&6W?0dB2kIF+KFQ*2~;!
zNq_?XJof1GM!F{43+!$Ko+cc(;nIQ^wmVZ{cb-Ip7wiiUtmf4vop`f8D4gAjOumer
zuD>N`nbg@-%VHj~^wF>mmKxcyW#qw*eOPn`S>$R`vi0d_4#M4A>0lYLEH?y}#_Ef|
zmS@sppT|5Z_YQn1bCx3m7ll$SoNx^`8kf`3H8pLy<zQDo_hW)P435h|4;74`L$hY>
z`x*Q6q;!<!4}BJxZr{!^>wgv|oc*rl#Bu<FGX3W+HQ@OAU_Q+V-{-*012rDGsd*Wu
zCKNwA#x=DWG2V#P<Z3QTcDr-H5Ng*-6U;5=%$Esd+RY%*6}IIc&JL^n0-mxFWYJ`p
z2q93svC#6b4YHEad~f`Ffb_ZluCgDc!=#)VmBJ&%;rF}B1}_+dEF$pv7Aw61U0WJ1
zwCcyz{WJSvD_Men6KYbiMtd5j;^Ct#@yaGmiN3BQ6v!lR2ev)qnb5O;X<3E8S+!zY
z>vkp!|MA}MPZWPUBxTgvMmftr+%=7Rfbp<r2-?jX+!QeSXE_wGUZMiI$|@f(!~e~0
z@x}Q*q?6KUFUDM^6H{PISm`mo?J?e`$?v%wj+?nff81j;a?FY;JD3z4V~ng6y_t6~
zpT&zE!aEy^CHjoumFPl4vVMOq(V{vgIn|yQy}QoxOqJ7p>wd@zCwv6;-8Y!bEoJRZ
zrh|l!jL5m1s=`I#&@E7VEVMmNZZy<dePZQ<aaq<dlLvoZaEo2w8-rc!m|VNde+)4T
zUE`>Yxc_3^taeqvZ^5$Qv2@-!#iZ8bz2xS+pklM<r_(Kn)}7FFZ;#Jfw56xUPZ6}~
zE+VrCd1<HWkU-NPGT%r`ks(ln@~Pne>fh`tNpWOm6F7EChdIqTQ3Vg81a@bp%$rwQ
z&fT~BuzU?rc|)gX*vh>6=SlM3LNv32I!`$?toSy&{p_e5U$RbFMY)Yq8adfuoSt5%
zgEDHW7}L{))$3z|T$)*D{YCL~O$aIvi@%5sYGloK^dEq<F5I#AT4XeqEwjs56kBbL
z=TBY;Jd0obwudi?CLzEF;is8)j^g_ts{ldydD|uny%$f7wlD5X0EBF$yRySO+QPhy
zgvbJ&G}a6NmFAUEh`(=~$@pROJZ_?a4D9WU@7Flw1oJrftKpPIACoG?DI7#2n5zrw
z(qLXhf24&L+A=z3=Uip<0ZUk_%JVO4jcUyk?fCdPt*$MWB}N@HP2sQsTiBnZ+T5Yh
zwQZFIEz?5<wV;R5pm_785*7iI*WUx3E*^Zc8CQI=6j>(X@7BKg*~oGA-Et6~uEdN*
ze$GTALE561VAIOb(ysMZ^pIu}|NbH?84C19w<Iz@dT6@z48Jt#3q1)x;Xoh^nMW1z
zUrN;aRf~<G={6K3HZ2`*v!Fq73v04svD^;0LAG$0fCt5egK;mSe4WCWSlhc84tbn#
z$|RN$>eF(+8{?Z>@?8yoW=J*7g4b-!6%N@~6L^3>S7%Ld@x>(E7$hApEhj1+p(_TC
z%-PF@M{u#qZd`&2V5j1{e?5#2mU{ed?|D@iN92IYyPVxtwj!&&hurrA?K!+$jH8zq
z@3N>CS-Kn<bvGK)aPex6ES-vgQ9{JVtX&x*EkniF-XL^vNq(+r?@?Q`^GlPJMh3}7
zndk#U9rHe##;o_P%Fc^5YV~E+=2-SSlYH87qiYGsVC$IXV6#a(foc2Y{l-ee2;(dr
zizB#gwF|N?rEe3vKVQ$#DkeQ}q10Gb1Kmr0Q}54=JH4R^I2j=ixLxP{fOBJYP&Tp5
z+b`&V{WY?tA^u0g00|h=-w!<@<=dqnCERUVpKAq~AN~0}*BTtyh}OE8xAd`|1Xn}4
zIosNBzF+*d@B)q(kg?K>bks+31HLz3&r0jvpr1^#-`?iQJ(BfgXepW0nWX=yE$=*u
zR|r(87lC1vofr_*TjNL)pDb=@{1YP^1art%Tc!6FcdD606FfnJS8MZUMJaDy;)l4|
z8*ISCQ}J3`+>G%ML=Woi+R_rl<o8FP<g7jTcKgb5EMHcD(E<Nd(rJF7JK(p-t4O^r
zw7M=SGm87NDsY$O?6R_ZlNS=a7kPYGSwSb-A9DN3OmvB%@Q?`(QP3(k6SR1NP#}*u
zKQ5@Ix4>&HsFZT%=JZv2XENx2)*iFNsy}G55~H!QNX$jVak={KapAt;moVu=Ci9^g
z8gKHZ&gvvDI%m$t`F^KR!nmI1N4Wk`KWg2DRL!&_i^>|tN6&+sLuctcBkv^XV_go_
z<TLV2=7A*y<q-y*-$$**V{a8m!_T9%su?#Q`a?vrUo90tnD9+CSP|dId$vSQT`zj-
znwR_&(HpGX8b9y6?_}-&5F3}5REjeHo%n~Q6YnRr3H5E4bTWhAGKeodW{=nvJRxt`
zuSyhL>6d9qmf3K7!WsWQkmr=BQ^M)>K~5X}c3JpniOX*#l4p)Lge;aRt}j2oMCCn^
zq3ZW*XafGewVcGavmjp{{?igE@;L#~O|9uA_`v!l0^sd-E6P~f3mr4;#^hWZx~Nfl
zhWTo5s?PL5^!)`{EAzr;wtP;YL=-E`;hmufbL-;Zq9WmPz5_ESYjVb+$F2<xy;09Y
z07|J_($m3}&1OI8L`>7h*vL7Z_#C<XVRdkL13rRp*v!cpj}f9cr`OUo(M9e>V^iH!
zACX+dNJ!BbmRv#j4a{W0+#Ka674Fr@+S*vv!!1lgN^e0Z$*7B~q&^Cst@ilZ5H4NL
zZL$2lBGG)`#ihy(1xu6hmdV`HGo3zs_vJ$(%Ej4c*mk^oL6yi*g>g=QRf&FN54}>Z
z*a^Q-pIcZO*o*($|9H#FCR9tF>FU)~jclQS#h96oJ#&ruLAM<b7!lyht2{zb_<D$>
zf^70Gohq8hb|iS7nY^0Lnd^3F0!~%<1tB<?*=YHX?UvW4oI*(bu&d$lGq)Oj+yY5C
zhc>4+5$AWO2lJeo8>-o%^n3E$1+XN6PcQBQI4{c5ZYaoKizGrDK-Wb1O|hN^Mg4fi
zKV$tqWimbM6X|JI2AzKI$mBlRcK3XCecJq={u1o#wWy=duBNWzeSUA|zhIpud@f;%
z9g!dbI{4o;=dYde_!RBL)NOrT!_Kd7r)jdp_!Trpg^U>JC@?2%GJcNKpBk0lNS5ZG
z9_RmFU$2ZPSs^);Ftrmwt`|)H9Ws7zV&dxN6^Vbown%YNN9YsgOA7g;;<6szaM{^4
zc;9)1Sk>baPFIm@ktn2ND^Tg^LFeKcS)xr=lyiT)Qx;KkDX_Bg-Z)y`c+=}uwluBR
zR^=GO>wpL)*V%;d$ddaeE2+n4jlq$TRD|uh8=ck^&8>S^*#<i;Y9;8zKF4K+kI>~E
z{L<MoVGSt^<xN(+;HqI<_NST0({4LoRzjy^XyvV?GyDAIUFi6#io9pvZjS_^g!M8F
z*=1glwkWlZuXaRTr?aRb&Uft#J<BQn<L{ZM+#YKhMY%uV`5vkzi@mByF@5+Hqj>MV
zKffp7m|x>X5XgMc^1a2-{>@(7TaQl)gO}>ov|T8@Vju8=A${md^qX4S%D)TwWfAAi
zbc$u!quI~#m_OqXE!3{1P82%QiT8*WI&}V#HRzBvsN==WJoPIgF7!m&wVF)X@_sRa
zQME6R<~NCb6<?<bID;#2mphp|9^DhWD2U$P6>*?TA+-B!qFFOt`ol=)Z3*@5c`S(L
zY82N{e!wbcx@@$?O;mGIKrc`T(~@XSY@lJ`DN_1${Hd&n?g_CE_m8pGKV9P!G#^8Y
z^Iur9LbAel^3`<GoI9n3X~wpU*EY8P-m;Q4z+)<eJ^Xlg>_!cKaJ`Ne9DCV!EYn`F
z?m2BhYu7-`GIWcQ-u1{GJ2k?;l+@#;V=q76g7<x`gtD`T3%1J8#KjrbU{<bO8zy8o
z%Zm<=Tc(`_ejB}%kg1EuAxJ;UcPAw^XNe3f_e&NkFlWqi66AB*v++=~Yk6_fSbsB*
z1tGPPHsnWSo0_@k7i)i+rQ}hVV|6>~s<7DL;z6!4Hn@mNHyuzora#!MzS2N{hNtw{
zZa6{K)cTd*5E9jT$g<Arz*$fj(IB#SL+*U^GEUL<+R8(uHy^hBS?b2DJ~mvUC<43r
zFfIaHVh(GzU#qxthbpxVcy_1aG&@9O-%OBZAYs66jB8_UXnT(I>wU?F2)9oM>Nvj(
zBBx-i!&o!=kHP$<Upv=<1D#j7er)PQQ8^iHorxdyC4#h<8*W)7r^GlVI|K>ed2KN8
zL0)>yzDE>UTUU60-MoN%z>f?+E++k!E5E~nh^A*Wg73O4WklhE$7LlU`ce-2(eG`=
zvE`nM0?XK+#1`Se@Q7P7XRFcEe$Aqv;I+D)hAD6bWA@IZd>@@=lH0xWX%$Var;&{_
zPI`?#<ax_v3?|+AUcs#k-){ara1mSKy>)tl)K|5l6B=m`$xZF;SIYhx4LQ}yk|0u}
zKh;e0V}*PTN8nj%D#O@Tihkr$q%95ctJ*NW4b0bgLmPDhkwpN?UYFh&|NMp%!I{?b
z&B2g3<Qzv50*_j+Dk2&BHNa(D_bJzEnn%nf_NPKQ>96XUlF=RRl2Pusx4JEKysB@Z
zmwS2*QmHj#5rKsvlAd}RS<_J+wx+RyKHc*pYFo~Xd4}OQg&AvRVSSKOCv4S^klc&h
zwHA^RSTy;_&^1yru%77FNaMoKo%$lb$i#<UN+25TdmcEcrGNGWJIXMle`Mh|_d<`f
z6`w&&ta);KFN{p<%whZ_;ORl0>QXp`R&!_P4Bk58JX-j2-Jx4<SLn7r0JNp5@%5aW
zYiHR9*f0q@(wyL*cbF#J{VuswZ*qKb@fEg0Z<%fMk7B7{LVSFC{`j(9vz0hcCc>|(
zaMdlOYd$k*AlP$=VMc>d^GEX6m6<g3S}4idJ0QPg^t*(*w9pvXw*@_43Z$|tvPmpa
z(^Ad;ELwm0q?k|Ea%BiH>BWAu<k&BBPNU3!DC^lt!aZvI4U=-KmEtZIZFkyB3$K>3
ziu;>gctX0_#)^cl<A#in$m>5>%K2+?Ltt;=CjuTR?}Q0^>Sk=ERCB``yFLt<wz2=C
zDu0=yE%);I2~*9htIL{>mMPzRLutL8T@&ANCIdK&*iSWiB7_T}<d$^2IZpf5GOcb0
z$}>+;s}ealsPu~X8O2Qy*TAaEmHrKsZxo_Cp*y+F_*?O(MuUU&t;wj=UuGKlVF3^)
z>2=HA_+$fh#x})sJEqgK$Dtft-nmt@`$cy1O$N?|YQ)TDWgLczhKlcmW_gDfLU|F`
z#h9(11y-P~k#Ag3+8v|!q||@5qsN_zS13lgG>Ety^>~@9-AH_21jb4`;t1Qj@NGod
z>4<txKia0qVB)?Vdr)JH3v6C#HMdv*yk!}VE{gtmNUl8?5e?|8Fod<!l*@t$nL75F
z;zqMQq@nMja8Fpgg2^FKvmXb6qKrwxw$5l@?D4$^>gi!$(J~*RvcnWvzx74}217<$
z&z#Kp+><RgTQ4UdCf@_qES~)LJzxN42W*a95-#xQ4AhJq*1m*$>v7^5GR~-DpF6#0
z<FbnHTGHHyb>5Mba>}K8oPba#bSv+878?NzMZtnGw*)|{{olWSGixXpMO>pgs!Cv?
zr5aj$`K!6i`B>}&j>Qk3iI3^=Yq!@ElQTc7jy$uaYH5)JgDU|qq~;?D8`i4Qw!hJ_
zr4Q+`T|53tXzqlCU9WB$bVA7)ln0=VYe!!;MptLJUqC53f3!r#RS+=!D+5k(UiYTy
zb_V}O_r?h9#LFV+6$Xqx&vodC=VE+e_4W-DPo0T%+m838%X1phO*CUJ&K$rss!9m0
zn*__e>`yQFG;n1*lIdx)Z}f}HW<AFkh&X>XQyCi385G8D`NXog;VG7AFVo!1wMNry
zVClmYO`EJVPO^4XfIK@dWVc7}Q+D*b_iWdupt;#QwKK?6iIp{M#Ny_3LjZmsjgv^d
z02BN`@EVeYk0vA?$xEcOdI@MGYI8P6=C>oRR(=)s<cV~W4nXJGLtxXU$6^aWv1fwC
zS4@>ypCGQP7xn_Wy^0q59fqAuE1Q~#WV}u;8)0@(T&U}!FVU#zoX(T!>H~Zq@hJMw
z5X>A;E!*4zynw@ua0qmO495>cf)4U#8yo7v@GZ~Es-(&5S?^`a^jxTESp2%yoSP#H
zO;hhpS2QhRBBK!Vh6`Na`$~vQZv5-Btu6MWk86R9eQu&xVQGtL2_DCtqpBWb{v9p(
zw#5gjD51~K<*T8T*?dOp0)q_5<%VNXpROUt*B%J*rQaaJ3ZmDD_lOA6uE07jAw|ir
zyXBFZ9->qnV1+`A2kGF52Fo{7;!9(P+u65}dE(dGg4?vGnlW8tY+HO!1c9L^NF`kP
zN4v`+!6)XxV+E)wjl%|L1y9@7$HYI0)$hMvse31;G^3e6#B1*!-zChno{Gm^!0fSw
zJ5^Bn*`B3wl~EPk1?mVxbOE|1G5ydw``^-h)nWbPBu2klo9a&6JRCk42w?K}mVR$o
z9mekXkb>Eft4Qz=fMMhZz?KyJpNEmBqq2bmGKieJ`1$Xcr<vfjBX#(Ut9j7rOz@93
zXi{HWw6q~>XhR&?m-69FVLSRE5gz<&e>A;bceDNa&RSTB&ZrsRxzKK%O+lF#2}{5Z
z%Sz<h5r!8>C}vst->G|KT^9Q*zDTV+W`U8#;6JZZyq>v}t>m?S=YzcB-}AU|wfy4#
znlV^oC5Wn`h%ph$&(d2`T(F@krA&>#1H@b!Irw|$w|H=zMm<9^SueoW_`aANSE=W*
zrM*|RvyJE0Yod*uPz=kj=y}m)8`VPAn>+w5MgjfZH3Ci5SC%+J>4U}iYv9=oGy$o-
zJk)H!I*jv=$K>6xUNJhK%-ZYW|JK*l*F&!Dl?e*Cl8;vDcP=Fx8KwJA%tjfkCtOcf
zMF-)Sf$IR96VS>nBaV%>WgR%502b%xYx&344`x}dr7xL~bAJ2twM<@(mDM%1CAcIo
zM3(QmrLOW)vnIldY!6lb7kQS%_daVpn|A}yIWc{xXbz}{fKSyb6lNF>+rapqvPF~Q
z9ln)0yod-gOFnU?+<IRMZ|SAqmORqPpFvEg`!!N&et-SzixPa?=U=~0JCf_Xw&#|3
z8uLB*qn5kD$tvGItn}S6#g?$}EQqw^oig+9VNwgyhwmUkVQwSI>8I`=!^j_~*z*x5
zPbgsjRk8sWp?Vhr2YIBp6j)BqIBr%ZTBSkqnA*|otV-mh$=f_|jTE$HAaPL<FUs%W
zt+Bqi^gyf~ffZq($bd3>Sel6sQW=`~(d96g>fNj@)5$Dli>C%-g$=fH{Wke72$NZM
zdg=b=p+V`M$*|!MD@3QDA`?C*-+DRUWEk0j(R<+ZkP%AsCl7BSw=XwM=kJ2iBY$3w
zw)1lQhQ<JLcWpm(Q~OSzrz<D$S;cb9=$LBZ^vGPg{b>xV(FfRX(Q7C*G|u1#O3{9w
zp`F?&XPFYk81f#1T<c?)7??=)dV4GOreBao3siP3DRJk7Bw_+j?j`w|KC~|@tIR~K
zUIUk=51#n=u^(oo;j#xKWQfrfAPlG5W_gn6sP$y=;gmJO08_PATivTC5|sSd&>Au>
z|3dP5<-Erur(8qmHI)MwL7WB`RPsa@TZ$#Yfy(s<=$h3!5_*douC@&|1dIQ;!-n{9
zsQu@a^5b<h6U2ruG|=VFX9tIb>;M-Fj5aP#A2DycHnL8msV%ldwok-T2%+|EHn<7M
zd3hwZ8+xOQiJ9~+ta4}Z#1R5MbLrz0FFh1+yNnq9kF}Nc9MA@LQrtO@6AY=Jf9<=!
z+Q#5zLl~VZqdI`pBN9IH*7Z$<5nVg3^+=${iI8?;hGVI)M8O+E=y26R-sN`Z?fF#8
z-|-}o@H#3$>PkXlf%o_VC*n$Hjwd<vT%dVT?)~jp^i66Yr-w<m5ArSU67lV=r{Dot
z{62v#M$1nk`-1hw0j%kKIT-%)vmN5mdli#=dABsc&AwSFzMAj)ci1tyDNNM#)mH0+
z!th%I;-9GNoJ|C<KyH4`F5JJm1+0VV7;{bIIv5t9%m=VbUv{YfNfR~ZJ(;Q@E%J94
zVUJ!&fQ(5b<}(8SQ&;2pN_9hhxAduPz5tk6aD1H3+J>M=q32N!6Xe8#WzAmT!%PS)
zQr%?(*t;@K=^&DRdU}nh)6#9GfY?mn6zQ)?fT~$cCM+(FIVJjzF-W~$^*dRO3Is_%
zaulYROLtr^wcVeYpCD>Jrr#;^0UY3hT#CBo22`URixD$YozOek!Tay9nQ{Hyc{SY6
zh@wM-gnHTjE6o=<0eKAyRNpm@-j3}k(TVsx^o;3P?ajlv-M`K3tK)o(G!_-*PY1Iz
zO}_^f3W(78mh=e|JyaNYDTp(zEo<6b#g7XQfj)%1JsB0g0gBrMPALAwL7S)IwlLLM
z6WWv^oZ_QO`!N<w)T@e*a4yRmRg{>|&_O8jPoF)oxS6ef_7JBr3A75~NN~yn25{(Z
z;xV2Hq6HIxSzG@%;5hCz@b~Z4n%&;lx#JI*<K@fr4Keyu9Z^K+mH!zn4=o3V>*jgl
zB7uL!7)Q<_#VL$!0jhB=b<?Jgg3&^3RFg61Aut{GJ}X7uB~?rYP!q3J+u^@`E>aJu
zT@~5$dcn@TYfD_JDg6|44Q_1W<WADAVajc*`2bWP7zgFOJvdxF)m>g2#UTUz#$EI%
zJmpm1lJaLSkQl8-rqqo-_j0Fm{QLMUE*28~Y-RhwR>V(ruar?+%Y)b!^!o3YOW)XT
zPB3+rn(l0Ly<BYlumEqDQy39LKkWHpKkTKLD*V<E(~FV6xd;)U1`H=sXf3|G?&5hq
zo6YlFSG5O*SaR`5NDTi~0ntE2LxP~~QpC__Efh>)0XV?)&Lx}yi0S%~vw}0HI$J~!
z>Ls4*=V%|UU01Rv@#+uJBwdw!Rc39MbX$py0>|D)IY#M3yaLjqc)}YS^Hn^*jh+h8
zV)y`VYT&T8r9M)>Jm>PH$$7g9D+1ju;yYN4bK4pV=CTJu0P_8nlNVoJeL{`eu8DVD
zh%jtc+-Ss#t)gJz*dU69vPN_C_xG;#oRAY*Uw{wLXfMgDV|0#7&%c)~+3;w#hTvOC
z?psWXYdV=l&K`WOYe5dwC!PV~3C|O*CT9bhOMmk&#Qh{iONWZhCj&KhCsj*)`>XwV
zsu29ECG#n)EBz^>dx#Y&B_>V+&4ux8OW(7r2UbsJF&uze2sp1DrhN>h0i@0xaSU;S
zeY72>=C+`|AYJec_eODV_dsuO&6HXP>jaOH;6OL6H^-|m+C3GqaQ16QGYuI7IL=yy
z1>QyAEw99^{ik`t7V`+GH{RjN84gNU!rk_pUfspyUh%phkDeo8j7O&xVrlp}ppEU+
zoVkVrm+D%YqcSHuKv`ed?zc-WLeb^K`;!?57HWyJxmcC9D_d|tx!9wCGE73B{}^mB
z%h1fB|4p`Zw-G*w9etJ@g(m6zElVz?;1a-E<SDczj^RVA53Xp369`WB1D~H5ohgne
zQ9NNLksla;U=ExCNA%aW;b@uS4;LXtc8qk?VIOW&scwj<5SHO10yd}z`RyKm27pnk
z9Ea|<Td|9~<pI+9YfL#1X@<P%8yjBQ*3L+`JdwJCSg)D(pQ5S~G3RV|#&Gx$PFq9&
zT5XH&2wGQQqZbjGUGU`SsD9(ds>_z8gmrXon_t|wxy;O3=6lBqHK=kqOfgl4l$J;l
zjvpQiY%Hg=R%;Y#!ztJSX%u+$?$m$`ba}zuePV;2tKmc&XEYu30ob*`brm~w0T9&!
z7FCuhkzn^%379sY_{bfiE&dJc#eW=+%3J%vxBLUpfzY2gb_kax4g`mWceKVBgnL;z
zDd5bh6Mep8p!45;O=>*n5y-}^n#{ar8y)+{WYt~<IYmL@Y{-88H9qtABp^!6V+tup
zns7@ngiI9HpiFB@s?_EHVsYIz@wA5A>?}H^!!96FsgX>)K#1kx<Dm7_FwKzhXII<r
zehNAYE%Y?y^0`<{6Q4BI;;>h&zVTj{U+Wr@v$IY&NXm);R_?g&QDaB9or4mqKwfy~
zaw$Nl1q^1!i6q~3_U!EAlXyP8)3Z~n0l&E)eTguWZW&AFU)IjWX}+SCRguKq5#(^a
zqeniLz~Y3<++rf_DIUemQ#Ait6NUtIsA!9pK8vig2mDc2c|HJW7ktfca5AesD`n?W
z50!MZ2O?MW1{R=vd+Eb-?VeA+*aM9^@yAoUv6yML_`Uu-ao$l+uc@{1BH_6b!|}wH
zdljDz(}79cfn94c8j9{vZF3R&{Xm1cTw&hK;mXSB4f?*TH7&n3g`9k$fKT<0a!XPJ
z+vqYT;W5EkN+j|^v@spbM>X9-l;kagCk|Hqg(cD?3JJ#^9XLW=@+-~69p2|alc8_s
zBbVS{6eqwa)5+bby<0~QswIG;v3~b&GAS~z7W*%;u+1YiTzsrb2tWcV{{^f@^KziS
z1BNa@caJGYV7<>Ol0P@@hHvGO_}s!qgYM<pd0>@1a)UrjF{d1P&cGF27@OtyqXa=M
zb>>_PCnYEE-p&gM?bDGSlyH0SM7<L3eSfNS6|J!xIr1~SK2p5t$}kGex|Jk1>U?t9
z76&7AY40#N0ZPrOud7}_xn^G6A!4zPq4B#~6r_+v{WAFuI8XzBVz2E=1O9=o_x?$i
zBlaojS-zr>jGIhzvDA)~bPU0=p~GsJr`jG4NT0$eaW7FF6l=tPOebJ5{Fc7JfmPP<
zRiY2LcBdoqsbr|J0wulroi;QHYU~wthSznpy20!RX_2`V!wCDDL?l67PP6n{0ytm2
za#F(qqgNChz|)=LLsS0v?uL=oQ`#V-*zu?v_`pnt;54xds~>YD>?T@Nih<Zaj)^ht
z0`%@md0_Qfp92AqF=_kkI)FGAv$!f?(a(%J`V#y2zr}0G%&Vj4Kr083iIGkh=2hQv
zgThzgGla?)rv+jx_yz<(oR9U8^47B06=N)Bpy#7+D6Hfg-~l%}(EwPlTK}2b+r;z>
zEC4mM15mLEb*aewG1rs3)qHF<*7S5X2_~||RgI&@P?tQvO{XA^zexhD)`0;)`ovTL
z7yKP+**JbQYrV1tWv5x%a5*C@_-*T?na9zgL<ONN<@#q)`S<mi=p$TRKLFGLj2N*(
z*qc<$?D7@>9m>P4s&_hOB8TyBlTzLBkUO(EJw7c9{RVAWKP=nMfWt0WSsaFTRWWzE
zNbh!ExH)u>7*46tzE*WjjU#bB$-@G7IeY^^+LmXtc59MIKk-Qk(B#1hO}Imi&L_a6
zy`!OKmJEej$ep8h?cI-Kp~GNQ)zzL0dQ{|OR6_LewPK(eX6WLqzC%(|1t;<^7=c$(
zy+UI7>S2;4ETr}KUL5+_r<(`Q8>ZqG$~BR0+FdB`59j}q$G<D$&hOapBp@81EH{NQ
z<vh(AGktPOi-|irHfCNj>twq6>)5JksQ)#8An4>x&R}<c-l$VO6o0ALUpBlB1x&4`
zA+=rbXOwWE>sxX-3+Ehz<GaJo#N_ltiIIS_#}od1X@smN&Upav0zejfc1^1bU@YL6
zux^0#Gpd{q(-o~s#I24@{J==7p#h33PzHSIAIB$Ct9E@H70Z<ht_xUXV?$7ECkWv#
z1#!Ufs}KS}@>LuJ71V2LyWs*!hB|L}STS<|>>b>1*f7xE!(5t^QN#GmnDTQyR^n8#
zE0V(v>94Uw`WA-ON~%kfp`~JS&87jS9^+{ZU!?pP>;MG7_X?K{s~f=0R+&D(HsyH%
z#Dh-(1)MXv^48zfH^;l8XXh#G&qdQc!{|oOh2}eyw>;!}{7r5BkxZw|i@Du@e-L7j
z=sq_t2>FjP;ZtZsw&C;Zm`k^EF#;eYrgbbA%qn7Er<v2l2vJ3wqE>m@l}8S4M4*LO
z*!cpod8VEmzagszGRdPidEs<ND=xs4JH(Ey+TDZ!h|o#ZP)jG2nnmwq`)(RjbiAwG
z?f6PAAm^^xDB#>X)?BV)YRvoxz*IlPUDz)7Nl)_L;c)$c&Kwk5*^vs6PHhq6P#H9?
zcXdv<7c_(S^N`jOpN-|=infq*z(*SQ)7zg5F;g{Grn|(~ViUNuaon(z@2bWA2F2B4
z)HR|z9jJN%q>W{<#^o*K!t4S5Mytw6`a3!)a?PbGjQRS=-<mI|Lnof-Fs6uQe5*HL
zJ_~a$#QFro{3Tn^<M<hqMD-(GLY^AGdEyh+&YOgw49^xHz!r7aX~o`!Kf+0lXNpmL
zE)~|4$DM9*{+(Hx?=f{kxtsug6*V~B6ETYRi(mBsuO|X%Dz`Zpx>+@f1$+h}uMJM_
zhTo$R356I%)@<hu&{pOn+)v!C2viWP)ziUv0os602$};`&*CuhU9mXRHdY)nl<X9`
zyB%U50?4AWmEO)IDfP1hOD>oJn6<a*9udAs3$JFfa%yR?6{EU8dD1F*H0<NpcTfkt
z=x%5dY)<jwB#UZ%<*M@C@uel=!&~oeLp^l=m-MXp#JgJ$H3t0sefrTkot!x4dSC@p
z+Pu(6=)!!{r-ozs4$Ddc2WDjDP?4VjAeU>#N<jOoZEgn%v815s#PTPgG!E0N$GGP&
zNusDn#uyKhng8ZEOI&U7Py0kHbfGs%;`i%+G}vEqpw(6d;#-a7PC6TuX--oNrlb>&
zP#S@KN`|co85CTLKvxzQT`<hEyE6r_%xj==y1LP#(&5lA88!|~u5x>F;uy-=?(k&V
zm0=w{Acn+C0>1fN5MexpfI8Vv&EKF^NziHpfSUsx?Hw;-HGbmC24%kPPrCJEoY_t~
zPUTmcLZFkCE2)=6|J9`x3UQ>%NJV{N-?<GlCi*D?rgB1Q=PAut<SnsqIsr|qDFZGP
z-?16;s*2v&0z=F)E(&(5b_eVYgX2$tNZ1x3RP2p`p4m{bb#st3wZ9i%C|qf{`Pr8h
z{Vj+Bt~}I(L65X4q6?1qt%TzKhwc-i!`^CH9nIE-Q<ikl80g!jl0mp^2NxLy$y=7|
zrmf-k*1H~`HakLmA+eZntYgnD>p34zYy-wvg}z|aSS$I{uv0Pj)!2~LGKU8aeOips
z;K=xZ>%#r`iPIg&#o=IVzBHKwQ1pa9@Dn%7c#ww~RRI7YeipG_=mPBa=mac1N_*j=
zc$*++otgxNds8Rg=G;+*!0MT@%N~^IzuG&=F@XcW`g2?KftkR6+V@{T(U(2DbQGwj
zuqHPcw)H0iy4{>O5J7+zBPuRvJ*30}OgDFw2QAVBw&@~(hEea$@+%t7D|k%K8rcVq
z^q88T0mXO!{-TQN!bOxA)P8l(w|FBIkRhpvAyifrtQvVAe;s9@$+cxT21{VFh#%h%
zUD^k<O-C%|v6p73XGFiV_t%_UMK@_A#7%XNsW6a_O@-pNq6g*-fm^)kl3eSc>%|Ee
z&pQkhtAv@jV&nfT1gy4YVk^Vq?i!T{bRC#R9id!fp!QUI_bxg&z|}z39b`2W&Q(xQ
zwZWJYc^tjn9k)#fBC(I4NX#35L!wt7qU7Fh;aTNGD=suqC!SVR4FF<CyH0`7d?7~y
z9{va4**T<%KhuU@V}i~?lXkeHMNTvx-^!~k#c6ZE*vXRj!#{CcpX#B#rB*C(VrkcT
z(${;qOyJL^yR){f^8><uU>IBtZ6Cl*qg-4>sj^I)O;4n}zVIp}2QLJp_4cebG;YTi
zWh*@4{RK1+&(9To2WXq6j-^xiGbv(ZE6?aJVuAR2=(?egaX8Vfk`pSgmRW7S?(R(l
z+0%JRHB1f=0=`>~Q5-_=gT*?&cAUh`)#4uO3EI|U*>R}f#ejGH>{?Bvtun56n@V$3
z=87GF`r}-joI!n1s)TZaKaI~f=RQ8fkl-xhe@SCk13JNbd{()tZ*Rm<Kpe=ug9fu`
z2e2jr?gKC$>uAj4y^Y0n;u9C^{Ue@lmd2EFi{vS*-*GpRU5i(yQIwFu46rIY^ahCB
z?wVLiLIZ@*11HS+-3YY-yI;Y<IQE16s_HpV;*6|<WdK8$oM$BJG$<I_<m?g}Q(r$T
zXpMQn&Fm52%q6Z~PA!q&WUIdQuuif<kV}*$&U5u=>$dO`7|QvIs6}>f_bu+C*pUR6
zyU&#~fNN61J*ADk03Q`&%hWJXL%>j8ZCR@N=+RbCtmS4Fx@$5<&>@V^(P;DNiV+Ko
z0{o<E0$6tSfwb(bc+LfYY8`M22fRhmv)EkdgKH=tb?u^fQd9EyiC-w}!W#)yxFg(I
zWsQTVGr*i{8?(ry^9G7qio1&XQ`U6Q9XyEm{9C5)3iL%v4YT}LGsQIh6%b@)>n623
z(JK}zYsJj-J^*a@1bG=YJ%FGeIc<Y{)XWehVHP<u@0BMpH{arfwFrDzMFh?M-jw=X
z!@>X;m(&L;41os*%8do=lEk`<qU$+gruk0IgO^|y2AAvMwDRrwqssXB>L#J?3$ld#
zCM3zH)Q-^f#Ao7!zmj0)+1pNIEnPoEa_3zYU$x)Lcr>ucDqh0eR+ohZ9jfm;SL(RV
zN)Gu`dlt>~#*~vjq@Z}GK1e=g%lAv8gYhYS)kd!VWcuAXmU#${^r6ms{W&ieH()xO
zU!j6P-}F(&P(YM5iLF^rev;<A@Wr=V+hx#-sc+5Gqm$m&XS%i;QXLyCQ?wniU68sl
zomysI7QLQ7{Xee1<+!mV41SymBt$!>!E9~KV|I>}-gsI`S9-I)n;4IKNI~35xm8d2
zqw{s27V5P{?ytxxJR*N0L>5cL_cPt3AU1WJbWeFVUjMar0I*oo+Jr;~^PISRAeF_Z
z*ENZKM^CT&g6OBT2Bu>C!L_&LgYB$UO3rYG-g=Xqp6uf<+=EnD%~V0^3Xvnr-q09x
zaczPh1_(uK+^ZF}-@R>0y0K!v%^h9hWaWFWMj>{5E8#Qcu100c5tl1-_x*2diTwTH
z%AHWQb9fAHnO=RwtGf1eUnN;WAq;I<Kh50p^ZH{clHc_s&#$;<;k45+o^{ktSHL{|
zXoW#|S-a`aic?Q1*P~albQ`3$y&rs{FdB%yeC(dl!0KR5gXg{4k0v(cn0wLGM_hoI
z+iL6sI1}XNi(w(|zn<oTRwl?3$jAKyr-F~^>!D(^NOwgR@<YGJ%94&V%r*#Y`#@LB
zm!*G1J*}q5CQS<*EVPbgR2uux;k4=55#r0jyt-85vGYuwjw_@t<%otXTJZP@gfK;p
zqRq_FCZhit1(Rg$V|=XX+WvTZ6nT1AwpSF>=nQerr$3#B!+(;u?YVWzHE8(KqieIV
zAEY10rML_LduCS9#~=)l%d!Cdd_z2y^GGgG`<WB?e$iB@`__q5`G8~$6>63IzvkmC
z7o{TdFr`wPFTd3ohO9*f>$>|l)yj?j+|wz{&ZoKI??8?>D={8x?g;1{5lEgezxJTq
zgY^{wF2fhy4L9dLM}J9BE(AnGy!59H4ldVZMUgWsBTLSfB2=v=MiG^nVq?=7;#|Kk
zsR1Q>#qOf<)U7SGRmN*;JI^3_(xAH}!61tBxbiBR-tK^AYNKU^PBYi<=RdFDswz<a
z-NT0qI_kTe@K$jYI018s?8uk`!I>JOW!{z~;`Cxc{g<i%oE6bQvepS-2P`Qm1s?HN
zs$-n|#VHJxwCp}>GJMkX=Wf|24Ts}ZGAMF?BHmxRc!)fy>o+%<(Q0EZqa7{;n!;ih
zjBbLY4@^d|RImYtLHLGx>_v5-Zr(Q+HR9vxK1spMODjPON>Yp5QUspSl`6S=M_B6i
zqTb^A`fW*}kT-aQ^;^&TEnJ2v0Xw#?^gEOm34MR#{wc?1A1p?1eYVFzTRaAQgUL!i
zUFJZjO$=l#4krJm-B{E@UKVmE%i*@~Yo+ZBUmd}>9~A)LsA-k$PMa5}u7Vmm0TiV(
z%8~*CvkYa7Ljj9%ZjKK{%MHUDH*5!>Qk64RovlPOHNcnkcdaK+R(-~y(Jt*hmx=1}
zUar+beK@kAF@qN7b|=vKkKvetXO3dwSNxw46kc1t2<n5J0`2Ct_P_*uY*k^UpY$or
zzM5{A_y=!Qg?F;s^v5>P=ih5)-IzOPJKvX}pMuhaMcc&%JWim)EEW<;2;@c(E&ZSJ
z%{@!?U7P>ZdYWTc;`xd77RCS0$JNZ{v_#9|hntY;wo~y>3F_Xw2v^{P>C&A4j2}H6
zaZIR%<;v{}3Q`2n>PS~7dYKa`HpFG`@vW>j$GGvuC70WOiWQFc+VVN>(7(E=!`JHw
zyWT%GclE0IntC1^r6H$yZa`aF$X4aq;N6{!=vA3CRoE~;0VM8aMBeRE-`xPtonR;W
z|I`Wd6^>z)@!x;FJaLJsf2#go<NCGu;m4qO-`7oPT1122ovTYm!J|@%9k7dfR-Tdg
z_r8CwvI2mv)znuZ2M;+s$A<KhvoOWKE!#*}s%>*5!+eLf8WK+C^UOy?g0HEsVQ4lc
zW~10s@MlEMr^5N|!p#<bFP=AsZ#JgMc*k^Vk}hC`fbZUsyLx$9om*~LJrwR~=qMJ8
zKKr{31dzhc4&NGY5=zzD@gs;z3p{N2i$!*SmZVz_F0FJbr~s_7m=rZTi#A?q#v#2c
zhC`KfGRZ9{&oy&hYqm?}XU4sIf+JtDCpQV=u4Ih)w$=nji<}&Lo62G#4A<>xzb)H+
zpwtE4=fQ>HbkyGOb}3r5JPsfIV$O=WX~XKZ_&oZ;ylhtD6_Pb_g`3IQCzD__8A@uY
zV>~w^5|b$_{1M6||Mb+*laRo3m_WI`b&9dUN2Aln(rk4;w!Zh`T!qfoYhyo6tNGZM
zasu`X3w^l-^**G+-ErIv(4Pai<oiCRNRB(xz<-E;1DH0D@~B4f1MLtyR#}9h{9Hug
z{+BmA8ia+h5}HA7{1QAXy-1bAkZx{tdsC}@WxE@Hp#mcGVk!~Fc*tZG&$pK<2>6GH
z&+3s*Nyk8$^)zn1IUTu9Xf5?}>m^;p^OmRMJqt7NFHU?7cykpEu{=z-e)w%?EUs>%
z9lXcCjD3YfvxF~*2sHkF>xOo}y@mXKX&m|E67+Tc8OI&yev~^Vtb6>cA}Za!9FR-Q
zYkNdyAJ9>U2-Uu{)ao+QOK=aBdx>dr!x->w7xAE)>@3vNN|^F!p^JY9u7*XiRqDL~
zvi5yWBSMD>)4paQUawQETtw{zaOkm5M*sOUZg->y?^8p@HLLS30$9H=J*~FHxKH}h
zT{|*cOP7g^6)$IuiZ`31azr)Xa0?h5@omHZqceSSCR}N-8v(z{9b7j1N)nh)VmUj|
zou!p|^LiwRAp*imz_ywE^22SHZam;shpTG768KL;t(cW-HO2MG<Veh@S_(9>qsc+S
zVIppiqx?vDgh)X2AuA;hAo|Y!`|sPIRyflDQOe$zkTH|@IEvBW?uU*SM3(6@&tRRR
zQAZ{J<fYrB>9;))NfpNvnMHQo7{*;;uEHuW<`tQK%<4C++A_-!<P1?ze^YZ7Y}q_a
zO}wzNLQ8We*dssxLBW22VY$-C{*Xnv>chKM0qK6z?L*4C6}6Lo??qk!L)v2Ad!@=X
zC`I{UzKd7E@*daUn4n%m4PZ2V{Y#al6LWkvitl&p2gWyyk<l?k)vj<k@ig6{^RU5X
zMu&y;E6(M=24gWTNIN(e8k-S0qtdZQeR|lN{yiPawry3amsL%pHK%v#OVk-s`z*6D
zrItCar3~E-<zn0Jv$x|5d|JkmN#VF96fb0(`vJ?)x7Rx4&1!+rEIo!>!Jery2{X`i
zM+ng@2(i*BV(CcYu10$OUpeK87ETUaW^CO4XRCwQXjuI_528xMSaWNGNjiHq?$4<|
z10swM;*Q8tilkm#&%U@(xdr93Jjpvv0z%&&SG93tI^Uga;Yh@;CNDB>e?-ARuJO1N
z(%Zu)NckD#9V;yKDBJA%@PPcKyj2C*Zj8CuY0LevL@#J-Jp7IV$QShI<QLiRS;^5D
zI*>zb3d$*HMhsb_Rrhp;6@(%1!H}oi!-2Ha?Fy?>wr{Mf|1H#fEP1xK<VRZXBxgTD
zj)?#E$9G9{M%7~FOkna4rPgRs-g10uIRUp)8)Gl+ifpq6Y8bl__``?LjY6Kg``6Nd
zr;4+;(6{tiZU;j{nG&)Ha?$SN=kjKbA`g<`E#?Q4F^N9{C>vTvpZ_BPzktd-4ll7y
z9Ct?Ku~^;YNcsdy?#4(fkkzIhJiBHLJ>IFc=H`79r^uRVFX1KNY4FI9#9Plmywf73
zC?II)`IcUXCdM@VG6}L}o_*C%u?@~rZKdFX-rQpcbQc&_7`Zus|B_MAR9qH@gaCNQ
zx`nIRsYr)snh6D_h!Ky*rn%*(dzuJ3Vuk(z^7fbO0TCZioC*R#BMS2KLq3+X#E2~V
zi14}+i}8J*B}qq#w8CmL^#_L;VvKbgC6)_Qt^p@iwBw)ZANAVW{A>^4?Eq}RHlwk9
zZOyJyQoIb_qp`gq*TQa>OZ4d(l%q1y@U$bT(D+)u4`2w$A$qmbYbLa=xc7MO!>bJ1
zIN$NdQF#pC<PjV(+OxHs#{96&0H;1kR9x)eHhQ$4l()=bjvyHYf~+s_ludy5IfI7G
z4JA+As+mz%OgSq}zRV?T?q=Tr5%mpzw;;rY*f%?xNYRYM^tMAvw12vR@(yxa(0b3T
zB@T1NVs5w1)<eY0#kS4PqX~T64&-hc6kRvoK(Gk1{$U?tCDT6ke;=X0f_7udUG#0!
zIzna}tKdd6j3hQ{-5c}X>+AL;+^6U@M0h%F-92x$yqGGPg+9Nn*UI`?hrVtvySU;?
zG^V<p(;1oc85xb71lXl0iJDyGa4i5Ea0YpU)fX5E5M;3hvn14Ogu0?Xo&Zyd9l%q}
zkant2&~2f_IPK=yKLXnUAQtQwU0lOy_*d##<~2BjD_*g)RIvZn?G>^5ZT@RP%^5t7
z68RSC`@TjUK0I{cY<b*s@gZlrsDQzZsaZ`AwC9sw2X9T_Ix>KamGPUeAt|$ZX>F=`
zbx$%7Dv;5oxbI!Y{R0wFT`BDdG1GOyDt0a;AF7!7t<K>8w7vINQ%}%13<L-mOh8H~
zA)yJVbVOQ!P!v?EilR~l0V}<i5Ta5<iWQ_60TrbxASIy}>7aCi5D2{m2qops=Xrkk
z{sYf@&OJFNcX#gWlsh}Svv(%F(;`wZV6CO1Ak62V?uz}2t9;MI^R*Gp;xMnAt!767
z*>&Un=UDl!$rx(uaq{jH;hopPR?Og#-#^}=zID^5-I7E(akqK`W{RvEob9zupF#Od
z7$p0;sr5zl+P$xTvAt%{dHGOoNO{fV(#})I4<y-jQoz2={*^9^nZ92xfE}gBbNWiz
zXpTeZzBf&fz9FQ~2z^_?5O0Hikec81&(MZFYqi3r?AvM$;c}E0eYew0tv~*J)A-NU
zEem~;BQ7iryJj*t{vjpoDP+hBVdOZ6vuUmI`FEh#spVtvP+e1!9OH1XlIai})E`Rq
z+f;+mDTEp?=<%+F8>2=@PCwz(deWutf8}P!Q$XlU_Ol!945C#VHR;9lCHhF%5u>|O
zszdrJUq?TC&S_mR&DJmNLx0nsU*k*Xl1%-cYMp&eOF{c0eXY0%Z5jU*Hvda0pvvgK
zESi+i>MUTNaq%Z5|18OH#7)`IqA#MjD;u*y4ZcC_HNCh2uy;y`CW`E!WdsmnuXbJe
zle5$^b+#nF_-$OIwQ&^p*qI|3yjLm_5~NS^za119%>BS>$4)(c%+X>@eU%7uk@c;I
z(tIP(I^`$tg7($q%<}+|4ml}uvF5d;<2PjFTX+4~=7g{mw2x}7f=f1`%`P0;Jp`A@
zK2CiDlv*!_LzBWJ4v%@=vK`~|s$x2WWH$r!<BNkXLO1fW6!SI|OdQV!TnK@medC^s
zObK{D@Nf9#*i6#U^_MpO(H*cfY)PqW^7iR<?0yzqa3NiyAVN=MIXlU8@q>Os@Ac&E
zdGmcjM5lHTH2zf4aAV#Bzw0Rh`taq;6baHU)=cjd24P_H`m+!;DORgq_}!UZpU<lY
zJiqPGdS;n8=w&C_*$zmv_t88mgXtS8MXHKE$Ucx-a7A2Ud9BYQ+D!Gj{-|Y=rbd6`
z`bWg6ovcC+<cjZfA|42yG4$-?Lo@V+02;&FSBH@+t>z;h&n_d9Df2P{YJtoTq50*?
z(Ct!W0a{Ozekbx@(wZx4r!G7yTy=BH6tH7ioMQ2`)*zmdt(y3wKUe_1ZMd$A7C?|1
zwnNVp1y9erT2nJMD{$E+LPqTphcR*WO5{)ZqHp0MFT3ZjCr6sR+<gN#(^jHqgCe+&
zWsd90tNsnv9Ee8<CY?KqME0k}61_U+sX<lG`CXHb8kDc$_KklXWUCW)4qE0@x@l)(
z^io1v#cp>XQ`!lZM^XZDPKk@P-)n;HVqRvGgBoY`B31cFkR-!7%^9<!c-pacz;<Pv
z*?pwdVfd-}BN_Vxx7S;hQ=}xG_8Idle?3ZYT{`P3<LK3nC-cd)1(}srnT&rsT~@ME
z+i%LNY!pMz`24$m(lY$t^LZzW)+9t?+c@OU?z@dneIz$Ye+xejK^}&ctTx7)>}bC6
zroZIYI;7iWFgz6==^>6p5HtMQhpt`)5lbD9jwIjd`254l>mD(Z?ur@I-%PGze_r)w
zNS!~G%&tbswJL(hTVzD%U2l}v>7B|xbiNp`JaIy5%ddrvI^qreNS0b)Saj)bysUVW
zTf)E!17%eflMk1jE5JsB+guYD9<^>7+k`$obtTb4Cdo_y4$6aw#845RJxuKk!qo-u
zo3<qHxk^+lL9TSw9~Em|U>P?Z_cAcgaUy@xyG}?<ymS~;G6;UTdf=T64T#qvJIJ5F
zePZaSao4(iVCd1hs}1F;9~+m=%Bbmk%Y@+$nAq?8D<>%TEBkjE_gq$!%#}RQ^D!k}
zY_&avqYF{?QlAl1ktVNTE0;E)$oJF=c4K(<(X`fQQz$M^53zWmoO<$@njTlk&^2k$
z=^;=jOzDJoRmyVG??{C(uu5Xv0K+2^jv}lj74*~O4c+<)jM)A?zJQhaF0JF+GGvNb
zMiH$rQU@%OZF#9Z=3l3P>a*}S`dqEGNky|$a`VOf{67DrXF5D0d$v=_5mzz`96wzv
zy_N=$MVn1iV@Ii8&y>Gwe)6w^U357Tb<51WiWa#{SeGNke#kswq_Ildxn6Zu?X26A
zZ1l(aV(B@w0Qh`Q`@6#6BaaFfn7r2IzeUEL!*9CuxwPc~A_%r+gwbb&B+(=nA)nJL
z1cQ(+z>Tv)ir8#|rC=9LTT;--)KMw)HqIxJ;S?6MBM@|_oi#iU$Y6uKw_P!WsJq0W
zi~%lJquhjqesm92AniT@KUCWGCEqU@!oJtsn{{zF8k0nE@m6Rw7;a7vfWn^jn9Fj9
zX>w!SUw3Sh8*zmg|H7I(docaIdK`tan#HrZU*EeUG#k>hF74hUc@RoaLLIxpa+Six
zO2sY=%N(HuXs7WWmM)Rm#8HT#O>BRYjR$G=N3ha)w?c_~Te6E|>8W8cRyNLK&q!N8
z5GG-e-1YKB;0qPF4OG~~Kt#dh*qNF9omW)p2I)G-U}4j0uwMV<Y+Xu{_KQi;5LgyD
zCp*NNS-WM*(@T>A*j2nmI1{jLNbseQ^{O<E;wvGm-d?JU@%Kbq>Ss{~-j3&Nvj=OE
z^J{<g2s4VhF#FX(w<Mg19@{+q%irqXzjb|jrpc~jHO$hOFN-w#Zb*1P*z$$82M^5|
z2Fnatf@q(g2z{;at*FsL`GeSno9hf}_~%c@^ff}L^cx%+woMm(LQ{`_C6V1Sg^W;!
z__0p>eWCBHXi_)#C_oPKC{%S_VRYySTiJQMt<f~98(z)UJ-AUcxbewT?q;5MnCa8(
z!liY!+1QM=n<5h>#N;i-EKSR=Kauh!_u@kPnV+D-H^t`s<S+RX_fNhOUhi4gOVFu`
zh?+!OgJ7}rCaHjaG}eWug0)I<{ugWIM?g3L`Mo!lq3V8Y@x^H=RzGJ&cq3miEoN_d
z$>QA~jdnd!Sc7K~ylVZ&J;9EeDc84kwU4K)ri}^Ct_}mAvweIbJW?KdJoU5f{}l{J
z(I=YLouYqTvkmnBSSI8gfwy;laL8JF?0e0FApyzpYOenIMhs_n_m>9l@BvhjQ(uDc
zt@;d}bXU7PFn`@usA{so_hU__sm<LBPZVwrZ(I@Yl|=0ByV!$sEP#Xb)R-3I`(I!e
zwJiuqQv-QaD$7xnBv1Ia&3t9me8=g!r^qnzka5=B$b_&SoME^g_8=7aJ<XQqOby}g
z3+<l*R}zXjynCs$tL>@%oh5nIUriQzPfWPrcpMW<XRe1$TpiLmui3ppQ`>h)O?P-Q
z5Fe7)%gz!1RfBPkcxJwY!?pIb(}#EAqK{PV+m3t$pM(p}*Qcei_+3;#Y8FgDPgOdG
zF#TZiX4LzJ4Qgrjsr6q!H6xsqEEq4tJ0IwE)=J&%A-os;SboFRQoc3(xbP?<xxtn$
zBHPnzH!o$UncpHz5)Mj{bVxi|D1a4-V0on4F3@qNr0Y8NlYG%z`4q9XD{oJFx+EAz
zGA{R2x+MIFGmiJHfP@%r8XjowOEFq&TqRw5`x5-rN?ddPx=P;y?F(#t0LUNLsjIr2
zY)9rCvHwkwMsh0EG0P9DfTzVXu!6;VW5exyAt7Q_Xu<rXs$AtN5n6NX7lW#$MwP@1
zi~1^mn@>Eb7x6=>FFrpf`^P{5aX%$r`-A?EKMm!k+nTSdFkH%-g82})7N^%%MF!bm
zM_E$uosx3F^M0Fpdz6;N4d1Lx3sB;{qu2ejcAxVW?9O~M&$L0Gd(Ci^AjO@dYP#MN
ztZF8~+qxf|&RbSjIXlu1Ej8&c=r;k~dbsk@^*P_UW$u-1^P67?R$;I&1*XSmtKtM4
z7#EADtXbdfs@VAm?<)hh`)l-hG79@?vx;HwN;CA0drUHMcmv?V*&?7);{7U5JKsku
zAJI42w3K#8i>@A?Dd?x|w=Yt}ZO`s_d<Iz>e_7K4c$DKzs+gOAq{A@5pI5}cz3ZI+
z{otv9PZj6kq_zpZCbKf53C=?)@py20_YJ=%gL##%mw!l?16MIm)c%}z6%a!sv_b9q
z6Itk>V9rO2Yu4I|Q=juPfqT2YtKy;Mw~dugkvzTf-@g{4)*=6R-)tj_=49Br*miID
zYOI%R9obZFYOiFKdPgXZz<$~rPg?a2Rd#qRjptVtC4Cth1#alk@~(vJbCrhNU=}LZ
zbzXIO=Og<rar*|f<n^iSn}b5IggD^46_}ck(-qTiujgLc7PqKhrDo`xTDAc$(0`w`
zwRb4~^y4|Xs@vma->6g;K)-rP!~N0TAy%DqlC=NLTpg+IVEj8a9R8<5$7}0J=&7(<
zG|QR!v5?SHdNIM^0I%T|Z;OZ2e8l4nEj9Y}1IWpGd&U~)&L+Y}__$DeZJJcBL_m}n
z8x$8_)NOx3DCwix&*Gq>D#PJ9u|m}6-Swj;WPNcGQoimm48d`J^PkQd4LMPgj-#+U
ze%E}IwM!dOSO3n2RkBJ-=lQI*|JC&j#|q3k;6m}RiK;0^ZK3DZJ<Es&gt%<H+8X9w
z&-+9^eW?F=t4fi<FKv%y_tGAWJRBm$+&R2m|02Ilo*X8zlNPr!Wdhnyq3oP_M*V0Z
zY*gw8Epz%<b3Mmf`##&TWM7vt?0Ojjd=CK?Zpw0hMkce_?k5qy>b+&S&-nB)6gfKn
z?gaa~uiXm9ZQfHl)bDlmQqdiY-8)vW;K?$AXTH)?{#!#oMQm+5duw04q&Kppf!-+9
zRwT8eWjcQpIin8f)*tzlhp5dMdNgv*{q#H6<K)$lrr{RCZ+V^O5*LfW0QY-;uTkoT
z_U1E;gFNe#W|A^>2z{{zRv1-cu<JT?T(nAxbScrpYp|xIG$qEh;b*n`n{OJwlN%80
z-?T->Z}hZ1^6UD0&py9o<3|BeG-1b<c+`8ld#Ukpr{wCNjQJ`9U1k{)t#z9Cq2g_U
zig45Qd8w97XXO`DM~*`K3tmzP6z-BEw7oIyBF`YJ`;|JI;?}LlJlglzkJq{OUHI0j
zrOs-epN}Z(cLOYIVf56|m(-i?&+9zX8)JV~M&WXCijWtLZYVubU>|0iyqEeGMM*i(
z8M}L27v&nH?^};-^Dz0;GH_^2jI8bCt{fy(biP`OJPm(`CFKGp{8B*CXIFKCbh94l
zp6IX2l(}`B9)Gv(O|0Al*xtP2?T#mmDdMkb&E?SSwd#kRq3SUiEq|09j9=+)-2I`P
z5tMujOme2Wn(u!w{5-eNY2Z=V9=y!jKGxVr*+TAn%|tE@+c?Yz!2iC~8|G75Ya8*9
z9p%vrRWAA6w(KcPqV4k4qXR3xw*+tY%oE~NOZ;JpiTlk7YflCqnj}PoeeDt^$vQwT
zyoJ=>G5PymO7tR3PXH8vwDxGDx;#;m`Tb5xMJ#?u|4dKRYRaZlR}Atkd2dfn_Sdm$
zxSP(?W9ZNZU~?!uNb1zlvAayLq`u(M>Wx+!DKKfTI+vS(I917yPUt_N3uNy(J_8L1
zWMf0Du7|S*93ypRZf#!+hu?4WGbD>zXI)onCrt%g1zrCL+T3j9&Gw0KMr9UV{?i+(
zRFLn|P7}0`u_&4gexv4U=3tc+kv#ae@{In1=QY*#`*JI=J0nla8cI!WwEFE|-Eu*?
zA8%PFOeoE#UHxd=QLM3`7Jd-lxwJ4^BLKOp^!SA$R8x+XZ}h_jxy|yI9S@F6`5O*2
z2Pe#Se#aPC$wdFqwGub5d>S?r3a3P@BM^+m5wb^hWZKlvkxki>y}G-JoTw^Ind-Gy
zUC$bPl>a%O4C4-Kah%F6<$Lq^_jqdfiM#T0YO}{aFDO^XR{2)1FHF~biwqiWw;>FR
zf?IHM8$I%AtSj4*eZQvJ>4-V-$Qk>;0qM*fj2E^Dw>xM7Vd9Hd62rf(Jyr{A>Eon-
z6Z`pVdW2>XG{5n&$cGtei4R6VHk!WF8QvcC@z85~!;RzITwNIRMm2o7>xZkZTnvF~
z{ar7qoe67P385T*Io4Wdyn^nkq|4Lw^4913phjGh&v^e2ezqAKXwH-{e>$*WmxTqK
z{{Ir@Hc9^9B+OmUy!$wv64EZltl*l1igI`iz5C|=-At{YGFOmsrZ(>vwBD~3EJxj&
zyoeE;{K#6j-1h#wk6w_Os`Kgwr?A*xjX1@wQvvVguSlKQ4lgvjQqbpdRlf~*Byc%!
zXYEiv%r3B6p4DQ$;Q;g6T!auqU_=l|r|*lb@_XHTGAE;r*}K8--n^IE)u`vacjtJM
z?cl}M4eW1zoyUbEHLq>v?-YtY>N?S56}h=KF#$Y=8*I*ajAoxE)G)?hU2`hk5<JGe
zet-`hN_h2_`@`_15;9Vu;K}v$tDC81>WqVsV*!R0e(!`+Ib&hTW}kMcr1gp2<*_9~
z!l!48P44v4dv`gi*SG!J7}vS4fyFP#I(96w_ZtqqcvZ#m)kfnzzue|@J=F`%O1v8J
zs(Ex7S9&5n?vlObA6NAU&8PdKd%40;tw)F9^cW9KC7S8yQkw%&nrR)_S)=ur6|DWv
zSo9Bh4^+>KPc?8qpUJ&XOgel5%J<T(@IOvCVbP<tP!)|YYb+h!s*ybBQlwLp?v^L3
zIBOUgSp@xkjh>&fI~X;<h2CG%JXAb^v&fReZL?2d%YclH_nQ)9wTP>0tRJrR_F!LK
znRZomX=*Fu6JzCnFMTrgGGlh=#^R<!<#i8-(Z7Ai8}`e7kEi>-=!YF#{=8{8hGp0h
zSofngx60!ia4CnIT0+-P+9ljA<Exl`-5<SI23G%krC@%qvz@_H+nO!Kf5i8f6iI8x
z9v^ySM;-Gv7BFs?^3fAFn<MC_WsYrAA$F?SB`v-c@LTz%^B)<)7XX|aF9>b^bEk><
z%F@t7W(hsP$*XBW>yT<Pj?6w4vvrzQ+L3fGv?t1&^3^+nD%C#8;7Uosb_Q43`IaMU
zW?Bo2_*M-Cg<OGU!uK5$aIxdoe~KQ}2GMDi-+IVa_y{hiAx2v(-%xA_TKp8vmdhWP
zFqrjW_9sJC>WPk<vI;*KokSNa7yaIHr89P6fYm%Kp!#OTrsmp||NA5OGra?+IJ<Ju
zZ*r!o6tObiFUCWq7e~HbY<-F%=Iiu;nQPRJa-u=JEo0RVEIh6bhW;(o*J||ZzqJy^
zyFZJL6Y)~@Y&XJ3zbkunh=YS3Y((*QB*QDDCS33d=^@EZ*(b-!i23MJ!I07N=9Lj<
zlJJXHBOuLLw7W}$9uene7^UIRPJB@+CdGG&Z{vpK<{#HHpV}Uq^oMt<&uU1Jn5k3l
z?|iV(4dVaQ2y}tagB~$nD#R#eY~L&Zp6B}UHCcSv^}_8}_9dK^hSEY28oOTSN-XH)
z?aaFEYlQz)1XMy%KGO90J9;lum*AESPXS$W(=$!J`Z0yO63BjX*Xj9#>Q80wX+?#Z
zx+(t)a;5*4J45NF!VdXe6z8M0!{LZaCX-Kg2?<C_>*M6x1Gi#0N>2<cUH>;n1olEU
zUi?Q8oT^;bN1~og-5Z4l;k9uAnxG@aPRK*v%12|XX|&XnMd25ELee-IGC1;oJl|XE
zw?fwNpWV~r3ftZsb?pmcS7%+|u~V6^yw~nEKf0=-VQ0*UNe&WI?J!R!SL}BrB|kJz
zTmh}7smmdzPW~skh6BjRkilwtfWa#1dl6pSdUvzfu)gv&y%zecw~pQ+s$$S?u2pib
zWRPLJ?bJ6AqUc`JZ{=5%zP-XlY8-ubTX{9Y?9~m;u!i5U4;Ja4ea<cF#c<P-mOzgl
z{AFaK{`dV5DJ@*mH?Z8CzmrGmknKq*pNXk5qkLJ$32A4&xeW0DRN@ngMAKX%Tr_Bq
zIb3~HPtC{1fC8bAjl+xWw>ROS!WaIXyMv5kKJ9Ew7z??{6VM+OPyextUTnoq4l#ZU
z_Xs{}3>-ZaKE?|>RPK8q-`}HoygfpwsvoDTXt&e)za5ksO?*}?G0%7~Q>%E0OK4h4
z-AD-nA!^CBy=&y|jx`3jkMaGX`K#1a+7iK-k-@k^v(l=1zA%zj2%N)Z<XRaIq7y44
zK33{&erYL4|FN0t7sI&>-}{c<k4Z@H_xa-*dgJ*>&d;*9$m#m)$#kdofm)FeNTuGi
zW#WX&WyWo&hk-h7h#`0Ep9;Gm^;<s$t|4jgSlZTXh;eUjH!d@J1TBJH{`c!%V>;}I
zgz-?mdx67Z9)oJFSa{S`#wwDOg>FAoxz>-gHe`HkzBZR|pzM_#biw&G&0#w$mt6H5
z^_Gk`S1B-kI)FbaWqJ=FHQ-_lgv=wqS2|B+XdmP18Q0sLCqh!m)6ohvk8z-y^sYg3
zWPb;YW_1$z1zjCjO^*=^cK?#H<iWVLLvYw;(tuOzUtJ!%RzgmA^JC+@kO%!nIXwW0
zK=SVzG#&MyMC39=TSGOT9}<`OAsd$&%C4!~-%@?cfIQ_u9@Y85!&crPR8g%RqmI63
zR(sy;q#wjve}|B-*du3|Jyk3BpCm>k{20!9vf{FCOX%pnVat8CBEb+-McSy8Zi%U{
z$*U3GAh!FCJfB+`!-j*5w^&;CQ}v$XWl4bSl7YJ0eXksJj7`WlM3;-;DQvg7b=KXB
z4AE^T`q~kKIAhg+W%Cxb2<C3iBm9z1?qL%t67rFwo%1LNzIh{dfWbyH;AJfERM2{8
z*Z{N2pVu??y8EexoaSbGB4f~ps_H!r1D}n<YeX-*M_-OnJNJce*`1N3&geR3La8-O
z(T)`_GP+pBPXR=m;R9BTH4gO*j@xp_40#f64n4ai%@mdY+70sGCn1dL1jE9lz3ICh
zkzUr0!wZi6cZ!Y$4EiBM20^TC=zDg(bg#wRpFMEal2hN5i1|eyf1a4bB^ciPjBY~@
z)?Q<T9%)LBdnp`cU$4k9jw!~b5ps7=h*bb=xdO`R($^UG!9#kN%y;dyHpl0WHT+9W
zOGXpeNx0WE=%ptaCqthB&+^VEq%KuT4wo7gR~9BG)uZp-U16$rn7iJQ<zB`9@ks36
z1HSdo7faH|v-cX8bb5}^#BALr6Xkwm1Z{Gh(Nffk4#E0s(t^@c6pls(Xr@d0L!Vt~
zVr9fahGZK<(q!3S?;@jj4p-8|K3(){`A0_-dz=qkp6==(`WMG{^*iTcMBg*RPYiYG
zUAr{s63+N;XN;vdWB2`mvTO26mE?HuL=oBKgUGfLD9_<iWJ${|FD>kGDr`vg_zW#u
z*QolrM<LrkaDa<WD-~@_{h(?Vm}W<rS`Q88T+*Alk1BY%H}mb^hDU3>X-Gr1y34HX
zKH9zY9<JDp2Bm3w5d;|bc_O~@ZGI-^3vKr~Y)aO>*mJ->Sj5>{NmooUq-eYS)2L;>
z&l|?OPSW(>h<(}>lkxui_^`UHBn_^KM@!Ye#u%gRVksgcQe%EBDUN-ymY}e9N%qOF
zAa%oSr$rT#xEt;kRg7=ldEa#!IPsN_IlUM0L6;4tOO12Lw$IY|<}z=}=$>23-a{b&
zR4I~i8HNQ#(eyuDPGu594=)us02$9k4<^s5&KI}6plG;hQ6e&dyek(Zo>oh`7(Bys
z*h(UT83+3T|Gl0O8qI68+zX5Jo9AdB`BrZ9;U$TjdnM>2(iCOF=@Z&@5M7@iO)|Fa
zJS4c%DIfl4vW5hBj*-(2Cj}V+T-B*6k&YbwwI5X58gU!i-SpA}1G`;B<)~z2?x~sg
zUyS-F!;9;yJI91>;l5C5k6P9y&_j0pzj#J2xbdb2B}!DG_Q@E}c|n30+zb6qB#N%P
z!u#ke{y=*l&9JhHWmIrB3W>Gf-9rM8YbR}0$(r{D(QIFHN0;gH6vk+>+XiL)+!Ep0
z5{Fn`4`&ipGmZ%{kI-x;IKJ%ZjMN#O_1%@ZY2+ou$V5k?URF8u9^xZ%?zS}yBzQm4
z>qngqC)rjP0X7?IBmPr|-Lcp;Gr$G4Q2Q&(D!LDi@`5N7{Y!Vw@jd!xSku7S)OXfI
zi(x#)PbE2Gm?eCl-ft)9(U(si9+s+c@+XFt6wTn<cKNaZp-&Au6%I~3%4kgE$p3c1
z#E!yRcfwWiop(fWyS!hs*&{!-4O$1`oc$XD<4(J?#*XNP>vAd@_nw!2I}!P5aaq6b
zSCX&N!9>+o)gSz{`?I1?B4<d1dm(clG&L9RM%HLe+N`WX_m>!NTJNSJp3K^NWE+&q
z{@}{4lp1nJ6?+(5(sjb4?)J7M&v(!d$Hs@B3-&*7o>#_95<=O&JB@v196*vQ{74L2
zL5~L3LGsOKsoYrcOEqiC&NWsI0<hy%+}$kd!2rK^L-VaQwBk@gSp(+``CT|uZNC`j
zw6Wd#DwGI1+Uq=+^qQ_&q!S30UO;lcUh6v#_}s(KrA}k{jEq|`&YQot0ve*X6K~C~
zYH@j`pXf_6r?}=DgltZ%KpAL=4mu?~8!1Qs4d2)kY#R;is2#4m3u&5kjQa;U3^~r$
z()8OmpyCL|q~oqhBJ<4g8iv6<y|A!{6I#@B*Jp$%zMx;lCXl9sewKb>kZLDuu02DX
zi(pi@(yp_T>5|tBzwkc!+<Z_9aVARh->mQ;05;_NL-zG=I_4bH+z9#V5=e39-CLr0
ziu%9I2;*O3`u<icIoPiq@RfZ+l4=znM`q8U%^OeCv(G-gKV=YRF1e~5aqhxSrFAFm
z<zj8^Ob4Tsq^?`D|0EQhg_LG?`>0QonX12daM1o&E_a?%+7Zdv|2D-SZ3Zv;O)2a}
zaD8$ikZR}_*H%_}*$w~yxDLJ?dhry-k1<ckF_)#V_5Zg_?UemLTBc?kzPVxho*52H
z{rTsT%?*?rzr6L-#Hag&Al7%bgE#H(Jn()p@iZ(gjp2MxO7@=ABE(ZONB+da{4sl*
z;)s}ual_}`U)Um3<7)i+8c2;{Grbfw!tYOxf?F+ej_x<dtq*BidZfk@&p~QeuhH;M
zmtx++@ZEo*=a3y8*w(oEe{;=rVCyF@IP^nf*cE%~{$-X@9BY;k6lYw>Hr&?MqTbrJ
znR0t>h~n4!39-mVZv7-pzBrolJa2VL%DH^$?A^1!x<51pXqOrJ+m;BF$9rH%CMCWa
zWt?`W@7ZqE9C0{fWB=5%*DL&b<j6Puqcncw*Qklr^eSE~uZV6jt2j8kb)k&=pke93
z6whP6hF6;`n{1nLGli6ko8^!wNSuLuq8A6<b30|}2|c#)r>%bD6je@B5zV=^ZMrhC
zQ=Ye{48C=T@o-zySZpV#op5UF7G*v#9LzXgN82QwMyvZ%r~jQr;wH9B=N^)9tc;Dk
z!#7U9$fMx!{d2Pyg7vw+SHHB}nzdh+Nn?C!N!9L0pHMlulf6MiZy%4hW2DIvjKjAE
zemQF*7qzvl8&4r6fM+ErXT5VCDwZX}g82r()c3HlP9zE#2^_D}TVmSaaW6^yb6M})
zIQxN`4CkjX(qq974pcs0icwuo4O^mC)kFOYRS_k=kMlde=~vOl!;`|P9;+Icj-7<N
zUw#$c#iVF44D$plylCNN{y7VkO?j>)Z=aw3c?Y*cRo+!(8;PsD(s3^IK`oYw+sy{f
z%4=IEDtLgAkdg|RofY^o@ARx;kfKhMLs<%P)?rI`1zTMJY&f;S7MkNAfk>3xyc^Ej
zaldyPBG`+LT1AZ5R%rCaotu4DaFHyh`Hej+yjBqT`qu7}bm;e*<_F|CEB1rI#Bhb$
z6Z&R4$+#=w<PTtHdi5D#qcG3ZnLEk-n*!^)F>#e)P1w?GHb~*+46wJhN{3I*Rh_Z{
zUuTg8FWs$-G}c6*a;MS~C!_!{#*qkoDBG!D41}({0yWGUDf>sW*?Ug)<UIA9FtYT<
zzjDo>NQCts$IpL+2rqHO91iA<i)P*2BRFC@P6SWIKi*@he{`JxJMOEfx6yX_IW$hh
zc?Sm&mbz0tefmz+&nLzvrW@Wqg<7W6U5(u)$}mQRh&S(G)dT)2091isgJ6YRxZPr4
z4vdVpWM&HrWC_iA*x-}nj`fx3!@eR$PXdNPvxd@pM)`}BR(DN{yQj9k7fFL>FP45l
z^j`L4^|(WAEUbB*3=v$qB&I#|)Aq>w-7uYR0pZ3uWqC+ALpGfVoKl>qCII0008+!;
zjY0P9`CGG15OoRi?Y9i`kAMl*2@Y=|Er+sa`>c}~k6cg{;OmH^897G`xpWEgunIT=
z)F2vW?xXrS^$}5=4d+xu39=VzAIwEbgVO~&&fgW@Z1oYLCc{SSNmiAJ%~P9A=7ux2
zRrSnNG|4nz0_60qCLniBiVC`4ga!Oq?A{K+2cjY%%-J6kw?D-<gatY2pAoOwZF&3?
z)WheH1)kN9RVatqxxnyB#YBfPA(&keY-<O#kC}W9uRS-^C?AG7l*k-#v^(E;A4AI>
zFZe$DqIqXN^`BbsKW}?7I4ol+V84FcyRA)kyGlj#P6|Z?(YlcR7;@wAt~lz@6)5qf
z*1NaPDnp@%c@#uXQg?PJMSE_JZyk^P&H9le5l|1hnW{7kMnV0t!0A7;hI^q#{Tw+$
z_Qert*6>~Y<77|hibkoJXtHK_@Nn(^Ls4``@YTU*^%jqPeQx`ilf%%7a`&!NAN6*+
zL-3hGeRZe&TZq|@Ij1hEJFE1v46{D@1#pI1gA<~trl;~-<D$9es7)R3fTqjHD6ke>
zLB#{BHqEh*w5D(mV3d}l_MJ6?%Wm2<^aWeD=;ZF7z6_}u04T}hV}9Kd5fj`KOM@nG
zrjJUZd&S}}5iLMld0pNjQ{GyIA(KYvOue88<%0lwK>z%ho97I1o@3Ya_xsjEbzp|b
z>`Ihm-jn0$$6y|HFdoMMcj8<}JqHtsuinEmbhjSWiUo=@Ya4%8nx({!A!nH$wBSy7
zlzP)LKu5Y$e^zOvBq<%Dr-G1~c&-Ck`5{)SG8>}jA)RsVeCZicM^{S~P0=T*2n*5m
z&UB%ziE>TB{n6dk4*a=3AEM&R>NH3@sFX_UXB}Z(!e98yI{6KZ+PQ?HFX82vQKod^
zE$0$qYZlDsDZ_P+wCQfWL4E}=g_c$3F0fo{5+5L$_@lA~(?EYhP3=nUqb!$LAw09#
zHIClj@jA)_p!8ezyV1~wlbx5C{1{vectM-WJHy)1GynUeYo{1K{l$FGOLwd;IvYHz
zjxPmN0cz_3mx{sOFyV7Lfdt5?&lWXu9OyyOI~#d)Glcie_#<F2v@8?994$D6O>GQX
zTjShT>U1a%{T6nsSXDi*=r2~M{%IMD>53LS>!3zbH@|6ir$AMq^jEg#870kz#l^AJ
z1kEW&*sX(w%|EQ)vc$cMmRTS`oeb^Sz#Kn{)?lp%5($^n)4F~Xd$66p+>Nc^ATmW?
zK8f{hget^Po3K~5a;}Bp?Y$QPk?+ec-LrT0%WL<4R&bLn6~+?<eSPQGUkp?U|1y+#
zALO0aWZ~6vj)z<?w-Io3Cm&+*;<rVqXriQ;NGTt&b$j$8+2HAwEZj6}Herz&Qvz2l
zfCFGl-iuK&fUp4!cWdzA?PO`o%?LZ{!QKsR$i4f?5ot`ugW+$L12)9k2QXTY8c_J9
z;-^c2`y4s$-(vrir1+1PsE>U2Gbok3^>8%(oYuc_@lytxJPjW?Pu8%F!fW=g^sF|#
zGQk3m9=*Z|e$cjvU3l)6FQbn}0Nv<gtXQNjYGLyqOCa9T*ZD)BbAo@)u#a=SZw|^Q
z&pB^*w)xg|;%&*BiExFAeYTAzwFCH8&OPn&W{C*92gqWhteB}=wXwF+Uq2&+HM${A
z4@a*+LLaXrLy;`XD<>QF#CT55n$bA-e!PG#%sZyPU_bxRF?vnJLYW`;BlzM1A`fr}
z>(9b=coi6E1oQ^_Lx@Q!t-8t5WCI&58%L{mNExN@!z!e5J06p|pOZjPwqrX~w!MLD
zv3sihZ{x+&<tFkFP_yUN!-K39TcNKfHQH0)>37^Rp9i)KfE+#FN^gbgv6a@^m3}|@
z*AE<da>`KY>IHdG&ESKr*H-kd{V2n?xG?DBNw@(pT)Y#ZxI6+L(0A5hLCR+o$fwF!
zMgvI@+NH=WVz2}Wt9`D$2Kf3uT*+Ck*m3DS@=-?dP>xu_3aJwSCg*bAKA9F~UDpoS
z%1f*SS2)R)KJ6KuE3Urn<vCD4TgVrZve)_obyydv*}D86TDtu#xSJ3GPhNw`8fd-;
z6KlnnOak)(u3Z?Pw+7a!`JpW0@}Vr>A+c}TAaQ=c>{C!k|LbGS4vG%|ky*GZ+{2+S
zclN^@V8U5&P44GQSwDjNEMEB*0PW2Z1F`c9-1UwzC77Pk&aW>+(|^3+bjvIlWzF!k
zw}Q)OP>ubgy0QuUMDWx?pHm=SWf%~$5rJQi8V6YkJOuhC04_qV3}7pPS-b{l&FdSn
zmI#=wWotLz<nz-d(DxxdT<g#C#4@~#@3@3-0qU1np>xHCSxH5Kwkj((&p_rJoz^iQ
z5RY=9!Y3bz%bg-I6)ZxjU?(Xkr}q(f?0-nD5+BQlX|a+~+9<eyg+EN|Hu82Y3p`_-
zp_L)-5qs5?G#rT{acUz*60DUJE?D019Ke3^=Qyj~kJ>kR=U&@(!g%k*%88`D;%zpa
z{P)F!$bPswk&>$p9~m!E8<GBcy5Y4({0ogCz%4pvk!69gTsj5rr0T)EcRG&Qj?AMd
zz%Go#1SX{$6aE_$z5P&VE5w@p7^T`54LmJ3z%hGc*>$d2b9iK&ub*dO2_L@LWJKw+
z&h2A<eZg`==?B$1v4%zNUgOiT$dallyUW2f*FdGCMz~-bN(QheCbai#XVzfZSrq*z
zpad|0sR0X@?0n$K?T^_=43h9l_0O61nbq%%6!2pi!&2h6^(DzU1Hs<c!mQcgZ_Q;P
zaDk$$h)<Q$Yw?0AVg<kpOEK;~NI}sVLDa&johVmrn;$%CPVG9_;!E$tSR6+bJW9;W
ztYGJw0AdEFSuqaHydTiX+FQ$tvgl=&OdzgITzD=Jwdl2nx7@ppnK#tmc)k%bOTu^M
z#j&pfwopeB$FqDs^mAJ3RF$hbynaW{jWGZrLD<-7dxf8~k;(kFS0%}xh^+acA|cNC
z4ratSp8`4XR?gq{o!q{Gs|BxdzEKN|+x37alN&1Fr7+_}fU%%%%xeIpy%*r0<4TcV
z#&ffaCr1IF&!+)QRD-P_AI%B`dtdp*GTIRA9_6sTWj+Sk(r*smJItq+V2+(_zTcK7
z#TmC5{RTqh&I(SyKK&G?sqp24>KZ9Qlqk{}p4O(mJ)#t(Yg2Ta`>=Gk&ma~M2wWmZ
z^)LcErwN$AX@Kr#ZYHk>>pFW~DxmY}uJd_fA~@l+KQ!S3&egXfSr^6RbTre%H*~sn
z#8m<jG3qV;oe|&t4MVLjKvO$@VdO~)-a01HexL5;K3_2TlsCB@Rc!P=xa_SjNrCKh
zG2&YXNEs4*i#dvVPkM*4+C1LlGHmRrb{SSzWyB{3kOH<Nv%r?7)L*awPxP2;aXUy1
zI-vai2?R@wALn3qS;Do#W>vR`tFKEkw7UVzQNRV*)H!XkA`<r5o-6H<kjNLIGmwKv
z6YtK7P4b5}h~8#Pn^bt7@kBxSp&&QeYLSJwsWtmc*<b88A&`ZFsgpo&u=GYFe?SA*
zjDU7f;yI?fE$W?9$-)Xbv1iT_@2unrj2~oW`gKi!6aKj-KJE<ZxsKB87Y=-zayE;m
zirtI6ehU<NDy*#7+W)Fh&1ZZ3R2lN~39~(^WkV|%(dgByNzgm=7<eF+>Hsctz!Q@w
zFA@OjKLNGv81N*9$~T0uT*k}as~YBLXU*|M<|4Ih%>8oRUSoZukWs6bXLlQ8+Bm4<
z`_k(2PMya~k1vXNBvL;bAKF7ovY0q^;VLkokV+@^tbtH%MDA$loL$BUrr=P{FDd}_
z+hd;usE)f}1>7hAO8|X0-ND67tA%f9$4pCeuSf07nqacH0}=Re`H#-)K>k-iv(Gz*
zjX%BAGDNn51dEIt$GDg93Z0qicN;gX_0#QVd+y)LF&b1%BNd)X+hjeQt(n`_u=vBa
zY?vfBYxLulFVlt017y#vt+1@4ev{q<f@M+ubkuR8WFy$$6DZTt^_(w{zYFmIIi?;D
z!xNNcASd`{xaZ#j+RVThuz5jkly0-k3i}!$*k0#$hVC=mS1sFC3C5Vs&g_zT$Sa@j
zk<rsX#qL+$z9s+bwBJc3xka_b!<xBy^FQJLGX}#Sy;A2GGT?&{y!gQ38$fJ#wsuY!
zYS9WzoVgFniO2yA9_oUpu7ambQ}wm}X{o$`NkHN^VyJ_dZ6b?bt_*Mo2-&xk0m^>X
zm;Y4D%zbwtpCOX<e!Qbwg(0$14q9h9oQ>XNA5cu2FT9udE<y*~Q>0&p(Y9Ha)n^F6
zk`GspjjC&OP{T(YtE_sUF-zV}yq*HX<t80P=W5wFo&tONBdp~z6l5Ur9b#8xQFQJ2
z25ln7=U>$t`=65@5G0Ssiz12RWg;dJgUvP#F|5qD=X=u=c%4}d&gFZ0H~S4qvP?Vh
z(jFUdlM2vs4#SHdD9o_-I5llT$|It(PYz%+Cs-xCijx4M{?5<T`7PH(KFJGPq*_{o
zm-GiRRp5PG`~nle4i0}7`Anjg&m!KKX}7InUz!%${8TUCKM|Tx+GeQq_T0VKx1XMM
zPXRjil*2C1nlHdQVxBIus(ny#JRl*4m=;<Bs9!B^P2s##ee%L%4-c$(H|yo^6QCRP
z5>VhtEH$Q*S__~U?n|HK=)pD*CNqSc6}0j8@^IiG7I=ET3&TxhYT)^6IyRwEqu23i
zG)QwAYhW9~+cZ^jSx0T!djzz??JFglw)~=gsbYF-<&~B}z24Hko8xOLKqp6UYmH@{
zB?HvB#a!9<_j4L&P=5dx_4%^+{%8-bM|C;rMCp}HwqToWh%DCU|5}F|5DhGC)Dn>_
z`<4>E5<w?}%OJ?H7&x9B0f_}U>2Wgo-h^3xt+d7|a}xOLl#a^p4ps={lvXm5vE?v3
z&H597|2fJ18zatS@2!4p9d|Y0Aw0EqAf>b~J`1en#|o1deuDJYaVgQ_h>-9(%+dLG
zz5?21^VZToS)A1}Lw7g!;810QYZwn$kAf17(wL<ElbDvzcIiE#d@wa@XL+OVJ>35;
zd%wZd(=jOn>0jeT03>WI(W9k_h!=0iOw}P_QA(Y4raBU&W~D{L#!1!~J!x(&<zMd2
z!Y?)eA3<&GyBMP&wpES;idU-H=;gV!rkL5DEq0=TZHPo#x3kjQGxts{MDOQ)4ga;n
z;uP~$lW4@z(Gu5vo5-4B5W95<taFn8Z?k75i$e*R0WO1$<%+ZT#+W4lC<j_7I8h8T
zV$QNZWtt4+2$xlX#HRw<K5OSJMl%t0dAkS0kf~B++i@$&!aHDdTf#iis`P`pX?M{v
zO6i98W4YKX3r3pBp>%F!nJzv+sou7VX)IO=;v#Qz08`*9Y28h!V$jZ3ILn%7w+h&Q
z13ZiRC(pEl^7rZtx5tvf#neBMQ51f;D}AtFYwE#zY1=31GHFy_>W@i4Xs^ph<;&t+
zbB3RGBxqE<aI?QmiS><J?-@>B_~ntnwB-NM(@FMa36($#YpZ0yB=!>|@L0Anz5;-B
z7we)C-x_cTkQ1D+onsOr5!Bi+-^rTe&jtu;hV8J7m0H8`-w%3S=Y}9SOs$G-w7QY+
zNV`{%ik}p`(m1$}>tzqgD%|&a4cHm1b?Ehyz&*jYZuq?p#HNiWTeV?c`@>RYw4Wyp
zVoTxA)myF;o6aL3RB<jMhVXqr?vTKK$hfw>tlWi7U$*xK5hK4=1L^@k+cw$Py#sV}
zbaDY*(vY2{M!wpCv?QKeYd=ZNTw2f^U=PR6zXEo>S*6rkiGJ?Wl;O0LT2$pf|7CW+
zr&Zxv7e~oX`U*<$NTYSG?=<K5@ZLHx9#xN^k*(2F{qZFWo6AwU=y{fq*|q02{LT<-
z=wn<aY=Arq_KgZK%mB114Rd$1-lKRMQJNcOS%R%qqLC8N6Q-{Ls6qK1=kt3Q-+uzv
z)ZQa&vUQsLTb75miL7N_Y$3_>JsjcToc3Blp3JGl`&R2Q7c*<&PWip~Nf78=EwtyQ
zZ~F3hZ#K+qhV?N#n7N>e>LRqTX5lip<6vbwgckCqF0gM+vmpvjWv|PdHxv?IdP!9u
zVx2q{fX|H6#rLz~z_9}ubzf8K{(kr~b!_FimD|Z<m@>GV9(od^BF^guJ_v{^0!S{R
zTES+_X>Xe+*!y3)X%n`n0GJ38@y_WdlbBslg^N+YF>ZbmQ2^E*gJoj&cL2XQaI2>n
z8KvGbnuJgLciAt{yo+_-Fq??T0Z3&6xG7CM{;0FMxiNqtET9`pcESK-vi*H6e}Ik@
z82*8V9ju#Rt^j!KLqtywAQ9-@Ryp4v9B4j=$p%DZqIdo9>!|DvHW%2|1h)dP3}`t#
z&y<l&<dIT6FybS+SUcGbN(3MDVjgd>rRb=1EwF6F939pgL%RCYGx|l4fzKk}pfB_R
z#-})rMleJVKs9p^mBi-TTd2sHXC9{A6Th+g{e)qROr0~L2NOv`#a3dbIXuK6v!R|9
z)cWPTBwz~B3M{h{<xRreAfax6Wpu?}OoQ|mG4oAs9ZZ70_NN~t^4_OM=m3x60EQcU
z&u6&V{4u+zP}SR6;^@LkSGU6=U0{@npTwhM=WH$BKshrcBp50W5g_5QW@M}T;AdkT
z6}6B;5-K^Yttlw_#xE8UtK+dTT_hb~vyQdV;r#-d(~CtiOu+0J0NOn$*F5Tk2R4T+
z;R*GGNcGSB#x$k^0DTz$I{a~@A+a_l6+j!ucpsyp$u6@Y$<%P(0+3X{UIAcvjg8ih
zDf7TekR{rbVv*z=a6tuV_XohG?}ZDc`8)}e4NM`e&!f8iBzDtsHpRJc{4!a<&mlSP
zM#S4iTEL2GklQ52JIv+CA$bhjS=kp0NXrHcY9-T=TqDextJBYN_&-^E*5xIVqXmMZ
ztsRoXWlgetJb`L;Ah#)uz$TxYE1W%zJB5@!sI3wuZgwB;hR1f28~R{okKi(E(Iy`3
zj5OYmpMZ9ZKp)40wddo@(ZNhA_eA|Ab{&P?{_Dd|0lg094Jmw}%PYY!*XCV<+tgGV
zJ|yFDpym)27IL*X*92?C0_wt|;TMNl4SoeCyqdIlax=`V3nM{^S1;o)nBcgq7rV)%
zq23bKIfgXL1<iRxZ%gdJN>2luH=aKTG?YcgB3m$vsIwHVYZ<`LL5AMU$9P3)Cr=!z
z^jIc$DVc2;62~vDqf-40+XJJM9e7EYb<|1(Z^+XW9o|K_%|WKb5>kc1@L%qGw#B^d
z4rs%?4U9hD2V{EFX~6tbp}ZlEfMyJnPfUS~JMvb6kRC?9Oe_y`bd&dyT$3DWdlEYt
zM;L&izIu0$qwrJB9;$gPMPes`GakT{7!&M|QfCfRXzByvT=o3VGJv^#T%W^uLoU#o
zFmIX6_J$XP;v`0%@@_H}gU1<Yw9bBc(Xd9TGYUvFBBt?h0fz|O6@whPdRwWW6w)99
zW=0AZc9ZM0gS!0%$m}R^Mgh`hxK9NdE~O=BF%b<wU}lGpdEMm9zJl0#*psFu5`ZZl
zyc%7w&K;x~4m3oT*rDKX$sEAf8!x!)a=ehS)QC+hw@|sxE6`_q65ElyhhZ}WvvCY_
z@OIE=+jz7l*P<bDGYBoiuWMMWWaf>N_zgFqJ5p{L>F~P43(oUb5Si_MXgPke#3OY{
zTSx*Z1&0G*+`w+1Kx*@bb9&nwON2>G^d_?N8dN5Y`%IW~E$`v6&Oql)sL6k@B$Ujz
zX%$^{$z&j0iTP;oje3}H6cdR)*uP>KHI&AS$I37p=nf#aI#6>c?-Cueo1?iy*|;Oc
z0hHRGH79P4|F~3GCjo-1KpUC0uz%g$4#-JDaxFHan+SUhDRk*~+iPpl`n%4jsd{x%
z=<M$SM`Kvt%73@EC|~2Kv|6~?5(*LKY=0mp*~&1X&)8e-DnnElr&Dlpqtb>RMuo_&
zKvpsXdT?B2N{_TJbsNF5P#~#5mc^vUany*KEElj_8vqUUXKt)@NTU(ogQ&31ET9hB
z&5&roIH|U1h=K;wJJBZeN`k~QU{^orGO~=!@Hp_JT%!9CB=jKBb<U)uS-o5@X-1k?
z@&Fb^fxHfzeP^J?MX%56M9+vlqex_cA`h?5_Wvq}?0O7QBwXmN*A6NlN;7VIC7*;w
z6O*WH^n_lNG0}i@Qf=9AfO;5_*M>&}1o0P{nNI(E><9n>ZC>oJ<)CMHL^0U+HNQSx
zVvl6+BjjXnby@hJ_CGR!Pk8v%!l<lnEG&HB{}=e|Uj9D{d}dxvwlx*$&}ian6qhiR
zQUPFOQYF$@Ww=CMu=Bn!ZI<R{6*+}6Gm?>gljCxw<iPH;a~G$g=C3mQVvb_7cfWk+
zo0-AC;HAH&9AiuE7sGIKkpT?K{Hp#m$)|3!CwIl^^NodFNeX3ibC8s=p_`x@6udMY
zx9l*#452YhhB7S<gC$<bK9iH$$u3yx>pa^6El=U3#dR|Dv_BmdDlBYYx)mlV#i-LM
zk)rSH8IOiA|8$9jdI1II1V6}yH}Hy4o}ih)rwk<{{X(Q#L6?suq!^SS)F#>bz3Jnu
zf%<^}{n$<hR&W&n_dRS|<X+}bRC;#coVUa9j@gxrvCm4$8lXCpjow3E$8qo-QtWmI
z<9GQkUMTfs-nae<kv+wDaN6@BxRzQN0-)uDu)0Q<LSAp9Oz0H1rH$7b4xNkA%b;CA
zJ8;)%^oWT(EhovgKtyOfT(&`yCeaY(Bl#P6Ih5IO+ajV~YWmTjd48BxsPPf&JRr>Q
zCAY%@HkKMD*n8%7rC_hxFUkO7iF+BC0BY37>+ty*`9n;I2Fk{MSFdGD^?EIx++{zO
z{LO4MJ1b;D#N|XSk_I9qdOem-E_0_tE5ar8BqDhA{%cY(hP=-K3QR+c401IG_}75F
zpbLo3dR|db)n4?KzX?)1PBOdRmqR;qt)dG02gH`R_p!Ut3EV@#I}x)xhHve7Q_u!W
zDQtJvK3hc>kOmNI%nrSvHSBFLqEkmzM##Svz;~z&P*lu0oh8b0nUHHR2K|T_015LK
zKva}#9s+J9<tzY3D+`{^<C^nKK@V#$X0u%;lp72{*Wxb5L$#E=&GEtb5X`E{UCXU5
zU@py2a~rLI`s(}e(?C)BYe!>rnerI_1?7VS)FS)c2s?L+0Yxgp>!C%?3tvwS&$2VS
z?{_&vjmT^yu3q3<XlT0rfNyjO$o??Lm&r7oqo)Jwx84DMiDlhg;BmAb#&Na=AJV)C
z_yVc1ltFZ6M3v*;LE`^D|4c^QtSZmBptYM;SU?y<EWMPBfvU*nK;)IJBccMN7THBW
zWv2LW-fR2<`%JmtO`o`iO4s|rEz{6G<7}0KD1d5N<v`q!$@=(=r~u4+NcR@X+DkZv
z@e>w$HYSVxiN<;(lNh-ICU?GCoRjs`&pC=(VN#u39z|x|2i^~pu%%nwt*Vq+x2|e9
zv&$cf)oY4GDSO+a5Z;cnNBDRA6-bFkY}qq-TS#ySQ<WSrp%;p>_+IO~I{goNC1>-p
zL%a|ue8*4(z9d8syE3rBqBr_>_suT*5o<$~DrTN_f#W+wMb5LFg=|<14EHtsF3g+4
zFyok59q=$*LhN!d`{Do7Rbh~uL)0P5qv(1#*Ef!9^9-Z7&>7=nlyu>=@_#z;pOUNr
zx<NM(&X&UGLlyAhIe*xdAf6g7XZ{-+R@+~qXdjg8HNdZ=>}@_U$9l-(+VpTZo?}Yd
zTR$0Y=RC1$^a1GoO5O@?7RPY87;|=j^sL|mQ$=I={OX{O|Cc%4+`^zTCHzG4BevVn
zUvtd;_p!h6Kaj~H0I41HnIgxYGrb-k<sLf@grh<X7uicdHKw!S_=vfv2c;kfI9@}_
z{#@W6wi#9Ch@=Hq?_d0wG{C<ET;}Kn-avSkN3F7@LCR#k|3B8g{2%H+c-*EOAtXnx
zt#t6tnLEcSiIPNz`^uHpeJ$2nN2L-gLheMV+_}$1a;{wx>t1p!*6w22#a>^Z&oAFU
z;QRe)ANzRC%xh+z^P1<pW?nPT$WM*yfacphvZ=75qa@uRn<De)>dhR}5s_<#P0>bs
zabkQrzz2%9-qZ91ML}1WSn#ceC?QJhRjzNkN0j0B^RH*@(KD{vu%d=CmCDHlhOk2^
zJf;xX9>Y8=OJ-{!0;64+#igZ*7u!44;$TPCzQL&^)*l_K37B}f2AC2bIqb~)%I&!4
zT6DA=Am*iL6W^E?IGd#{6%$br^c``tH|yWz!VUn-ZGWWcBeW=^GVV#vw%acZ;*<pY
zV>+*Jg-XHy79P^+2+ioND0iN$w_EtO#hs(|M)6AXsRBQ7H<|v`?}P#NUe(>cMR?I*
z-#H<&`WE2SU%2;csSkl1mFWF?!Tlji=lL}FwOlv%HL*|J;n~!jb3L3t#YT3F3M5qL
z_tRrEw6v&R?(!BUy<ZsHCql|nbWS>-1G?(#c-Xfe!=8XzKrPN0#A+@+AF~H8XkM_t
z@^^=Wb`)hrKQ;0?wHC}O32<%-T(ou9`F_DY_|$vD)FW4VBa_?1;K*kYZXC&CKO)BI
zsO#Fc&Bm47N$>b8K}x!5(p$n%dv0I<VwhW+&sTQReL4gX8GCrl`RcOf=NlnqSE^S4
zIEi!Emrh5BTUYW8Lf3>MIUalH4+0%cC!^01;>&enVSp;(0ecP#2aU5*;?p?bhOTSg
zR;k3+6L+&m@3-aQfimY|=l+)$mrerbG9a5f!ij?pu;H(+;Px1gTS^A4_asYcr)a<O
z6g``?!;KYQOR(CCtKz<$BfGE%u$TK6^<Wi1;*NIbsN{aJ9D8P;?KzNk2|3pRqvP=C
zd<QP}Lva;n#U#*#$a370Kme1m(3QZ>(ele-1hDly2S0$6o!j=g`%+}a^gG+nR|&fr
z$Wu6eb5x$FIW-amuovvqyn&NWJqaI3is`$6>Noe?@n%N4x}IB}b~z;|n9kj#y}Spo
z<LfVYkOY|I>)+#Wka>+cUM=_l-1%Z3&PHXcf8Q!STqKs+RIRob6r5i;<d`^ijVllo
ztHwd#a1`wS!3ldF`)D6t+66D)ww%CLIkje<WLJr~Vp^_2slE=foiUwaoZrM3_Bco<
zoOC(p90&LMpKzA+z{}6oB9)z5>J(X?>ocYI^SkvI-U1xCAE>B4vnO;ra_>Z?W^(#!
za^Pd%bh2?{rItYRcWn4&%Yu86M-;2YKS^}9b5wA<%6!b=Y!d90G^#vt{|sw0ZgSJk
z6c>rlFF#scPLETadqyY#Bj&ZGfa?2wahbAv?J(-5ogH_eN-P1;Al(Uz^;0q{11v+&
z+j2e?d_ZYTI<%6kUDE+gH3k%Jp1A$a`G^xCDj@PX$ANzJD)92=rH&`nDhK0VEph`b
z`ZGAI1*^w7n#B4U&0K*Moaf<GqhS}DZ1BUL!a~uR;8;h&51O%#e0AfoZKsxJIR7%j
z&O@!=Yi{u6Y6loToyLiGNuL9AIP-N&c0S>}0IvjOgC=4%*-se?z%wft-Of_3{vQB8
zg3%=B1Y>NG{niJrT=GBnlLs5k)N&*vCpf1Ij;QMO0l^%FVI1WEPr+CQ{v%cP(`um>
z{0G28Kw_{EdZefi3{OE@iX{Ssjonvt%^x5E{22AJM-nB1V3^zFlTQT~h6<sbLaYYO
z*c>EsOV>J^QNt~tfb<jiwCa)viu!bt^^^Mp2H?xshE?al>lss=`3ng0i$%#;s&w@#
zPM78W{y=boQA}aarO4<#XKAu?;KRsO7fRIumI3~szwDj`7aClXeSZf#4=6k6Tf`0Q
zxg3S|<q!@qYh|nIDHAz}Ke}F5$PW8=aAfYF3Ou`r41Y#u|7G-Pv-QdFsSd#4H22ZI
z=h7TWS-rIYJY?bTzC==40gXuZ9kyY?LDn`)n=F_|W0F7M*zD~$B3F8^vQx>RRQ9K{
zNF2OBVDl}B^T{ep<?(hSjMMQq%_Ugy^`nC^(9k4j<y{tBLQ-;2Pb`L=T}m69V7-lI
zXAjblCXqG+E=TT;&N0#v+XJM<rZxkSY+?@$rx49{(PCet!4<<)YC`bn0Vc;AB#Oy6
z3R9=pMZT;$ZwEU%J1qikJ?+rcuT#h7i=kthT}bSI6ru{2RmpVuN#(fb$u7J$MUma|
zB(j~pR+At^fcFf&XpLq0oy2~;8o<c)n`!vce}uI<Mzi?9#0@g?w`UR8&~SNk<YZG&
zF^WFjq0(UB3JXC75Vd6wX3!jOhYOZ#*=2rYWE&J2HpHs(WOP)jckZMdG<I}!h}cv*
zDjY)S69eg2>TA#W(m^3}D(>(JJ<wbWJLElX0?*&s=FZ;ijN=UL{guTRu>dVDE{-<L
zBPdGDQj+nkPc>vFAqxMifmF7_@?Wx~S2+=gL;+n+4VBiKp)8dpolxpGa{b@LwwU!l
zm-MwiPKC@uh)i?{kPu_Di8)*atx$k#(Joo@&jb6RRC1Z@(u{{x5_-QnsI-0F2R&5l
z5|N)Z88Vv%H5_V2gt_$3v!^?%6bjjYzEhAxrwf5BdpXufj%)>?ZS*4{Lb*I#>|!mx
zUbg(8uGS?G2R~@jzDO;NVU4|4{xj4Id(swe7$F9onX;Uor5Vs}=+S>ia{1nB|6r^V
z^&e)^kfyKr(VP@<s`}kQgmN2^VQR>E1`|KA+CkV<Ax~PuF9dqaXc3X#S3)NrPS^M6
zF10P+50S;hHq>CQb54efwS|+i%1D;k>@E(O{-V@v7`fd=6;~RFk3itzGn0u0fXZQ*
z%^ZSE|I*8T86PG4D6~8kdOl>p2_0EU(Z#7l^(l{k;+s@x`#bscxnLIEOx&l+2XD~~
zYxMaYnM3Y50`#&BGhBiC%@2hj>zhD*by!A2Gw=*uj)F^n*eWxq_myl=s9q77f`S}_
z7~=Dr0<Kla22}X0K$8u~#LG*D!>~vabgdbuu!CFq?j14+VWiSQ#j1E6%xohEq=0gU
zjE(r&_#0M%6U?T6--g>i^X9%tUQcTaPh1#AH8=b^xF0Et*Z)mF+CtIzXq72<bXE}V
z3B_|5rbI36hC9R-4F0d|%o3lKFb_jtdx}i+QeG*m0NVEDYX<p!JL<qL2stu~Ko!D&
zJ@9Lg56!C(YbG_{*_!7-VRf5A+VHFn;A9BOn>tHaY;{U&k4A#Pq|$!4L|sjSBsI0o
zi}Q6P7?mhJ8;O3{Rw_$&8!I4dQX8UN90+5p9qLrmRpfjfJEh2;?__&8?3E)e>hpku
zd&gkiKmCwWeU&5Tp&Q$>s*QNM5VfI+tVxrjxTG5!7wpK8gG}sbmnIS{&W98bhAhAx
zau8j-!u&&oHYUA2lfpO5>}LQeBR}GInoaw0?w6{a-Rvoj*N&9;ze3_!F%1dr<ung8
zqy~aMg0+#4s(b1~3%%ovl(tm?@7!M1!Rn&TNx^6lC~4d4j7LBcU17)Na?CFwC95;R
z%%A+tBwdbjpjt&F8WmZoQWuC_)u%=@d?f#%{alTC%h;yzjXyOxkEMCb6ADl{kraK?
zRa)+oJX2Cg&XGEfbSI0+Y7dxg{E0xp-}|ry^uCxB{HcWT*qYDxM|b$F&<tPd;UnMC
zW^vv%v^VtZNDzR<rxaCvY>KoZr<>`K+!k3N7m`Aegyd108kPoWTGXx(&^7i6T@9RZ
zWUoLw4Hgd(I|Ar;%B9%DUCz?`8F*cd##_y#B{^MZV71qLTC*SoMF~b{bs2`S+E=N#
zYLSh-58*`L5yfH(`suv;t}gj0JnMTb%!9Q*6J}p~&7+m%EcEk#FqhFNSW{xPU<#dg
zD+}pzMRUx!0qp%Td;%qfkDPGCHnskPUD}kRAp1VIUQH!B8~*9d^`FLTQ+gx)a|!)(
z-e{X%!)t6GUCr^LHuu%$0EAB=F0D}u>qnV`pvBQN-fcF><!-+wv(G{0=#%j7|C<>Z
z`}hATGcpJH@aS<t=l0axsIbt3w09>@|F!xI;I(-6b6NgBp4Ybkl#hL?A=Esk_pj`^
z?70wXA=BOFlE$4NBcmIgUvoz+xSyI{(+EvLPZ{7zWEQENJO{<Wiku&Zi3)_EX^*U&
z>i7$3t>pF}*b@hY0}M|%>EWjLrpVFRM21hUeqpBIVtV`+BjLaCgGlzn3QUP%mcI!S
z^4DXJyVY7!|H+{)ay+3A14H|p8@`;uOp<lkg3EY5_N|zX-KvuHKZRyXFX3{@;_Q2p
z)$l6<;;iy;mgNZWj`2Du?eHJ+O4tct+VsP!^l64@$>o5v-v>&SUec?KRMK<uOh#^6
zheQK*FMKy->3>JZq&)H*e(Cl8hr4>w*^-Sl4eKzElN#BNWMjl~c#^MVWG|U{_DWd4
zSIc=MEqNmOqUidsJHIUzMpUg?V?x$3fvO6iEq)E2sGJXE_}ke@_Hm8a*?r+@Q-a`L
z!eJVLe%QIh?B8ogq$lA*FPXSBTsYyvdV3~j{4FEeS!*KyJ0bD>VgLN)W#X@CuUm!B
zXeNdBnhG#4GkD6HNbBiQvf0D!z7f+vwtIAW(U!Zl%RcXSl|$)hxL_mFo5lyaw0B~w
zzLxyHWh@|VIvs3Ae3P-zk@q(*c1@|?ly<4)GDPh-NmnI~9vxUprHph)61~f)rgOxr
zqlOt%K*{q{f-9^vW?X3}`%X1b{$Pgsm)XmTS>7PcmuL`Wt@lrah^x8J9dORmt;tgv
zujjF(B>Y<(25G}fG}Z&tQ~xW1DgJIqy8Q5xKdgLDxHf$E{4QYE&_^*^JQ^44d?l>Z
zIbq@^*lccKe5A~^<L*g@B@gWjDjyw9y5NsA`sp3wk}uN}6aO5+nk~R|_<?M9ys0Zs
zskE|MmkuPTndR;h97Jcj&gGnmi4<4(-1rlj8#2(^hU2Q_EeN<0aAjKS-x{Y<L%xP?
z8vi<MJECTud-d%+zWvIcGHT!9EnI0ITVOAs&i=97am=2Jg8ZvXnJs8Wa;$d8oI-Wj
zQc4r(kG$ZICM_4Jq10^HmwHkH^SoQ(=V{<>HWNtywZu74KZi(7tAlp6iz?V-xrmS1
zt@?R5EgMCe7=qct3+297^40{I5hWf_!49ukGx&1Z{!J4Mh#w~z6waGCglu`s$Aca&
z>Z!6Hbj|2c>aR5GHoVPM>C-H1!LIT_B3IqoF>%=#VRevYO;y8c%nt*aX(+uW^M84P
zMZGnkD^rh7RAK7~VHrs?z~clhS6t3i2t1OS&DP7QQH4F+5WwXuL_QDb!l})Qwg3K%
zY-kYv{<$&xrL@1?E8<sqGqP>N-Tld#wnp0%0r3W9ER?+={^})~XC_ps3HztsFhIC5
zYt|=IZw`m?qT0E_A2k-1R37{_d^m}?Hu)=ty&yE>ykE>r*TJLU=W-7W5NUrbWFA4X
zrcB|*6oPeC&D-4xds=R}v$rUvrz&YaHR@rj3N6z#z&ujFYRjT<0qMX4{~C|&%BM>D
z=$mu0)1#C7!kj;<)9g128dGIv#_FA$#`c`Aqur_h^B5kkh2m#{0Q$YDf=c@LXD&WO
z%ERj5g9zM?jdx*A7BKo4mX$Qq8W-je_+=GBVK6lD|LSL-)V+Ev-#C4$W1^%kb588f
z#NB}zUWDOor4RG5np4$tXSlvCgk?FW+b3BCOi0c@{}Y%ai&JqcI(WL8LXY@JbKEd=
z2?L>1wtv*F+N!$<5MY{{<g#W1bX#>6UKSfY(_t-5wxaL4p~KjpqEI$53%u4rTX%wq
z^gh>By|_}<0D`;9!BupbK>JB_aM_2Pa%astdvJ0Tva>l~LlT-itk9t43jW(vGhJ4)
z&F?4b)44W{e34s5(!33E1I<ihi|I8!DAwz0Zkn9PMJV02>e{cNdA~Vt(h4_MS`s|+
z^_hwQ%qF%meUs^3cwmk;Dcc-@d?^xrVljDFDWcr<%isT=6g}xGd)0Z|->%H=Wf_s~
zY-NsnQqxpFck-F;6ZnScOkOK?!Jd6guR{cE5LVh3G1Z-qVW<Uyw0`@-Z5tD(Fs{^$
zXP~^P($-1JRF%&D83j5k$hxJOPr(tJ((FL(@&+=M@S-U>Co{oHA$Q;fvnLf;gxU*{
zdUm`iQDwgBzF(+dCRjGEhNr?ZDRbMo=Erbx)nZ#LNmCbNOwpiCp$r`I4}Q2)51cZp
zO=;UCsw=6wk$Uh0c4Ch9Ytv<VRHSWDBESi1UAHKk01IS)?Oti~e7@H5XhYQ4d>q~F
z?wuuR?}aPH(aL9o%Wk`0ExCuV8e1_`T2i`{T$&YMWA;(SEk7zdELb>%n<CkQ24Yub
z5afNef)&GzW?=sDJM^fHn22(Z!v87`e&vHZw)NRQI!A-;24mKY(SF2}^R(1|0<>#%
z0iv8O?J;y$`)u*j>58lh$%;(ZL+m(?_BM+BZr-|FxMLzB{QGovY+Mn<yj=4BfYOA$
z$Kx#STEThEK74g#V+x@eIuSaMkI`?kk;ymvaY|}aJq!wkySPmWdKEF%Q8L)r(*6~h
zcC+o?`wK9YNyq|=2*biG*(&f8^X4VaUJ98>MJ7*@?<PeCvibdL!22+Co9{E;2_BlA
z`R08UE4bgr?hwEpb<JMIBR~30V=OnFxfGgTsL=fDdZ7v4$$_wr`n3%k7k*w<LHpSr
zSQ-MC|Lw}=PeOJ$0BA#0svN#yPH@7tqekxUjeKS*p@WE=ydT_)Rt)|74MM%tBzm?(
z8Ise`f2uLr#QKj`g4e(1q&ZGww%E`6*52=ZhuetGJsZLJ9lJueTn8Ug;JZW_*?>YU
zAN}53ffj#s;?t&8bE@b*Ykq8b2R&b+Bj9CNA%`|RTw)Mv0~%<#uw78=2s$|CO{3BG
z(aGq*Pf8Dd5pP7zuV?WqX)vOS7K51JwudSwo&n5joI8{K2(Ls`%nU3gckpiGQ|cEf
z3bU}E%gUR5aqYV~DS;TFDgwXuq+pjK^xLtCxS#FJfrBcb>wNC@C9O^`&~55l_OXex
z?FV<0y`426(|@D)u%)ZH33yF;CD;7V!;Qq7+&E5OsZpEzwNH-GL@;qo$8u#zrXLzX
zR9!qpxj$Ft3Ucurj{ebJg-Ino+;36Y<aQv$m|WMJNAAqYm4rSNxBmHEq>)@+PX8Uz
zrozyq^Gv{oT*A4oPVn7L?kz|}x)|?ZEh#RAY0%LhfEePY39-Zc3NtFMlz!5^8DAZ>
z)w{F0<>jtv_i!>O-wXoMe{~7CWJP-?tT+Vw_*8}Qow*0&9C%sP`u_CCy|iK1j%<tp
zI)>Yj9yp6M46jA7-tJNCDV><VAq!V>yv?E*B`nRO?XeTqiYI;(XUK+2Sx`1Rn+gt}
zhUBkcMp53|UQffCcMYUF;GheweDqP*!sq#84y119Cies-`rA;dBxc3w$m0*wx=@hY
z^wa6g+XFE$-_2~c{&)Nb-=${#CV0*~dU1G<R5Ba#L}1b}M(Tp9h&tH6GXLXxG^t!8
z*Hg8xQ)Va^0OoL_^!=bH&#Id)qHN^VL5{V+Zg*-%AkdqYOIkQZV0y{UaRtti`)@+8
z(}@Bs;ptpP*9T3TH)qj7;lT|P(h97}@4C}V&jq&fxY{zVFADo#)vv<a4SVsrl-b$;
ztY6O8%hyiU7R~8qsGB+zL08S7s}}UgDFfzN>pR$(j|W#<V~p=MhI+0#kkn^N!yarV
z$^E1i!d5NY@kX^3k{Mo8jaR&%5C8dxDEzqcE<j3POTTqWaX068E>ZTQ$p_7`X-E8U
zbx4Hec+N44-kBVCk;x9HChrv20sQXnq^#6`&6I4A#*5ic>PY7K8K)X@PYV4TG^YjA
zxQR`ygX|T4`TNZG^3bA6&Uq8&ux8%J^O|GIg<1)hK$Tg*i!7LjyKl>chEGQrO`%*m
z5R>ik7$QvtBpwi`s^pK+_pi(fey2yaOpIZJE4xvJGdp0fj`}k9VeRPvS1JnBKwT4b
zzA$d(e|<kvGuiuan!g-CD%&P<w11`FGGG=M=X#mK=7Q@BHdU)Ikz&Qvy6N(uXU^F-
z%;V?^X>_3{EwE^WI+D#Byy$S9LVIt$(XnSZ;Jgp%k)}#b>mGF~Q(wBw^ODADZezdR
z=`dVr#ioMBPZcDsBZgvfzQ7+6-+`O*2J8?mmdh3#(9sBM*he($fmyI+R=izR&p{%S
z;XjE9a94LZS?ls&)xltkiVVT;%u(ny^K`pfob}At_rInn&ELu{8sk=+`skb=ncA0L
zS$HjW*!!oc`&8ynt?s#6lns4NUP?~Xbmd2B?mOT^|HK&PqxJJJWMIEn3}0?!RC(X#
z?;1l8wwM}NZj(0A<%7I2*(l%*>;xu%k-R^$K&BjLIIVJ*jmlwhS$J{CaK7}XJ0JZB
zj>zHOk!{{r-{?A5X9~2$CSrrCx%lz;t<5G`1W|C5o-m0DXa|;eMPJR0d?HT55J#PA
zP1)hmb9h!k74c<bsT#O>0I+#wYQ?A70J9A(&@6#f&3bb_RtZ-BTB~^7A5#zQj!~ra
z_R5cYeF>)DKpDELIwuTybsQ0V*L;=A&~cd@1Mg<Tpfd$%!#uobRV@})3Js;A;6MLf
z%=FwVHy4?z7)^bQun3J27`{9m87;UHsXRW^oGnG{Yt|5x{B?TXt~lQl@k-NGVw01r
zCL71uM-2s84lh1wRLQ{7ExPlkvPy!Wi%92s6-V);Y{;}a+&|`K6PF^U{jiO9b;Kv$
zJ5wnl^7p;InPznL0QLb@V*c6HQs{qZnBRnaxw2jca0X&Z2$}h201>C=+Fi=kv+LVO
zv7I6Gy<v9@ql<Y<h{78l>2J~FP;1~9T^-!(a|fMLbwUH3nfzrcyvM6`;%DZt>Cc~2
z%nel0@_#7Uz=Sf&x%|L^RO?kRerJdLX)>!H+|htErH8E&!rk2vtOF4PeUl|(aC;d>
zrY|0EZ`ZKoxbrhBfm!{1DcBtqWJ9(tvikTjXa;t?5qRf$%GYW_)jDQs6kLrY_O^M_
zm3|&E>*qnbG|iI5Y)>>=F^hv~4W{s^iAkMSNAgBU-}4|B)wc~Td1|ZD6UF&<4bE1y
z_a8O!ne@OIz4%9-Bj8%Dt->1VSP=Z-Z8+;GSUl4c_oj4Mf11*an@l-l1sR&Bde;J7
zpsOHdTj3eXNWi`7&cVA;ZOYztx`|;w)P26raYCv^$7R!{QV`?PMc7UKftUhTAIo|d
zPoSqK!Pis3`qKgH9zeVT!GNMNLCly6NRFJQw{X|es6cQQ!3jo^Ft(aBk8ItbfaSp+
zv5uRyAK{OC&uQEyGnxBZ=QffM8AKRAntYv7-%;#`<V6$L)C!{WxTzbPjf;1vm)R06
zGpw}Q6~kU-Z0q6dsC60~y|xfns$c_6MCQVT;2Vb0#S`Vi(W~p2MgBTN<Af<%#)fev
zWeuxLTuVV`w)z%^BRnkU$PMY6GPS^d{k-<E!W_h@XO`^`HQ}-?U%B`yOz-eQr{zNq
zr~U5cf@!bJSqi=`f{l;QTF%UO$G}}`*@lwLy!^hYX-jP6@A-nr=T-Z;UdXever&b0
z)<=~P)Vxqj%E{C~;P(SQ{QINjkjwL_Bz>{93OTeLet%wvf)#}bLWkHv(TqW~ED9<D
zq5aJA2BuJwI^O;@Y-?Q!*clGOKr%U}_#ob>$QQbadx0T270_Q13bN_})un}{C+7d=
z#e5YbsBLnE`a5`Af0#h8RiX4<$f9(4^hz<67AWXZHCzK$ZS-$Ohe{KJR^gJJUh}p<
z|9R7ToOi1OYilyY$>5GK`Z}bvY9#?$GZ5oc#!(A&2pI=uWGv;gugl(W+ccogJlS<j
zu3YL+gG3=WQZ*=`jXsDM6iS7=(-PLKW>2952=PZpr*u?2C+pF-DO!in=_8R13Hv-)
zO{d8F4Z7@XEplVU3vld)8wrd!by<{~d&?a%V80$TG!xFf1Vhx!v?K0P1KEkDW@l$J
zyWDLcH0#xlhy;I2P<m2~tW?=n*=krLqif{qDk*J6wq$X^H$}JfV~A5@eTu-IO49tP
z^W$eN)1<hY$T0)c!cG{w%w`@UaD~)8pW-hr-a(2-qdMyI2Ml?c?%tpaXex`PLor?3
z!|sO77^0j!cC+R*;chAiSw5^fFf{5{v#UcwJF(CVstt$Ja{CKFrPr4ftYba}7}5JQ
zY5iWvO?sl*VV*gTC5c`q-yNLNKIjWw74gM*#i*hf9Cf$ZoG=4ol{;XQ)-q|%`g~eP
zT(NG|Fw)C~7NACbl?|CB3qB{elUj#6+PU;^Dw89G5XAJAkCrrAC}Oe0(2FYD2Mt{X
zr@YN3gPZk;^jluQu9g6-R59l2T(n@+m7x}u=an$nGF2FDh`T(A8pCv)@(23QX-iXl
zNSPr0@Re5oJKp4V;IBQydFubjigDmyR>!GuQQfex@eW^0mCeJWvLwafdRn&%>0S&w
z_qQ`D`l~*LIm`StX<A33h3}~(gmi&%d!0=9=V2%2uW}mi&qn88=2_nJzYM2gF`t%T
z+R$eZm03Z@{T6kF{eM}!5XgRFkmZYZO^RjuPQoFo@97+E9-o-QoGS<0T3WYtX!>$O
zF3!p_m$WD~jg8(c-r%KJQo(B&OAbJUe%7nlq&qMFq0-H|*cbaGN6OWK6VBNOtXZ8D
z>J92a+yt}WnF`%yW{wm2<mWc2WK!^AYx~c_QZ#oMO=sFPsEzEU!cZYcoKv6zY61~S
z?e42gqI*<1aM0PAHcXw+;-JTc<dW~~$DC86I#&?{)KSsw{>`PxHt(YePg8nc$iy*V
zLC9XKKk9rO!N?IoxloUHvGJZCHU#uRi#CT@PX%<4^>XC^cvduP6dnD7y0<Hdk~>6c
zXTka@3<EB9ELCy;urhH=v*?x#-^FMb1lB*>%P^M7ikRY_#OH2umeap#F8tGWIvp-8
z^afgf^i(2FsiC6Nruri0Ajgqe&T^TrDP5mPdqopaZD=DiT78O~oB<=R5f3B-4^(JZ
zJb*$XbTcnQ2NXl(Rko!T6c0TZB~#`a<NjC>b73H%diVn<Yjy*5Ls`A>9#xRLQLqLB
z(F3;qq}(9=i6$6@OfI2-l$-+&JZ*H699E+(OE7f@9+iBQsVxf!N-{ZN0z96;0hG7k
zk~*l<b9KqL3e}F=1W9*-da11?;bYhrvQM!*nqxIaZJXoS+S#~wRJB<=`I;$`W)hWb
zq+S6av2M%?P@k>}a(_I?9gnd*LscdnUVqTm5sBuuAa!ldd3WqQ_?S(Luu$P8%25UX
zDOUip8Xy+2NH5?hs?<y!K4bW<^{2q*D@s>gLl91|pOguI0P$Qcyr}Is9B_l4N@@V*
z3H~XRdP50vjez>B9u1?Q6wM>sd5OiLSQ;l@3nzuu?EGoDfZ~o6R0*O25PR-xl?8&~
ziIgsv-+nTmNM6t^ERhq#u<#<Q(XrtK&4l(kF2~6U1%I%ShdM|jMjhB26A({aTJy(>
z1T2LOM_8lU@IlUPo9$R8_>NmZ6(FnQ4MuaRgFOBwTkxN3R69)vZLUT43bknf?5p=I
z<pklfX$FM>y?cfuP+cWV!NC~#_*3|B6Ajhc9!|i^`mZhpklT@vddp9(9l+OnnoIsw
zi;**ljs%#uVU9avi;DF>2CWfnYDsrizYh=B*CNDLmtvP6MRkCHE2u;#GG_tKBqjzP
zGBpJ5JkjQ*uLX{US)^d0VMpd)*uFHKEcP688vCyC?Ldu@XA7>T9cvG)+M8dJbV9;l
zNezwISX(D`l0T3SQkYF++O>H#SRQjgR(5ZWt%XY_fqR4I5Kk)*Lnx?oE^wj)YXMe4
zt*jMh$254<(PJlD6SEKGiJXNqI7jMb2uw<T?yvF>W-JmKTR-o<$|<wd%;fjK)wi$o
zYaZjcNA9W=e(lAQ3q*KSAR<;_n@0eqyDj&Fr%eA<`f_fum2$wn39a040Spxqg2h6$
zX_*1QiMf*#9O21EO-I+7xuJ-8Wi55!aC1AI-l){?Y|p9uqRnd-H%C{0e2<F83?sbS
zwV8_Z!+yZ45SR>4NU~8Ngq;m(6x@*QnJ4E62-1%u8$`q*J_0-;!%d~o&ieTSa!JqA
zcElGLLd>t~$f&@*a+6$+b4qY_y|Z$-I@_#=0J}D=o!Cmmae{1=)0T{EAePY5)@_{G
zNdwCH&AAD>Lyz;k;H}`~*s}IfwweuHswO9L4DYvJPD3AM2rQuQIxg4MIxHNCFa>SJ
zZG{|Z^hznLd$wNl+?wXnXbXM1&v!MS&`!p9&c;&}soo0qWO%YsNVBO11%Rc`Q0Kv&
zzd5k)Ov&?(>Oj|aq+gXk=N4oUG8~}_88b9oqu0548}>)RXXa;+rDXizOGp+Yp>oe(
zoqTv!Px#1lZ&}s_jQB0&_2CCi8qj^+2H#+q1p?C#T2J0Q8tgTAiU#gAZsv|cUjEd&
zgtK@Zcytf4Dz+FN;E&~b5sZ`SH245}ogLuCexb*$VMFLLwDEVk+6TTV-0J?md{34-
zie^+i2TnRUaVp0Uk1X*i<eP861Bf_$&R`RzT@3FYHf|yG9lJ!;T-sa1<&jR&ofo}X
zPjZ;@b9Qsv?7lAnt#2t^R~wK646B(VY-DyL+o^4Y)EE!@F#mN8=ip<n*TUs-4UVuw
zB({3uu3o2Sc98l;9Jdq6u*g16Ha8XinP4<K>*VdsaQJ4`)aXIiX)c>FLTpeRNclmB
z$e18XbE5}ar`3GLiCH-__tK@l%bk7gBfF^y*MLZeJ%fzGyJ-g};@D1y)R<ra{S4x{
zN|M#Ib^?57+LWlAbG!&XdZ<-${AfUnKl!cJ1R=_8juz33YdG>UNq39%Uck3swQkq0
zCQa1$(#0FXF7F-+L-k&FIx$cgqs#_)yti{*=FAMsR_a~HZO6I}P2I*XoX(~Z%)~?@
zu@jT;3@0AeiVq-h$P=4<B~y5opD#2}T=b}u==n?b<+IFL-NSjsIk7x4#+<ZW7ofm;
zQ(KyxwW$Q3ntNNxxy!r|C8PEzZ9)v*@j7ZLoP!Yej|489WevCk#UazkXH>>MX44-c
z{t7D?9dt$Z>Il5>wX?S~)tJmdr!TwE)4k^qxlI4H0DYzf$>sZK*r$ady=^Av;U2|7
zFSSgs)uE58(Ho>d94l%SprUZ9gv*dihx2N_od4QA4k@>HL(F<K9giS<HX%3toHj49
zIY_eD)FV!2IAJbtoCEHHz}yHyfBKa_titcDi3SwGYXm^?3}8-s1dLYM#->k(r{;m-
zeD#>b%k(BaMsX0tm)nJL`s)0MB&am418>HLfx}Z$rKxXdh-SLNGl5WVBj_$7qd8K~
zwT~wh7!m_EjW$@%+(wW$*(#5CJ&oBx8SMpd)8nMi*8){I%bK{rH>Rf8teI<*r~W&N
zx;~vki#{Ks?($EN_K(jz(JzZ5mQA!LUZ^<L!Zq_)xK1#S&13P>K2bG`4p*<ux89HH
z8{pvhPyGK@GyiTKRWkwquNBRRJR<CB)293cHO)-%^qZaw@69fXojv-MdvAO3^w~Gb
z@7}#jR?LQ5(ep?J1?vZY*?(W#hpZbuhGKcEmC3Ejk5=(8tpbke+V@Jm*xRm&#lMaP
zI{iKY=D@v-`9*muvk-Su2+(j+YNX&|7i2Dds^&%4_Hc9vW7{}@BO`b2#?|p4^@gLU
zQ1r|~ME+6h-MZ?QTXWS_5RLcrr`bn}B`(RFRoZb)A&&Vo^O%p$H+|I<7{<bsj6Z^c
zSO0=%<Vm+9?_bc_%k+3)jB?Qx>;GeD$)$SPw@jl6amVi}i$w5Gg^*@U#Sy{3tM`+$
z7XbQJKN5m?1xL^DUaiXd`O+*M#(!8;+2vW(C<6~dG*`yN#5*V`7i+pFywOseIBfde
z*%j}1szBy=nB}K^@;BW;$Ne#;zKWEf3aE|I*mS?=`sspni|*VLfeH<x=R~PJc}mjb
zveyDTuLvfFnFqNZ>!=%cX<V3I?F5diG-cDZ7#cGAF&r=X>SVRGA8ts+^)23?8^39V
zU7R3N2&LaXgjp=mpokQ@X8$GSn_uq_yycSsUzgwo=icvZ^NZ7+6lCC!{Z|?m_xN}8
zwQsMcF6oVuh2wI|WSX3AFS=pV6fk^0tW7SN*|@7(cysm|w(cvag!(*7_xiXu&O6#<
z`^s=|Q2N{HA(NrsJ4*_pnSb>K71y<>AsW9IJK+<Wxls#Gw?HMI?*g|LsE$(lXDy{I
zE?LUSYcFdpTrU+f;LZQB!5mBCw!2YG>rOBeXei@Jz4s|ZD;Ldakzl*ge{j>?Hve+p
z$Ghz!p~pR2l?SP(8Q(Cx>dF$&J>@@OK!5c?`ATZ*7X$Mx0R|jQc*GgW)9?7$(@b8V
zlm6Y%?xMyg-nluh(RUQ}mzy$R3%4n!te!iK1YB}>;5Nzxyw+k$*q*jS4(l(9{{RVn
zJDS|V5qrBpoS~|vy%<*k)Vh0nJImBz>*WbZxIzv_8>A?`Yy?R;nat&HvC#K*PQ(v&
zVi0Ty+FKjAH~pL}SSxA}J101O36fy{+9&GvqX@1zq1_i3h2yHg52!l1CXK0nJNvxP
zZ&+q;|HU58$CedH8^*BGUaib7Oul;G!nJ-=_ptsS?<>BEl4ClLE6U8JPZ|FTuY3NK
zAb5{&Z3S=Kc8nYwS#_d_4=<anSK=rBobQ41-xKGu_Ls=Fz}uqT8xU5}W41|mcdYvE
z?6<BD?=#YLA_$K=U)~*LG>4o<6rX-m;P|4J=hWIjVaioKE*SUmkP3<0LQ67M^0MZ4
zDoC`iwsW~JdcU1)R^uBwWL-5!9DZB#uzd$~(!1MJoI-Zzc^z>vzx389wsBrisNl1R
zt*Z}K_T1(-!l;muFA6~q_OJH4827E$`RK!y7v!j$rkH|HRtBQI{$4`M;1{<xsO6q|
zC-uKE_r=`C7>0`V|2`MwKJO)i6IE{(It0q^{kw>i;b=+h`}BT>iV9?E@cR#)9SU;&
zVB4F1z<U^usi&?<Oesg~nfbq?CkF%<-tCCRXqInxEwIfNq{F_y+1?-|W+C{z7D~nV
ziCC=LrbNc8HxBmVXRjOT_gNU-E!L$RV;(F2cX2GpNfc1T0W&arDaH>zKEhJ(9h7KU
zsybNz*%FGhN;iPV0~C3+d;7TRajneF4Sa||k74T4XZ8IC^8Bop!`D&ojBi=|*7uSp
zlYe|W5Uv1;H_G!TUxv&nYa7eO+=TMPJ5`OH#Qw{A{MY8#Pk!y}KWgyj;Fx&N5ALR|
z$Y8Odz60RM(yd8lsnZ6#L*{j}Bm&QGqdA+oz8Z_uoOIQd*i_CXI&p2eO%?6PHN6u7
zPo-3bV}fw^?LYGw${p_&NO<vX;igFzZ}u03jQEzzk&`!Ip(~)9^bJ0{O+Wc=n>&Gv
zaRW6wvD<u2(e2l=uC6f`9IrJ}&k^SpQhjH28Ye=1B~D0p8QCk}Y>zD_OB~Pe+<uvq
zoM8PBw;R`+;AiOZ@lM=2%a^sgbLGB(+bfTr&!+0lmy%z#$(1R9*4i)^PXI!I(l}mD
z7fl0^pF}(9zMIx^O5h7YrVo@xJWsPO;Jqrkc)AxaO?+3vq^rwnZ;ULX`gQ(V=w_dd
z!@NEFR#B>aZQ12Ne?7DKuESoFqtohYUdc?j`yylZ(WDx@z)-$JLW)X=TcLN;9Z_);
zo@*<LQJjC&ro}_!q^_rd@AZ!1(Oq0c!$j%+Y#zat;v`drRq4Ob)7papTSLG7J-+Ri
ztTk>eW<ZUC_a1yS6zT<cL*9Y0?HxsYa!rC@TXC<*_opXJ`e-@*H@*+w?AGlLvk<DD
znb&w=(Cl)2%r`By*V_hBIhi<jKjX}}T&9bEcA_b{B#<+ZPtpFR>}mGnuw1SFblR^c
zUvtN*n^sg!%-?JO7v0E|aH@jlGqg2tJFfq2Xg!RwniTyMb$&6vkDE74)GXGJ#HM@=
z3(WaBP2GDmRooh&{y^dNex1XD^SIX1#585e^_)_U$K}(D{~0t~>ggQxvP;uPI;SN}
zHoCfK<$uHVM?aHkLJ>}=PZk-*zPNVo?1L7e!yq<Rn_Co<Qp?92KgCeIAJXx>^>((w
z&HA?tSH5q=xY`Wk(jDq?pOnXG3IOZ%RD{vJ-#D&6l`oRmZ)A=&_LcfQ)xRck+brEW
zsQRb$ujbqBvMWpbG+XHwpXB9~pL>>m<)$~7b3%=~e_+=e{*&w&5((q)@j5YiUB+mz
z%-SqlnR91~{I*)Fd%LaXmsZO2)D=Lf`>LsBprw+4&|!s^jG6c!=Hcuh+&q8a?&{3r
zPV}bO^_IJR2_5-wnFHba-ccz_@79iQpW9~hsNMcA$S+$-FJ0iyPVyNJZ2=JHyzD>A
zO?kUufhiBqzLS`N+gyX<0xQ|Pe-1~#urq^yPnu;40f)NWHv9T!iq^guO7seGC>CWj
zRjIR+q;756^CUH!Y!CKU0{#RAZEa3!2TH|(niBfo&6}5VrP6+1I@Mo#F-}kLUmHl4
zX;mI9k*rl4_<&Q}waqR2Rnxk&!3*ve%mHhBgNyj9z3f^Azv35thMl+6<z%~y`P*Es
z2;lFqx@86`!UTttjJpPvtdsqPJeR6o)c(X!fFU6nouqphlzxAOib(x}l(j<iy-*jm
z*7cn({jfcl$;n&&QNG>V7tffD37@)3zRx4$B6@8T8N%%jn5k0Q2{wEGMETR%+Aptn
zCE4h|6NoFZ@5%47U&_Py76N#G60bY0m6y(FmEL~l3#ob4X7uh)R4-Jk<$+5!h#$}(
zU_y}kGbb_ftoAb%*Ib<NEBy1Sc^SV7{pR`wi$AQN&-6O6+uMiY_~RorMFB{Hg17ZC
zzjMDXmTMvv&wVgSh+xQreq0x6@a}v}sDCMbyG^4Yd!zFl@7jVG52YU0{L4h+T^!f>
z%-SH3kni$G?$o^-m3yaN7mC@1*O5Z5`&rBWkT>OKSKH$vkFr}Yu0m9W74ApTsrbdP
z*LJ%~zb|K$BGx)De6|zfHMnx;6%#F!ptIGuCIM~D_IoXLPbKC>GvaJQ_jwkgkIq(J
zx=F5Fm>fmjF!)xmk8i87%O2v7jyk@(b=~WYlw5Z?Jm6XLHQVQL`+aS9qSYoQebhy6
z$VVTPh6{PA{pJ+;G9hr^Joa_Yz0Tn>$!X<6!WaScCwR)~SJ2IepPc0a#CBN9zOk2%
zM;-h3FD^--s)uIxXDP<#_40>=2bCPJgM?%WVVAY#U!0IWc&vBxSW%kqxv1NO2y6;h
zZQSQ1i(J~!4xeK5126t3dlR3i;Q04fa__h7K1F>{DY|HKKco4_Ni)tHSf6>i6`w(T
zm=aLs(W=#rTe|7R-~M=yZ`u&I7B1P@lPeu#3#Z-wekobfOtq|d9{f#kXmtDge%J-w
zrHPx9QGp4l@`-TYFBQH9d<jAbS<!(eyUn$}$U@QC4$V+b=F6{tg!!<3CJAv-1y>nA
zWXKj$H62NY-|iv7eFjOv!TcJ@oClj5rHnHw-b}$9lg{ANtj?GRuOz<u-^4k5s&w>r
z5QP?%C8l{x7BAV&aE4nK+FX%(m3R_*+q2qFjXV!{2jx392kGGEyvH8bOO?_7cKPCa
zJS*bReR}`bh-x8UTuOoe9Q6%_cv<*xI_$TRW{C0H8RPL9zr1arj+sQllBd@;xw$u#
zKMk$_;I2J;H>KJr0YOeZ=$~b;$rQDUY2=$nS0dzGPl?0C-3r0N*7yIhKV8X+x%2pZ
zH+aVe5F2;zU9LTL@7BlT_sui9zMP<;iXJa<@fYV|dT7i!y}BH4)M@g)7&m#Rm8$vk
zH=l>E$&Y&N^z8wQk<Gh~jg*@!Ctzte!pvyaMJ1nvxzNWr#=Qx&k0eP8Ac^mrzIGzN
zWm4nye&V15%}*<g7*0)@>(f(DS*<tdXOJK=>!px){<3TID)#;%BI|fbiF%+p3vKba
zq~S}#w=yV2nIn|1@<I_O#XmGh5+t%fNcQmAu6~L#)~cl#R1cc))}H+&TOm7s>ZM}r
z7pc)o%l?x;^Y&DmpRvLdhmPJb()45vTF84APcD5Cq0({(f4mN6lyY<~cWdOENeIa3
z1TRnh<ucz)jva_-A>NSQy*_wNU|QI-Q~7&206SNxdSBl=pRk;;pMB-^*Fn;-R`Gh`
zrLJG5D7A@+Mn4#-$oYAtKk-ld*h8i&zi|{GsgHQi_LVv_um5oZ^@W9=^ND{fXVDm2
z_v-ePSb!TqaxI+G)a*pqg+;u8+FJ_vKGvN1C?2HvdI}$$U_hEpd1VXJGQKHfqczjY
zk$M{^{>_o3p6LhGIPav(UTm$v&feAVUq3S`qIq{L<OL*t)|vmutwP$fZ-Vj3yYHf?
z@vRNTIHnFeV?p`>;e%6FRKEBYrA4hlD&ot(7;z3p{aqb7_l`?(m=2N3M)%uL4@#EQ
z9F9|_r9D7~_Sy*YTlI}q$b+h-=F>SYqNq6ZIS`L}<jre6snF;S_6g0}(xyb%kYNNg
zNZt+kb7nN&tq61?Hf4lmm40cb6y@7hPSx0rFTun|_}SdSzKS&^R(&;Ill-&0sdjqf
zbQ;ghX)xAmg85F}=k7D9S;anf?BUYk8%gt)9P8J#-#uHOMSHsYsr{A!*Ta7N%fX+m
zuMqm@Daj5(1x@!MQ@agrJ+u1_pY3B)IwRgVT8D<`0RA}J-zCnuIm!bDW&7>3ADCn!
z2yD_hwYNF*#V3lMr7Y0$HzT=JX%9J4@15_DJT+{^`9Kg=AVie2HX5tBtWo7ws4H~r
z%C!MNK~K^`!yVCblRh2Wf&Dwd<2@%ezKLZ17g@l33dF#ZhT6s?Rpj0rjSHF^oz?5T
zW__g3A4!0{e`hBaEwwSk&(Fx|;~yWVP@H*6!N+$6T64ZEmC7%7d{X@Ky6~DrA_J>{
zG3KUSm!C@2%aXQIH*9BVEGiscdVTQssTN86rVfDV_~{I=mLsr!FK1x0%u|1(h4^s2
zStvDra&!9X*z=IBp1^l6ns{)R^e>+cg{p(z74wuw-{(>Y)E0i4{r4E0w6qTT>sI2k
zd`qO>#+L`4CI%>K-+FDERd}q^e%eqkY(HP21RTJrAZN4KKm0JqztY{hJ!1C*)S)xP
zHe9WwTvzjruuHjQxvKfP#9o)gTj7SIg#NT)0mq}jBF)64rm4CTUrF^Z4;dFG^Ed91
z+F#qIZoI_jmpLqI`caGaPw}@PtvC!Wgu$rqdjl~a|BFa|N*#aGLR!jNlJl{+T6<vB
zYJWbJm(_@=jCsEB$L_;ixv0eZqPjm9WkH&6Gd}rog>Hnsb~=}MDa9y-Huv)SDUD}3
zGYL%}i&{g&HpSID^$t&>-&>Zb-B1ZmLLNVwZGTd#R~TdcpJwOLmNvBMGkf{D+;KbE
zYu8&Y+VjY}Rd}o#;S8@toh&%R^#XfOsJyDSeCC)qA+h0myOw?GXs1rJt5mPn+=XK+
z?`#2TIG0$#s1sL&1W31)zSCMq5p+=Xp{?_=9ljgaL25jox%DPd&s5AM3ZC4&vd7KO
z2)Fb<4LplEe*580eA@v39#?P4Ml~5VZv`~~Khw_c6t`HZ>V()@9RxBQ!^y+xXiF7#
zgYs|c^mg;f-Q@e~jl*W<)>{L<Z_P6=<OazI?tGAa<7_i~y!KgKn;InaKz8X9-!mib
zW2NPz>RX<tfm*El(V1!D4PD=p?&w955EKxQQzz7cE5Tv?{CmE`FZCA{f#}M66^)*m
z7Z4|+zOG!lGym$w26z@FE^NK*dtN>4P-kH}E>Z+X0ExI1*!*5^S4g<=)cDWd`kv*%
z`QRwz3%AcDnh$G+{U@@DYhMe?b)V5^YN$sdX0<N-5dHMWnz=E$i%T$$i?=$yg1l_l
z<bY_CJ9(4CY`Cu-2MsUsTsVORTmHg68t@I0M?5y;+#0Tbh7ja$o%^68)S_q<T3t`=
z|1}<eZw4C3Q+GCl0eqt}`QOmJ3(03ZTq5~_WOyN1QG>*1ippe;KD}kA*o9FJ0H40V
zyruUlT07H0Sya0I)`T%p{6=1(@Uq++pT5;~UpAxa4W_W@&P4f$*bdJ>$~^EEGN|3@
z>b!)f32(+urc2pT$9s0dSH9_L_wUlVwQKlT=Y8atd%LGU2Gyv=2or>iQ*p-W><{||
z#_766DLmfylZBTlOg3FIeb21_xXw*6@Usnk+1;)bIr|bN*TW}q0j1d%LvH7EgC&H;
zFLH>G##_#C+AFm^qky0708auz#_1OudFjrX+o3%O8dBFo{=JUagSnSZ(+xSrSct>v
z&j5(8#5zW={+&n0U;ixZtK9&!dD&XB#{s{Y<k$HZ^0yAFURE8+<j=T+I%h#%Vz|h2
z9K^MqI=dyAmnK2l(p1noZKguqL)i_hB{rR_v9ex$2L>wo4^7Z*&Dtxt2H$L*dY>>j
z1Uh5LmdNq&N<R+3U)HLHGZ^j8CUDWanSM#9&-_>qo$rT!G0v*bZATfGA9HtYCrzmi
zz6_Vul=;;ip}(9!an!Xt);Py@U!-G4mSJkss#+JYb>}A#ltO~;hKKL<E1CRsHM0v{
z`eyeVj$L-*>(Rte5n2tpedopQ5Y_Z9qz=gLlk6gHV6f&&S&@5izvyRhU=F{Bw7};J
zPFol(dxp@=^SD{r4P55^^eN<;MtSmA=EsE2=St1ztWNzvb#qvHWGYP@pXj=N|LqxG
z57W}Dp)VFwtS{Ru`?R7vyifK`4Qg%n;pYLtTI;!xQ8RT)jYO*${ZHV9Z5ixLk~lx0
zxhTF^&Rgrhh%dUpt;!c<Fw`uyQr{-^8<mGRdsEB`mu!hg3R*llE`0D3YW{4m&{-dK
z8!r0BMnFb?x&B_~t3Kv+$i?XD0K3>cB`YU=k-@Uz>GO-{R|=ZxTemT?K}JF_b+01p
z{Yx^fACm(_dya^u6pwHPSKZp%*=7k|9l9UAX*X=g$WGB>K9|Z0=2RP169sGRANpLe
z9XQMIN^rbB>sdM9Mmskf)BHd;XXTF?5todpJxK+=x~!d|V^!o9=(*RoO@6gTeXV&;
zHz4Gz#dKsA@1FT_rO}zNw<Td>x~sF?H$5V>%@UtvcT7xP{^}z5tt$V#@lLTXK*Iap
zsdH>qjW;#$#>j`XV_#s=te~mddGN^ACSd9i^|UhNI>4*tipmHhzVnpKQd_Ojd6v_q
z=7iZOIjlj?-I9<#SM@Bo%{k8~{p1ICOm_s0)8i}Ys^5&(G>Cn4Bp>JayewU_>6oT>
zeMSV^J%#uPX}UX0e|&aVuFCqal(BIyDAXgOPkQ&&{yo=sy>U9FdxUil?~q6TK!O%p
zFh%}jF8Mj>=YW}x@!+ZQ<MaCkzM<$BzEvTglCJw*xQ1QQY<vpvbadiz+~FU}on3}T
zxZ&;vG$&b&!F?Y<%(as@*N+I1zP#aZc2gV&y0Cno%Kz<DO9X8r2?<<8$|;O`y3C8Y
zch1_MXQ<b=gkBEKC;a;5+!7X2j>!qS*{t(lSMu7kSks7iz-l|zk#v5V#}X`Jh|q^X
z?@nF+wfEMxnKH&Ohl8>t@xo`^mgrF9+$=RY8}&8r|8nTp_1Z7HNF<t`nVE^NyPRhh
z{bp`tp=(Vu{@(jHY<-P^1oDpK(-U9mN1A#ioOKK2<YfVUqV-?=!)5)%+^HGQ>||BN
zPv^@f1#Z+0O?I^5sV|p{sPD|1cnbYW^8E6<dU46Wf|K&wEolM9=@lh8jJ9iUPVxY0
z@QFpagkOPIF4o;n9YPwN{{vMm5P$MQ_^X}G_sQ-6&kDuwZ?F7FGP}}(q`BgQnQG{N
z(`%v$ayi3C_G$evQ*us7#I6N_gQwZwlo@%}6P<a~e`rQA6408li}JAn`o8!-?7e4H
zQ&0QvO^75Qgd);g0Hr9scS2EAL=aJ`6zN4egdRdiP(VP6lz<>8DkvZwLJ?G2=%{q*
z9Vr15vQK{Z|NfWfyg1KtJ?pIXyfBLwo4u}SpZV^YYxc~ZLEdI=r(@!t5047=VPNx{
zd%Ic9Z(C8#7ZEI2EN;j?O*@wM`lznxda+q1KK0>vFVlHcGYPs^-WNQ$haNgC4BH`I
zlJY_nqq*+?NKdpDu{B~~`s%Y)(GeZ^@vWm%9cXPZa9cRnP)l7qlE9lBJsst}xTBMP
zL-1^KIV0V80eM?w@_fy*{Ck-3Q0F#VW8?-Jc~597R+Mh^RrEJcgtynu<CPEhr~5)>
z-voYCq5VzI-t)UQZl=>bZMEO&lAY<!uuZ&>@Yx&gQl6!W#g)4CL8~bQ6m5o;<hAsv
zzAw6T^8w%i9|ntOR^OBLE-z))n$e;*ElN!^LWG$VPj1JfeVCrdG(2f-_>#OuDX70-
zEb&WHC%p-JZ+yzR=k~<sH|0xfKlk~pXcrhaZR_ha9j1P;Ne5SFECm?+Q0T5zvM{@^
z)Ol>}9s3%2dxP~#>k3SWdM7#ltu(8XFSBtjxLx0su(yJ{z^GTkI>0=#J35`0pYo#z
zgv(^-O|IN^krhT`-*p^k)TKjuLYukIgWl)z@~WziM_k1m>2J%Dr^#~9KkExnqxL?k
zb@=*ZX*J=pPC1Grw}d)}?y>PaLmJ*7q@g=yrA?7b)RSU3lKR#1nWIs-#}zQW9uHS=
zzeh{F%DKKY;QHg4t*2gm@3lA8w5*;a=UOSGIP9*TUVS;|!#6ir^d<IoqUA!Zw3>VQ
z_VlezjTs>8&c(2er~zE&jvvS*pxeW-{qX67T&-}e`i<o+)`I>TlJw70E-9>5DBl<T
zlk~6IT<1p`oBiE}puc)>4y-q?tNRwvFfkMx89VP2diHH@KBzaYez6AoRItJ5p;@tz
zno%IiUo~!7k$AOZnc8c7g%;^5MCopwp>t2fAK^_jl^MZ*px+O584f~U%hEIU{BWi}
zG1-ynjLw7KK~$rORp{4i3VQ9;ukbu+Zyw){9e0pXwCbCzaLBup#`Q(rs_80anbW0a
zskJm7@Q)W1Q}l*L;MCte@*=)57PC3bLATQwEm$?EOC$6!T8_N8#1DU-0B+_7Q;J!`
zt)8mCJvB!f1dQq`gNDOd<LfdxSs8gG=WZ<?u~W>Ls!Tr;s>zyYjr*n&BtW<Rwsa_(
zQ8lj1c;)=olZ*y7I`ljH#pkg@Ifw7}rdY_|s<U13c&OJ=W~ZJ)%jQD;i?yR%ThC@W
zAhvqK+|y#S;_X#)@@{o^*J=e=LJYWsdR2O-^*rlAg8^F-hq`8ooY7z8%{<R@=e8Fe
zQ9f6l3I8}+ana0Bvn4`C=wy15d%GTb)z{6fn2EiZCsrxpIc>1k*X~8OOYcn?X_$(t
znG+*l4KrR(9cIeLpGjAiY$ywedLVj;wy5@CSJ|NRn{TFXfGd5Vmd9dm9Oi070IoIY
zJaj5<Kl{2`s(bzxeaA;w!ncy2*2Q-wf^sdBR-qrro5e@I)zf9GpJ&DX6p5p_oR}qW
z<wZmFHP7)lEt}`_R#&TNElyg#aa<hk<4StkHLN)U$l2DOdGiLDKf!Ag`c1KjoBLb?
zqrkzmRE7B#WbYj!-6=e~^y3#`<&)7Bx5|f~zXz)9-2TkVlhAf%XL!c<R%@jtvhN^2
zWbV?@`$zKJ)JgW5#cq6sz!H1Ln+&IX#oF2|O?{goPNV!GVKKO$g@~64EG%ukGu=x3
z;`(aFQ-Vv()6L-M_Tx|cU#)2D^(EP8ddtl9Mw`B;lHwkp8wwG!7ofY;BO*3Cgqi~C
zg14>|`iSHPL$fUpn?G>-*e=;&6wSJtr{^LVfp$_!tMR*1oaFMCP?x7Jv3q?2FP)D&
zIhoXv)Y;?_4Tdn5mGCG=^Y_Ra(DyMd;M^@9m(VF^HirG*n*Kho3lbh~i7gce6kFoW
zJNCbO`2&!m%LkF?`72zUzOP8g+~1sT(2N(#%tLC;8z<DXUhV8^-T2tbbl351R0Z_n
zX}-f?uLmVpwd*lD={w1Xlx5&2w4#t>_8iG7E;PW)YQEJPD=Sj^Hb+l6?M<lt1iEkh
zO%J(0Se{mWQT%p|J8)+Accdp;AZvgcwDq8tr$O{gi(2$VYyZ?<yra#o^hXvRVxITv
zDt;v#Um-1hFnv`pZPwf@8tu(e;&Yz!e%6%V6S99NX4!G^k`Q+h<UJyqN4ME&zjL@j
zEB;%KlV1j1ORrdY^o&LuN!EOjdo+`cT6IN5(B;a_Hxv1mAY(&^q_0fYdTY5_k9%J^
z%|@8>pZ4MwuPm*z?@@cQ)zKdir79yn%}G{h(NF(EpM9pk^bj8fde5j~4NRMw7yHBs
z)Pl}2moBKWQxC;mee_{q<B~nJp+$u==m#<;tfapH`8Mm~siTXrbJf!1Z_x78vO}Bl
zN)35M#_QhBzyeupAWYXe=dsI*hK45aOLi^MrO;g0J=Iy~rwI6K17k<@>(R`|2NQR`
zfa>DjjOK-{!kA=kw-c4+UWx3rXd3FB&%7-w7YlRRu(z-NXyoM$Hi<O1i+*M*@<M~J
z2SB{_|7j+dV7q%$@8!1w?@a2XFEKyuc7Eo5s`B8qDzQ|KvM0zW>@VD<N6{H5IEghr
z`qkV^A^e=JgeewMpo1IVlRI`*)5Uw9`LI(YcpOB`!d1hrg~+e8GFZMo`@!bbV0+yv
zSIuqzy!MBQfwy{fZ2I<WG4VmGL&Sv$n85>Tm>&h}#CfADK}~TzlkutxSiHSyJ6rN4
zPb<poam^ecWuh%GTE&q1@VzX;sg*CfW7v35sN3<!ufu&FmYtr5RL?9Qe#_H=#>B77
zJF$rn=?umf$G_cj1-CshH^T+t%9%@-$JEQ@&IWf~V>e7wTP%<Y%O&r^z|AasP};{x
z-De3LY15Fm^^;B$u=r+p;j(@Z-nH+k`i@`r(v5mj<f!w@gwJ$T5wotA#v#IS?9I8*
zZatO)!OFn2tnlv|rJAM@-h7pQgBZ;Mz`c4#3qPUX>HDd+>(hWc;@S1xFb(UVUB7zG
zQ>(FETQOa()y2cUr=VBKRpfd5O6Xd?DVQzg2l0#b3+C116(nt>?g&S%0W@Kr_1MhG
zMoH~r7%v60lFepLqW3wYed)N^pW88}B&^57+34)_M^2h@1*gqTDlNbI&<yD|FGe#0
z;e-?(g29}sdx45~|2XOcAN>ONW1I1lc|W(wCqBsg7p}&?2<HZhS(<_?U=tGru?OtJ
zW2;rS;s;3Tqh1Y1;|HTsVrLJ_dw|BKg@|X}21&fpU4*mdnZ`AVrE&`jm-$mN!VwVx
z@orW2{`oqizvm8(BtDsooCX0CPf{=~6Ct&lisz)dP0g-Oy=e8~0w;5|)Kw<QhR`RW
zA4N)0qpDcnF}B+)l^efU*Lz>=%=cVs?ueO=@})`E-3`v;UokCBXOi-fZv`afs%OSi
z$V1iEZJpNJww6?(hSlk38tp%OR^HD9f4rO-`Xlu1O7qOv0PQU%eza^RTJ7CT&%Ldv
z&ETLlF@Bx1%Q+7|NMP2fJ)FiTmfDYn>0^f5t>6=d1A|*XH=F~+7zdi`>T{9elB@@_
zvOh{s8uzwtRc$Spi9EN5ML!u?@H#YL(%nt-`^r&lS0_jmlnS+nwSUI$UaSq!HPz7u
zF_&6IYOVEKC^U&prmPTxPH9CfF2=vFom@di4q@jXU3&m=F?k?!LQ=bc@4jS@ZC<AK
zbjMd`*x)BpGj}U+_%^{%-OJjHYB~H1J;Xbb&CUS0GsX5;IpeK_HS~UAN@rnCx6W4I
zWsRIo*`4UinN+9p-5d8yAU^r=F<ss--GtBjRw~yvU)BBmMl?c&?&m6VV_}xE-ovTQ
z^dOn1r6=N=evoIrc6tKAfp@FX-gc_S3=EVNE97~<uIM~Ac%N3Gl@QzShV-+5=3~Jw
z=%y#lk1eV1j*gF8YSJAdyq@|jeSBNVPr={w!N~fWs79+-?Z)RciFaMi)M=w1)vpCN
zaVXZ@sS!7zY&knm?lKRQf?wEw9AL_sE%RZH37Zn6o{<+gk-??W#(UI06wErkn>h59
zzG?H+kvLE?a4&A)^8+@6)br6L{n=D_H?2N<MM$d!-`U0c8nkjP&JT{xcCno99s1G8
zm>T=B=sSbA6lSQGS75?^N<MsCA$uk!mKny+Es<Rm{i4^CvXqbIgG!%FnYx{RhY4oz
zd)~SnV)V4@dj|6YO}jp$^TLPr8Ivo?V$cpy`$^A+W_4-knzb>iTz9%uz2=N@X@GS|
z=R5n#r)AZB&{82?`sOo<!mf&P$9GtIZ?tgFW~6KhT`OI_`-)3%u8G@{i4Zk!?@Qg;
z{}TO3?f289BHoF%2wqa6T=@4`_JS>!;;I4FO5!gMZ!L@6nu|}(MlD!X{omczjA4F{
z$Met^><5z<%}ZbIt`1P>;IXWD)i=HxM>nr~m}4*3)}8l#A2)r;AtqB>kUx9|aMbAW
zvwP@L)BvpO$=X}5@GfU51t{+~*Jg4p)F;eF4c=V~j*r(?b)pq8Pjyuv*Da$}|GK~(
z)h?H|^+>4b$<vvMuvVSZ^kLbiE_|-@{mBR->A9)18^q<#b+fzl>1hJeTJl$UQND7J
ziF2XE`Kb-9#jG_Cw?bsS$?m}XvuS??HJ%;GTkRP5V=F?*;eLg2seccFKHDyguG_V|
zuU{k4L?YMJOG5()O3oatY&LIR{v5heS4Z(s40_{3sIznUm%s?o_4^ELL(6{rn#ki-
zyxH?a^zv#puU^*Gs!|Vr#V2uh^Txw;Jm*=RcB*4lBoGgMF1~ADjn8;b_HIUuA#4Gf
zP>E+LDwV5t+8fQ|jeUI#R|NIAqn(9=%_?-4fPnZrrj3g^J&$xnle1-{Cx&V{67SU}
z*g2cNnKOD){lVvP%Sy%I=cj|T;C@!W3_eX%$Hz>>lCyy5>#f%-ul%gWO@1vLgwBB0
zE1YDV@rhoPL!?0-YeuAs^+vM*Mdoy91N&<S!D<$E{=qo^x$swFp54KwSM>v_Q#xYI
zGS108iw;vGJl4K^2DAFELn*;kJw;r#hPv8}MxNV4T{6gHeg6qE^Vt%f$FutWCj2bN
z+~qqzMw-P!#|kJaC2<)G_b7@8O_$ZwV3U2+CGW6aaNfd&nU&Sic*k4bC)S1liM9FL
z$rHmsS|8(&@E6R;YoM#LglY7wy4PunD$zE={i*RF7^yQmxb-0;pyTCF4)-FrE7{Cp
zghZdgS5HRIpNZ1HCpJhsH+YHsppt^z=cmGNrv|m(ZCGQ5<VhO*AkKihI_A;p`Lrct
zRo<tqUzgQa9;H(A4p;~M)@JQZ!L!YL*AJ>F8$s+^rv4hZp3^#Nu6>`?<n_yif>dZu
z$_ns%NZ*I*iv;C-OQg~G4<r}R+&QA9XLXQk+A)}fo?K8q`AI)0bBDJ&Ye~y$HQn4Q
zZ#G!Hr+in>4d2?g)^j*k++nbd;9N-dAC3AL6A0EtwGyN0Lk2E?_|~UQ-_>YUSNMk5
z+&T6p7XNKSq4$S-j>1h+{nwY@jc09qPd)QJ!tQDfRx+wFoh@$Y4SV^4KdN4}@m#x7
zj*aw`^P1nx<CAWk!g!!<5l*YIeos3wVp;5C!#fU4Z?+8}<dm~G=H|U8Y5HyS@mHiB
z#fc%jIk)N>-8tDWKM15&&~@AF&%XjDPgt4`>!sKGOTG~5s6P+l0o2lEo0G4mt-mrW
z!~Xw&$r1@br*k=SkfJZ~6ciUi|9gq(;rjpA<)QzCCod7rpFmnoWR^L<4OC>}z@DK@
zen8Qccg0xX;^oPwnJ*k`pGHqjdm^{9g7bj~GfvN%u1HhpzO#ze7NC~8{%pxJ=R%}5
zi;QR!pHqR;81d*EVajj&3X2g{lrEgNl}a-x_}lN+CVazPsJ9A-#5Lj0UT%nvHcr6y
znhTLtqsY-b1To?<v{>Y0-dN~TRgN4nb@usE9BBUg$;<AA1siq=sk(;)&U1P5&z5N!
zoYliTkE63tC=nvfaeG_T{kgW%k+qzsw>oMKgjS*>Ixnm&+)p11eewLw_C~ho>4d}F
z&|c+9+Z=s8wFbmd)`dU+(cFn}2kXe;+*-$2l%ziGjF3z~svZ4-^EzhSV1Vc52x6T6
zgPc?ksU<lI0Nq;|(o$g&w-umjs7tleSX5*F&=)JJn*D(F<{AnD6SD<!FC1U!flR&I
z!G9ZcWRP2`hc-9Cqw(ZDvW|)F5#UEJ*KD8ei?)Ogf==oS-2g*Hyz|e7!X;-E=vX+U
zJx-MK^B~F=bb`WGqtM5~t)L}y&g_$^wn1d%MSvlk{=9>W?HAz1eBtq|K8ep2==xCr
zVNC&`huc|$Cn%C4Vkhp?J#f+t5KW}>2K^Q$5uO5s15}_EfsZnF7!gspul|};-4{u<
zUKsRK>v!RhGsQj@WepRv!gm3=i<%QWQiZ3^$0X-3sH>~8MRyM>vxu5cNdRrLVkOb=
zFyH3gEFg9=m<d2hD^CG6;)*Znd1&Q{_rouM;1=q*h%mUq<NXDQUIzMVfwgVWKpX(|
z1>Aj(4JA)(ezh&4z5q{wi-0~Uizs=5iaA{ghAREo2@^Z9G3z@fv|R*}I#DHzB(3s)
zO1Y|OwF==`JRtBfv;+Qa_{=)sCQNOCbZ1@wdSsK$Yl4YJw4WB9^1dV|g<1@Oia_zS
zL{iEgpQx`Q<E3)|Tx6C08BtOj_>>$%`qn1!Sqa8SiX!Y6uHaE*d~v|JJsUdU(@Qkw
zc{`%WhytLh2Bv4}22jB}Bv;}}HGwS)IM)o#NI)?To59YoyCF{Sp(^#WJO6~exm~v3
zcDxN3&f8F%!_Fk#htZLj)Ya=jTPZtDsx{g0`lOX^l)th_EMWX|b{p=(PR9wH<x|-;
z=czg0H&CrPHcQMl0PsF|`@LhEc9?FqgP96-1BVGo3zN-7U4_!I3yCDUZ|@_4{ol_e
zWlsqV`lL^q32(ePy;Z!YEYCcG9dWyeny>>@`Xn}rSCcm|r<OeB%Gi3MJMk0i2WSae
z*N>lI3OheFyJ_<am^;;r{d{8t5kxh$p^VMuu;qtc2#fK@c^HwN--I^g5q^;fDD$b6
zpPI9?r<NI{e8{~H!dW3*F(l_Th`u_B<4CgRd=4bH95kYZvTv{1uKqo^|IOxCVG?jR
zd7{nDwsZ;A@YCkSsj!>|&bUvG?aRK`kO*lD@+<5pi#RBu^L`8?rhdK@l3lW;xw!_f
zpZYaePuzfH00McSM`FOuw&`{c)Jj$bV1-1*%@QoIDBxleic$Pd6{_9Gi^+-~CWa(l
zCYE-$xna(1Cw&4IAtq25Ra{O1p}HAwNSFq~H3&)FsEP<uEs6p!T86&s2@GQ~zbfqj
z&Ym_m3ON>i@Re-{ry8tpc!#Kw{H-mBScwMY2Y77(Ry$YhHsXZM9~=SW+|6t}tx&d=
z$C)i2Pwdi(T*EV+ge~LYr-pmegBO5(>{1`Gk_EVVmZzln9@MV?k|+O}sLSF&)T~=1
zu>x{qnsgWZX<>FmOP!Nqlh5i>HmPODt_?4F4}h}>&S~Cw<0Mc3(RSSAqdPaMh4n<J
z?25+|mcF)iFU@*liV_M@d$TsCpULQa)$!`9vnN&uB&unIE_lbz*GtC_I}%xVol(43
z9EXs-{ebAsS&xp;Cb~~*mWyW-0lc*SAn9f+-bSpc@U+tYfji5lO?9^`oWr^wwn`tx
znFpmD#onkx6(UH_<)_*a>0?^<(lg+sPpC9^`U4vo*i;eZSNIX%_@)hxG+_y&mH^!6
z!X6*9Y2O2b9<UIgG#zyv2GwQc_KF*y`&)qaXb^V(18N$<imyJ}T=82$-adxDKkP=`
zTm}(^1z8)TQ{%YUO-S;)J8XFz@Ks1|Jt!A^8Y2rn^UKE6p`XIYUEbR#4yo>BcO<C-
z>>b|hi{rX>Wnz!p`-JT?DwTuIf^^z-BX$IlRE8FE!d%~pI#-7(e)qYk^2w<xY?NA`
zw3DtGPw?%QhmLnPL14P|=!{-~(?{~mS25rv*wePO49&7&!|5XCBPENO{AIJ)M{&iV
zXbk&-gV2^)0lYdQX&(Y!s3*Wo?{McHY;Atr_mh!(!${;FgFqvQ{{pEUUPN@4@t;eF
z_(OL!F+~v0w-C-7$td)xC2TZ^GrL`V$6OUJi2!ChPPZf<jR>OO?%MEs*aGi;o5745
zISG&b85Lk;74O+B?+wDr=`HRxo-%)J!@*&bq(uE;#KWOQh!oh9*>7*44Ljuy=#UK7
z632UZ@S8RX<~&NWST}ed>;g?KR1J3AD^`QLdIANL0XopjWt<J6&!{4(%p1QyuH&pv
zfJm@bk?TUD?dM)X>wY-lxu!nDc%^<y&+B3PwESJ`SDfr_Z>7lcgj>A3cBQwUmzwKo
zQEeQ?pOKMRA*+z;vL}bD>A<;guA1VwMWjixa^5S&=u=6{Uz)K5QqpB^{8yA^>4Wmx
zVL8TqjA}xLsx15!(fzX~HDe=s%!&0lJLc!CzZ6ca+z;zp23l%sH90XI%`_u5pjG8L
zE6}vPbSKAyixEuIhsQ^Xv--0dmf5_tJJ(XKAk9gN3q(U~TLJj}BRW`Qf-B;rj*I%8
zJQgKmn@<y7G2eQ#im1qZndLQTZ?^JP9cOtuqzm9F?jFI$Qhq9dIQcFQE=?&iG%ugY
zKWwi;C2f3j2A+4djiE;nu|AUJ_Zfu%&Mljtt`wp0B7Q)eeC@Juf3Dt((H|U8`VzD?
z`=k96Wf`a}dcSx32cWIFKQd*g@T@buwzuV|GT{ys6Y!#zwgl}GThAuIWdgnh<u?SQ
z1A?|EzAD>3c8y#{^se{8<laUBnQIYlkuJM(M#5Ib97YR!3=*tg=y&D3XQr_)jt0sC
zBl=%oA8hGGO+6k#ycx~`zbbc%-lWI{!)vV5=-w*AL;w!XU32o2(kli`VG)A;pbB(|
z7Wtf25@D0x%^&PbUvC@Op1;2h;Vs?yJvyLiN;kNjzjM@<^PbD)qQWEKMl<x>f~Il%
zVB3+T2%zO8RelNvEcPVpInBgLy;Z;`l;FH8W=8vD0w+PP<U{UyTY2vpIq!|B56uy6
z_F6mP-%pOO7w#qN$!ZcdQAw@{B<7U%E6U`B>g3#>jps=gB<efr8N}iy7#y=ttdxVv
zQ4i3WdH$2d`!K6BY%+>iQkMET(%Lse`MVBNC+VlOd;Rn}Z$CgEKt>Ry#}f%a>YA+)
zvzfcRZNMB@{3YmU#}%U`2pdgA89aPN&drbttr=`eqelrg$s_im#ZVPBXKX&6zw>MM
z-85I+=QdCHvZt-cB_7>N{SpYU1P};CA1;g8D$dEm=tKbRbpN?ulCaTAsqIXZAz1$6
zGeBTMqCQ+h2W%r|=|8BUjz<n;+0O3tHSRyW`u$@EaL*l*io(5i#tI>!w_V{6IIlfL
zjYXD4!>Juyuw~}a>Qz?*Bg;T5+_s!XqyTW1m_Iqzs3!D}6<R7o@K@l&m7joR?S59N
zUeu45QzzRisOQJdao=lD(~)CZ9_c>|xh6$O@{MR%Hww3o^~#5qfQ}@UNTx|BKi<vC
ziB-tGdaiO&mMPd&D?EtnxH$3@HJ<~f{ieA^eD>89NxA^M2315K-^&45-y1-BT5`Y~
zT+|`tyiDZKX8^U`ZJwAfGvFx%D!K_xQZWSA27;$R`A}}aoYOm?DpGT&G{3OiGChm%
zg;b|ZY5=)Iy5yc$0aVUVGgyq=qK&;%<TRpVuaY0YF16hxk{dSHHxM@iF!KDeTK0QT
zZ`%mPRut80Au>YLNn?w!hHU)-*gp%Ub)!(Vtq<@8zgl&w0BwjdLW;ij!BRuD@}Xgi
zvm=Ooi=){-KR7>$kVy7z>uWmpQ1LjJ)#!?rA$a91nVho$nD!I7ke1H0!-7RUU8wEt
zZ$O|cb{+1?PwuZKTgD=1!1FiQzM_nDNQu6I<W2-q67AH49&@9=P=jhh|Gv8ok;s8e
z0=~U%W$68=I&?MN2>^aiF*S19;stDluJO;lY+h;gBgF$u(-cgozAc-6&}P&mYkWol
zh^fO3yG`K-XL;i<FLjM{y42k!H|LLAaK$8Yk$a22UtaRwiCpzMO8>NkMU%&Pg}K9>
zhqeRvopTAlC3^a%evnum{&1vUrJrjV>-Z7r+xBG|5hU2euy!9y_6I?RV2vbJawNDK
z2m$ptkE#?6p`DKWi7L$nA>B;PNa`@F&%;=@SQ#Kk)4|P__c*W#ux?D*_FX(K$eIS|
zds<)w6x~c+FlQ=#am);eb^bB^^`16447rAmKghuc0dlnoM&dzU;D9N_?8TycoJKqu
zoDvz-1&{%jL_j+aBr7MyH*LQ>t3qyRVP1tp1BIKY_@-#3li4Txn#7*E2Cf8iBQTRY
zeT=+!CDH3V-ibj3Ai-q7|2TTZRC1Mf=rCwd6S{j0(u2x5wu!}4Y|VONHE5k&=eYg!
zPX=+=mEPs4I@%S9p5;7^+5B--ylsHgR5vM~1H+Z@(2pQ2FcmpMb_No0LPtAtwgb_d
zv+D@5bXnBEc@R$d?`?8MUYx6f+`v?%A8)gn+pYrU)Y)xrpYljEg%i_I=wXGjQJvjF
zWgX9oZp;7ZSH@&Y!>&PZ_p>U~`hvSraS+Pn3ed4(O@<{Is9Z3o3*)*0cvDj!EN{-K
zWb>xJ^mGF<z^X!8xdRT%vR`_mhW7-Tx-?rDw*3_u3yVQMmBb&g7)3)*#10w}o1`@h
zxzaI)RncKSspE9P)pM;@TxJF;tU@C-nZFKWgFWXyl2x`}2@1&inN)mWf>d@|Y`a64
zrNTE&lKtSaO$v9mYv9iJJzbiLslG&7J43LRfFC%TbhdsL9>keijrE5qk%Z+b0Qt)q
zz;&3F2n%rWYg+|ca@id0n`NK9?c29dJvfNUln48Rn<i1u`zXD@(j$mVS4nl8_!&sv
z3hG4@2(c~y4wB&H$Li1NGD^;5B{aaCvplC5y<fV3icrLCdnr83uFco{hRR%{*w*|0
z8Ib60lf%8{$LBs=Aj9GB_(6e`V$UuH9ZJl(Vx;}RUr{e1Zd)8x9mm2FyL9yild!zY
zPI#NX6v*iPC*P;Y+43Nn%5L3_DnlRN+VkRM7qC)AaP9^&vVI^d@(1iVTelZd4fO{b
zL`_HfR=($~zpw*7QpWv=BrAr3w1yOvRoG2Et=I_|J?+o0Zrq<9MAa5eZ&Q@ZM)uP;
z(?;E6d_EPq4i|vM)NulhXm}wNBs27}CuW%JHf1z%xmZ*Z85kd9>>f7Sq!odjNq4~J
z%h*p#-%bnn`_RhSAlZgqX<4AD{ikPnv_kO>Ckk2Jnr*pqa*H97AoYq`h1wmb<wDCm
zMF~8gFGN5=?QLm$Oquj<^m5Frm{;7}6S6@7j|+Uf9Z=cESVz;Zmy>-}?YCo~nZn-^
z?@S8wXM`lgkc!^#96kWbE{N$Rmv>>dem1h-8}vA<T!eY!32sG?5>4Ax-|J%_fl@<T
z62_}qOlc5=aWrd!R$okeq>v6ngd0YCGL;uW3w|B#>xq%p0S7E-T7sVdnc1K~@SK{G
zsmcnhyoo#OSKB|O2`vTTTJ@%Ox>YIc6ph{z8R}%FKqqX)s%M%&k-@25+}FO7JEUx=
z8|xXek2;}>=3+NWMOA>tKy|oE1oUu&eB_r4#wn95mwVuqEyzu%vb<E%_F9XB2ZXHL
z%{F_<fymGXkknLb&o9ZYL^rL$sQYn~i>$^>La#W<N2KvHPQ7L?p{L|r!|u#bfyFTm
z+D=$V3aSDf7qERFPQdJ(D;K{rr3FSd!>r8gh=%_R&t&tUPQegzDUJs@?jn_WH>Z$C
zs0r(rBmLgcB2~uko4l{zD$jkR8OjejqM619R-k*50T0-5qCs-XAQcUk{Cpd*OliG%
zX_ZeSQvglSpThD)aT>i@U@b?}HYgZ*$V)YaMh23Mdb^>m`_m+RbokA(tBSURl!nl8
zZuTPu#vd0+$glJic2`drf~h_4&>2HPbF-IV9`|4ux=@sW$7I_W!u})VjWTB7jC(Sw
z3AAKWErW*)40wXwR(`f^bWWal0?@gTHY(zXhBuC!p@&h<d?LWA!ji2#LKbl{Yp3#g
z6cIV;)=%LH*E+^&(aPeI>2s<5ao0?;)p1qUoMwa0XqA_hp5j%U<y7*HQDvY9-^p#F
zNaGIqfuhgT+{QN%b51nR(033(T8i)*=4>imuJ?KvNPTqCy<$#MPKLIM5lLRP1hfX%
zo#Wvb!J^KR($6Bn(HazZ-C_<oM>oVWLZ{()1@fC4pd1)SCghHNJKYFJEDmGoPWSIW
zi83cY5zJTTXT2F$+QA(Wd9+$hPK6Ij9%HQ#U_l>$M9nrUFH*|DR$>mt*6AEwME&!C
z@%y&z5xM~Ufi~<Oxc_GH^=cnu@aZTbJh+hob};r9sfgHiry9equW@;|aE0KpH@Q-q
zxLd!GQxcTmW1=t9V`&7#ELN+&4ayc741%@lY!9W*hfd|t=nP$=t%*cFFOmw>W#@@W
z;91~1^Hvt^$BW~MeAZ8&=rn>TPAgBPztq(>9qI2)0~>X6!)$7Becs7OOwW2`w={9f
z7869K_~c>!=Kj3rN@D3Z69YG?`=Zl6p6<2|8f;k8$}zqm57V@3auYGpQt8^z>vE~O
z;4kp^`7~-T*}JUS55Roy<^8~Cf7z@Zo66zDCb2wL7Lsr5-6ot|F&^uXfvj6TtRj7D
zexUCs>k>UWA%ndnA|EpjF~P)t2i$j!pfw-)p#Cv})+g!SG^LZ42Sf8t8a*Nd0cE1(
ziSfBdaea(}dOHIWgAM$I;o~{xO((ZL&&27!w-MClJR_I2Q1^r5Amg`2A}0VjSeUia
z5*;SLa@*Gy-~VX<NkC~U4e<E$usp%G;jVLcH!%87!noyNT@3YH2|LU|ymX)BhIKM5
zuNB3fx0r*xE&)xq-x`T_vu%<iJJNDkKU<6|b0Mv;Jc59g!}2(|WBSGr%N(gipb0v%
zqj4UU<bJE(H)>4?r-k?NEZWF^a9%vw1=pUi9qEL2LvwhlihNdcspiD(W69mkB@y7}
zQx)pFG~+j^8C9!jOTcXfG(X;ORB>!8TEc=<%F7^^tY8Nddt&u3xXR51UC3?rU30Gb
zOnvb90C^`JWv7ng%sacNuWr~HsYeY`-vg|qi`PbZuF-CFF`Wdn&up<GHb7cO+8%sU
zJ9l_WsFzT6Wgy1aPEmI|BwTHUIT272&pXIDam$=#@q6IsZy=Y(=%ipC-o(ql9%2hX
zqX_y?vM}EClj`c1yMgq3VBviFtCE)fOonDKZhIB#1DvPdnVA5nHW337rxQIYyR*t4
z%W9ma-r*VV5W`8)?ndj+NKWMb?i2d3t|Na+0^Cj!rwudXF#^LflY85`<>vaLJjl&j
z&I<2bHRkv89|uU~)k1>RPT2E;kvY-b3sagoym#|%qy*At8Qy>zmP4khu~#itaWEHp
z*}cxSqGSGfIFa@}_>swZ9^FGV+o$u>Q;5&GP>|eQWfXA#D@yx47!%FvkNaRC9Z3&4
z5OdK5!?<ESU>Sx(ShqpC&r^tT_1&Fb6R=>>!~J#gnAR{_wl8z!(a>c>8kZu=X!^!m
zU&X3c!=l!(1(AfIBYR2}7r`eYCbWBkufY&M{y5h2!&?5D<kw749XlbGt5E(m_-8uE
zYpuH~X!F!v!y&}}Jy|HHR`i&bH7pengDqA;U3M6T?mroKBQ>e{ZA}#b>VwqeR41vr
zT&eJp-@MJEIea)mi<5Bbk+*PYR+W<IK$uh1f=h^5v`{@5_lW*M{98z$8vCPv9>HRI
z?QEL%>EXyekW@J*qX}5G9_{3Wq2ZCc4t-o;Ze+9OiWrC)C(P&Ps&VZ$GY|8;(>-lG
zI5Sm&R6j)KEYCFk1lFt9_GWf0&D5==W$oS&L={JM*vN-MXRIN7v_q88;k##=nR{{+
z7TAe=Ec6xhap&XN*Wovepad4tb$E@b6WX7#{29cB&*VBZn<~Q1RyF-CXd3N_kq8Bc
zr*nRO$UvOm*A5tTQL}ivtZ{<m%!ZOjA%1bzMNoqSOzh<|N6`y(BL;a+jR|EXprZZr
z1C-kF%NM2!ZNC>tbp=4>XrEwDnYPIfw@Ea;z$!dhRy#;@>JB(mC;-Fdt%Au<wFjRn
z%i9tLagj)dJaOY{Td`(Yb*9iMAI(`O>~Sq0_LR0`f0rSofEm}G>3kPz|HwJplhVLS
zXds(j;ts9hyvK2gEi#h)!c#%mfj2uxN_UlY6Vd^l2jBxSqgaiKX!Y?13Q{=dG?v~@
zbxN}vCbtezSeZe5tj5On!Eiij{J^sV-7rN>VCZ%5H*D7-+X(Z7bqs;;?6A%b?UZ)q
z7^PE(SxOD=trl%oA(~Luo^wDQ@dRfmCp~-}6Hnit?m|&h_)8A%$Nz-hBBk?G+B+k$
z2&^KAi8fipL~-r<!=1qa3N@VU9?~#6xfkTK={f7jjocR5v&q3z^po0OlD*#-8Y{QS
zi*+R1HDx0Xuh9DAE*~_*G4YoNI|08r^LjR1r;DQeZt}@>e|7P^c57T-paqVg6@U}s
z>C)O@_BOmmEA7vyo{ZF~0By!x2&7bm^PItBH92=pW$%$+$lQ^`X+8wIm!dbxt^rDu
z6wds%>2YWo`u_eqbkl^57Fg1fNUe%SFibcI9k`WKZQI>MvxXF;^S%+1cmXw-kfl^v
z9ktc)H{+?HeYje*bphKTEg{}Dq_tV^29GsQEv6TUKYv}GmVQ?adm_c53kJE_DiSy7
z{lSgVH(?$=GYk#rN1L$G5H1mWU~c?m>2<txA0csEor7rV=xHl)IvWIoF;>22m1Ca3
zQL1FuKl+BNXn%)*GOKHv@Fa_tkfj!~I<6LP4!4RJcWd{T()M3vMW}|j$nwzj7=Be#
z0s*7w0;)+GrfoRKGl(`cNcc<x*b{-YRjlfDg{R#C-`NBW($xn<TnSOC;qgOBhV8H}
z<LT-{_YHO-DsXFPrOc|`Lhd%A{)Vjr<EkCEdU8?x`vN{+T2rw`8>8e)1@XYjnHt<l
z_&i6Dus;-YT7wlQC;f^(E)Vj=4|;nW(I36bzbB~#+kDSi9w?vca->l8$&G@<b*an+
z-boipC}CJ>`53;oH$HvmxX?y+Zo9uOo6GC$>fG;*IQPo-(~6GE-j%@|B(1@exq_fC
z-b+{J<6p<#hShqX1i#!+ez|Ynhu#5CgIqK_Yl@<e?*s|MZ=ut}Byc>s4bdhsTt++4
zTVO)9oUK!P2_&}!qz>l1q&~~yI0LR5tTe_y5&DiV>sk9I4BSzWLOEflC1=s7{vX2~
zmbYk`y8^to0`i~<6LX}yAfAKPVik5joI15>ef;hIB|d7GqK)8r<zwLzs$7~a!=F~y
z^4c4A<)#Whaw_qYdjz;P%$7BVp#qXekHf;r{UstG^%XjQ3Xy)u(QX0@!ec*;t?w`R
zY=$0grlVBwhD(~y93g#LV0<f4nlp%DW|pNSmdoq15H70oP;nlBZO^@4mj9$H8F6%j
zt!4NM!(ja0<u3hU%GGX<fBh9;SQ|@X$xST%`N57;e7o#0Hs>I99pR)^QOj%s?wLC+
zm$U7Ungh8aoIv@sVPge@)SHlxGO#0axIS18>`l-9>1JD!bKz0z_fuLI$P{5+crF|w
z?#(dwa;)nyl3VCq+r{R@987yy{p&P|@>pJ&MZLDIhKLkfBaQ34%&IoSeNF5>KpA9>
zp_gL#?#NpXp>~_Qeyj(TmSYUpPSb{p=2FcIz66_s2{MYfr!DA+F+}MY#mS1Q!@)Gx
z+gtCqXOET|A#|E_NaxKN6lOgTxzqZaaT_=n1}B~&CE|0QH&S^=2eEC>GR-)>bH#lc
zt22TAJiCsFKEdthuG`V6i)^w=w3V=Lg-EMMvEGC_@s`tnuaBNrQgOKgeQWqjETN!^
zQ)$^X(z)H%wyg>D9op(v0m_5s)}v!Im~Xk+Mi*B)-B<bcfo1TNzkasTsqIMf6`Ff6
z%5NT+ncyqiIXx1b*ZFk+fPAPT1Z!~MeWwn46GIgjLNOt?y0HG^kxz|O1dHkm(`ca(
zYtv*=>uA&g?lEhlO?OV@F;h??7bX~YNLh#5YnF@5$HnDiR9)1?VK}ildgD=$JcfLv
zX4P2awrynvS~sZ<{qPp}v=JTRiD6KU+|D(u;dA&-f71=SS9k|}*KZu#QmN?}p&v$x
z>on!-JSCrHQnZ^jjP^w5)bF)M^1gGQskTVv#n!KD<$7BLMXT0p8_`zI@ql2qyvdm8
zG>LVJeGId>KPHxb4RfkoU#CKTYi)QPTsHhF+jFjd3?%RUIgiF4ObX+aaz^hRlMbc_
z1fQW?;RU7W>=r2R;9Mcc9rA%4r}s!Gyx%d9Ghe5f$yyH`Lq|`8>uEg0;24@3Z+s)^
zI?+tpr+=jD?g1`|HDtDfK)D*sygfTd-#DbDICw^$p|K?T7S1jJM^YPzZ3ItaU1w-s
zqtiLKt_?d=wS%KP5i?9%kkuA!HElV%weiuxH3#Hx>xtW&r=0~H-Uugb>lf0#n{C!u
zWIuQkXd_%aPCog~)-l*Vh&604d&lddAks?wJil$QptGhGE%XkB3CT8U9<(Az&+yH3
za#OVIbO5Sz59Q!IxSLfL>@CFC{%~>s_26AfBJ*f_wHochXuIoJyZ7exS_b0~K3oI5
zZ-vrx4l{<;C_~3opm_#zq=xF{r%l^L^UvOEKb5Z6EGAZ*?mUER+y#e<0opYxdO-DT
z2XdAOJ9ghSEnwB~vzTjuhaKKo?lp`xJdbQ6P9oH&aAD;k+4Z-=DDYi|Sc1!3_*nbo
z7?Q)SNYA1^ij@otzp-n`)7)W1_8`i^eNlD-y<+Uo%;)lG5~SLMg<M(BFc?fDs$C;B
z1aCc%oZ3<=Xw@gFrqq)sfL&Bww$S0}r*+$VeH2aTbv3SffR#j^Lcbe^C)j{0sLJAP
zZqODMMcZINu*T!B>mcB_M8qZBr|IXMH9kM>^OzD+u<@&NIu+3AjuaMq^_{aW^JgvU
ze?7~EzPJU=hTeY}Jw!i^?u$}zXRZ0FO4%P(9-u(S7mt#Q^s|$3AkPUe`cW2^fxd$m
zTTj^wPSd>&Nhva;h^8|15%?RI%h=|?0$ENbORK159^#%RiI>R3c`Q$4cA0<Pmf1iN
z0c|C@&sXBGlHwZGA&jlRmG_qhv!AanERzAw1I(WDzx60B>f>17eeuBL{7YxD;zrPE
zuOhS}+H?M{vv9J7Gr8+?LXTiwJrVtsn=#adwDoAgDT)@pGdICB6=?KJp0)SKX$Q#Z
zvSGXmxdvzoz>VNBY|$53m!AuE+twI-Hx4?A<pcKlVo`If5r#w_6^dD+g#|?$Av3lq
zaVJ8bI6OaN+U1;0HK)J8Hn*~CNEA~?vkG}e2#A}hW1@W)E(Po|dlo^(59Z!rRDyb@
z(KEEGwk7*|1Gm^+>zuUmRN0H^AB>A7<jq066>i2ov4C+K@vK51xyVwz;`lj||L>Wl
z^tl+h-L3;z%^GAi6hx@DOF=b7dyw`$`4s`w??Dt5KuO{t-uZ0^>=~}k@#V|{(IcIF
zZ9_y(Bj!<1M8xUk3J{32qY}i0BWc!hMgs1OfKDASgCFCD<9{v|1Zv@*uiunf-i*r1
zL~1ZM_wCU5Yj|WK&*W0M*ou(oID;&l&;b%yv1<!A$wPlv`23F-8k)@$;gD27xLtGx
zbf1H>>IfT7o~%i&Mdztxj{^WSuw>Z~dA>w}f<lJT)x^x-%;2Vqz1KqtI|r`^juJr+
zz5lfyNli1z+s@wI(I5W6(b?5gU0|oFO#trdpe|r4Zy;sht>ft8dM(7)@lMDs6Z;T%
zdu0a!O^vf^K`P`39y<El!Gj)pc>1XXsSEroUKR57Ker_X;Quo5cUKqC`X@oSnSl{p
z$II6dE+-))AtP=tr63EJmz9u}msXUO7lX@4NhwK6DN4%7h)XM}NK30o%fSEr6VNzI
zzEbmba8fbWz54G2$^WSfxcK{ft4K;Bkw^)otb~`Zv!t}Lva+O<%s<hPEyVqTJ^k&1
z#6A54|09zBj7Qhe&)(P7+uzm86aG)Ub`QJ){M7{n{z>S+e*Z4nL+}4e$kXrN^vE0~
zgY3K|r6r^!|0`y4Ks7QI6&+tkJAW@<6E80hjsGb7Raaj}q?fO|A6(~-oSNkSZuH-&
z{&#psT|0kA4H+qEDRC)facMadX?YbX1r;epi~kY+KaBpj;09g}u1>-K8C*tON?Ke-
z!9+${MNUdZTK?a`|A*1PgR4pY1LA*x`4{zn%Jje8|9=DKVE<pdyaRka{#7jw_L7bs
zjt?C@{r$+yr2n7H9PCw`ynG+p`D?g7v~zZp^!9XCll<?S|Hu6ON3O}TAZt?czeM?;
zS^ux9{J(PX-`xMNME@U={!QeM;Qqz+7uO#V_#@-L>iUc8j|lvc@n3cQ#q~!7{>b>R
zy8hz&BLaV9{8wFnas3g2KQjKSuD`hch`=8i|5evtTz^F1kBtAS>o2Z9BJfAXf7SID
z*B=r1Bjdm7`itw22>g-pUv>S(^+yE$$oQ|i{^I&00)J%uS6zQ`{SkpbGXATszqtO0
zz#kd^Ro7o!e?;JqjQ^_ZFRni#@JGgf)%6$G9})N?<G<?qi|daF{E_irb^XQlM+E-J
z`2VP`v;X~>hodL?+YTi8BaWes<82BG%lPZMS|&l?w^~B0#J`#I9TDy!cP5lq3J9(H
z1!-4PGE#NBF2_x>$D~mPUb5}E+x;$5=dOr0Nc1z9tyHN?^l3(g)m%q=JHHP#3#9_c
z>b=joDjx2HA}J*`D&v@|oR6KZ%Hn-y2tzHc#G#6b-{b4GA$a8`e9OvLkFl}5>cN?o
zCS>L8@7CF!(4iK9pq{LEf;re<1mL8r?Sxk+h8sI#5ke;d=X#uu0AAi6QhagYJkKB@
zJyiV_T^U0J!_qI%wWWE>-{HbX#l*sfpI%3rV%CH0?#f}XCct~aEcw8FacAoAgc~64
zHX5&RwK~(L5YwW-c~-FVmZFJQIKa@PLeUSC)9-^|Qp0mPv|))qv&)j!3`~|l*X)cU
zbpfYV3kafrLJYVZ;dItU$JTQoWjczrta}0ZBu}HdjqSaXFrHU?U&~<t-0rOYeVWTC
zg`QnqGG8L_pcd*_o$0gSC-O`7?Mdhc+yjs{6%kTQOR1!^8lR*|(NZVx-|APqpSKa*
zYPx<BvL0N%Kc(NeyM>w({|pJ$uFjOdtiAm-Fl*si=_};zeKm08(Y8fPqR*UfL(5Zs
zvFi_Z9QIXXe>i#zz4$VsWXT;F;F0u@t7x^ZM}6w8L><&eu@_Yk+F(?qnyI3ocmLuo
zNty@|-ciSDIMF4gkI2~i>$9_5(DK=DE(|9J3TCE5_jAj8oD5bbzlf4%^%`YImaRLS
z`4JF{Y~APLQ{oHJ2@?gv)!i<?FNkn8>|8hl2_Wy9d$Zobf9A@K6z{ZiFousk%<$&S
zq@yFZ?I2lA%8InGS8-p<*bNFP{MMT8gey_MEz&r+Ha^sOpZ{R@o!<T2SlLmw8z8A{
zsJW=z=phrfvCAa{YW^NvWo_6!cp07-Z!0WLtL`9t!o9tqqsAX_{U|1hr@@`cW5kV7
z4Zli6=4)lx%Y~~j(tX@48-Yu5OzDTj@Q!}hBPpSy31Mi4Px%(e4Us3-^UTuj=hH1k
zvb>6|c%o9D+v=X(rOciZ$D7^WY(~fQG6o%Q3a=4Y;Yd$!obq7g#IaU*6vyIefLu7l
zU@`St4l3pVFbV(hnYbTO7Qg)h;73R>D>ZHgo7`DB_^~SBj1!B9y;mb~<LsjQ55BKf
z{FjezCwrocrXQr73W6vlqjoU%Hu1;qe5_vepY?{XWKBGiltg)8_Y#f=LV`t)xo$^^
zFh{A;`Yk69ZHi_rfwJXAtVZ)JVjt+h19^?p6+1faF(tv*JH2%PT$D#R`e-=*!KJ%p
zD_6i#vBWEe>Q_YQ3l`5_dj3GJ!Jfok&3)kwXzX5e49mWi__^(S`vt#|QLi5wd~Mw2
zGfovZOWlf${{7~6!A7{Oq+kLLz0lkgomY*#GTPXhTQ#fI5JoZH+tvE%tqqOG+xP&=
z91q`{m$jsiz0$vaIj>)K+Sx<F>YzF|nUL(|ae<PSDd_=MWnAv8(yf7fmSkcJP5hno
z&byfGDW4?O0PA5H_TKjK1$J8S(&sP{UfqQ7PJ5i$#pl&W@hQajlzL~`ll`jH&W7Qs
zDW3-2Y3CCWe;jgE@8wMQU`~UsbSv8JgrZUH{nuyCG@PM}rxJajCI@|?Ur!~t-0uiW
zC3W>L2beTfIDF7l4$=syjw_Z88(`BEBfd_$|5cDuudu1z-TPBJWxUr#;_1C|Bd0Rw
zK}K&bvFl$HSxy8E_r3}nX7|~?|3ORp$%ulGr$ZX7mMQq;v%_z^aQOK7MP)j3v$VV2
z<#;6Utprpqz7S?6Gh->Z__<IN-UAY_fTl~wA{HEdX_b~c4p=*K&+-T7u>5+!@NJb-
zPLBP|OV0Cjdkwtptnbp4BtIaGiW6?828M3aJ8XNMv+~}{YDI8(Bg00`LQ7a;b<O~4
zhN6(<xR@x5yN@!gTeen|yN|v#$bTr%ZH;<x=^ndOYQV+hp+yD%U~cYov9zNFZSjdD
zo~j%*#WzQ}lSuJz?*mVy#RkW|b{^J!PN=umu}sl_5IY)4AbL+o(z;3Pz8?;^z2Rz5
z#9~*i)c8j00y{^D(UjwT0cm=AF-}<%g)173w<VowE>2y2ci52bU8?=fwBF(jrH=@c
z{;=Tyj2ZE4y!Xn%b&9%`mM3E+nll$phjw8NE0wZ*bg%E0BH|vtwI`qUo=-|i)%k{W
z)$I>$^3)KZn9SYDj<o|_+6x?)x3l3H7u_>TT}s~+9Im7I^||29<p_SyC5sOxlwBJ%
z9uBtGH$UCG0K3~&Ttgy!`=lk2W;|wEaB+4ftuD0_!+S#eIRE)ihsm1RE%7f{&YEnR
z+PBw3g*{22lB=J46d$wfi8C_QiX%1(t02QIa(Uf9&8=VGFx!3io%Pc%{&t(+89EVx
zUz)Q(+#83M4?alYFLz}!@Kz?PPtru+-wR8-k^{fy6Du}#{kq1d+qLN0tJYRJ4Qes(
zmW6z`t#{0mB896aKI-0R7o$vG95%L9$zlY*xhXPYn|aDCB*WbWe)lmf*<&v4XJXIP
zdl98y!rh%Sor;;keD9ZRkAfk!n&_ulnY&q{i_lGUKY1?&f$|ac4)>d%Xtm!}U40~*
zcFp+gi!TN>!-i%jU+-b<)u?E-z9v?woKKGiSLj|-_YynXo|W(%cCE2AY^@$I#Jc#j
z{_YnZjf5R}FJYT+UR#T=u3X<xy5((hgQHI4G*F}{j;iwE@p*TLPrSb#u#rzx(&Z@5
zH!W^g^L<w=4IMoz)8TvWfwTO&pUHs5d&cGqb$c~5M+%s#75xeI6%E>DCVuw~*2Ruz
z(JG^&zj!iO<wW3ZGl{7QQ11mlM=N+&w%J9QPZ!2NF>1B42P~y<uq9YE<uBVvx7uO(
z9~amNE{r7xO_Zx`II>7p+?1b9mpPl0kP~wwn@&kVJRN5y$YMtC>sfZvGID*ch;{Uy
zIJG9H_rdOoDAQ<?=Yn`eJB_E-ySKJ3qA`iKuOdosoC+i+Zy&c_v1l{DR#+1CRp@h4
z_?u_YCrM{2`b(c(E4nm(>pe)?{GHufostw@!AbXA|MMdD@_l=?4SZd;<I7>38yBAN
z%L*txRA+au|I!Ji)>Nrr0zaBCeLPOs6s_*0$aqxGC!=#A6MiL4$?nAu2qx9)8I5>|
zZn#3MNd^3>M(d*wYE<=6eW0J?gwoC4BMbGqii>Mk@7jmva^EkxmEe4qsg%wrPp+>k
z;H6L2g|IJgpw*rmnpyhYrGZHUW^QkTp6L7%zc2doiJ30qb*Tdf1f>4!i~IQ;-6X6K
z^>@h$$c@KF$?tM22OI80T;aDcmh35UJX?*9ow_`7=W~O_t!(P(M+@0U`x``Q0k=eW
zYfX3q1(Sx5D#3{1P;49>t-PdGs~kI#0~g3Lwx!`L%+7nn=bIk%!tz65@BQ!gXXK^?
zcI-@e6lS!G9-JrIkAK!P#eDFAzlYGu2hjcMjrkIadhTxY+2(;`w^m;l%Y*i$sdy@z
zL(7bb;5K(jEw|Zk!^4jE{}%v}KyJTbyt|Lt^*h6n^tI=E4$DD7+@dXGb*NKBuDleI
zEV9?F$6I`eSShBC-7L7_^7CZXA1F?V!Y_@DJL0D#wnKlUFdAPK1e$_OUutCnlOW#C
zsz}QMXBosM!7LILDIn@l(Hasfz$b#DzBNI&C3!><c-9*Ec!ufTHE>l9ldL$DTV(@y
ziX*Zq`x&gEe@}_ja*AYPact4XNI19R7u=+RizeAo3dCr~0oI5S%;{A@05n&~@)@#x
zimIHqg#+TCc?EkG!^*xKTo-d7&5}4)16&fx0u1N|Ty?4s%dOqGR}AeLTvIUM?6>44
z9W^IOLO-~J!%$g{oolxo_j}}SRp?StxahlxIGg0Wcx)qC4jTc6YdaMHah>WUU+ZP+
zwv+~}&kUd*n-HNG@1v>~KKdJViQ<ciJ+LSnRhQ4R07e}VZ!6nUr>GHAN7BhId<iZE
zkPZsBgeuMikamh&L+trcaS_Js+AXMjWSfiUm&S5Xgj{o+h`0RyU~##a71x=osn}gi
zf(?i1ggJ&%{*sDV)EWsHG>Q%Cz(}0@Gc&SmQU?#0$07wqW&v>Px`@s591*<|Se5c@
zNWB`v!zMii&{pMRfh=OO_rke^Y^{k`@#vCV9y)nvU156V28zj!>%5O`)`MIW91-t^
zTa=yJ3thEfl2~|@uw3NrrL~(NW!(r!1yEXJNj%h0R7H-Wrbh=BW}p~dfnLp9i=G3w
zskPuqEKl~ccZs+)%Y$V_uicciy<qFQvCK(hmjV;ik+4Yr4XT2{(sG4yo~`nZ3dM8>
zlf7#mI2vhUc<XUt%YkT0wZH4dTYCCIYB{5j+lv(Knh_4RR*K{}wFH7zMPPI394HqY
zA?Aq4ynQ<~GL$NijrLrufqW&aot<llbSs0mtxOSw#LJ;X8_LI$$iqeDTlrss%Ssd{
z^duFwhZO6`N0?o`9bMJ=<pD=*m!<fId@id&kykipnp?7`JdH@ZpX9eeYxqu|RB&#;
zkK^W*S)(X+P_52P%}Bd}1eu+k?Qa~IgWzKtMa_bZ{h^)Sz*8lxm$4?STqZA^1!gfU
z(Yjocmk<n{cfOpoB)lzCV$Nab+AXLo1M7CDeRVm(5%2lWfs56$8Z1djTkWDC$Y=Ao
zI6~gZEtKfDK*8z~0Y-+fNS?aGee`hg+D>x=QW=WzZp6~Lx(|5`ZnqX$y}=I7Rz7a-
zloI#L{GEbL#p}S*<beV4j9@7R!_?f196Ne=mV{YjdgVIu@hl3Ca|7hzfdhU@MzE-p
zn{i0!TR}j7$EQ{S6}RFbUT`qXhvcvbi(AT3E3O!89MS7v{07&u?*z><vx$MNfS&Rk
zOA-wznEYZa3WI=NQ51r}Yv6V<{1}iI6b(V5yw|j1C1`JIL-LN%Y!8#Yt1aGfAs1_F
z;9&G{V*<VrjcIViogz=xa-)qAk)<e3qJxM$g#k>C6h;MOq&bQ~4xK*4uq2pwqLhV@
zk9IIWTX{z!QFF_!8O*MLia5-&^^2WDibO4mmxUTdO~A%l31$-ufpRXJVdrcL+nBWd
z{@v(lvK+EAR&qhhkriXiuHIUEO!deuM+zJiMyQQ3iK#qMt`X1v<(2}5NDfsXliFjO
z#%ml&8IcOLz$`I~#Y%{k*EU`^2Rx>EjxASiyIaf}nVKNaCw3va)DGv)kchmqac!(s
zkro13d#Cp-GI?bOVMJaE12%6=3Imd)dj(c#eW#v#$OCn%X9khN?AmRRQX)8eK@0Lw
z-KvyQMTz)!B)z>MfqpL_3c46J12<K|T99WgVphzZCMTrauC(~z%Z=uD^^cY@7)7xI
zU9LRgL9%fxuX;Tj*t6oT-5at~yx|udRz$HU5ReJpvS>d<4OmPB-gPV{Uj`OMSN1hx
zqfAM_=d~WQtGAJlCY`QTFAFk+K^qnvB5PN=C_?#Y`zo=_+a}kjE#<iVQV<??$>Sc^
zkUXT-(2e61f`#=6-xuix+$JwEZc?X6WW}`hZUX6+6H+d<8ZLBXf(7NYF7L{e-iDdZ
z@|q+sA$bl6Bx)6Kl~_i@#T!Ggt}x!&$9QMI7dScrGDMV|Qo*|Of`d~j8lg)S@!*N$
z@+<sgU;+hRJ_Wo-COZlRCPwh3OX7hc=F^2<Q?S@5k_{Vv2Mn-g6vY%u<v!8vktysp
z4S0*2fPk=*?K4KfV4!x#6DI`t0n<k2l`T%;lX5usDs;&R>$qwP92(hZg6aMZaM$3}
zJ}y>&aP*j2OzoSwiRF;VJZv&Tqd8X|7=yAGnZM=9VicsD#LhWPloY!ucf>q;CM?D+
zUV3AYt|398C}vnLP9mxR!wJwbr9*T(G0m;($-A!Oi4_UF1V$MtVf*>)_1<Db)a7SU
zP2#P~r-R;FHs<c!FWV8)2@vOUwYYA8v2*Qq9lzNWAl)~vE8&jTD^@9R5WK}14gRSZ
z5hv08(*zG=H65*ho!}To?({Q5s2#9uWfS|?<ZT;XdI|@Xi}B5+m3Gh6W83mul4kh?
zS-FJPrP!$fm|nlm9fjXQlo~(3;Sh2GYNvPsE&;LO7Y*<%&@@;#fm~e&5ViTcW<bf@
zFeN}DGK=kxCK*x&7tg6|F*A0q-9}c7d%IRifN*lB3QvZ*T`Ltij!Ib$@_<4KQrs%|
zA-Oooi&sU5*-)Z8uX3c6sKaU@rM!`E^%513ehh~q0np$sc#S32Me?QWJHujr>{}65
zeq@z}C`OSQaFWq?WkEz>Fsr&GfnC(R28T94%S;-)Ok^b27ubS;+-yeES*3PcwF~&O
z$=GK0k4d^0Nb2IJs~FAp&JB(hK)7r1bH`1h6e(Y~$a=U{t?XNBlkr$-ZG_YjQ=*|l
zWcm|Pw_7NEXhK5C3PB`7C%x0&J9-48T{puFWZ4*5zCu~9nyPWp+DF7igbUTUHMS8$
z0|;&xZZ=;vld~RV(%V@Cc!{<7tgK$pAX^yb)tOKjAF4qGCXlM7Y%jqPaKHm4hYmHx
zPqY0S$w+x`fTY0@&kRim@~A&Jf<;W|igEaERV>GeNjdih2+|f1RQ(B^B1o(Evxmud
zcUleODAK|`(f=xzDbTuPTI*4@3pIQz&j9jb2Ca{6ewEmg4P-Q2LJV7X%k{@%yK7q>
z!tC>qhPOyz+17N_3!RjPX4kQijN9idj<W8S=kjq=OJY)46pL$t%MM^2&0emHpE6VY
z)O3HmAV9jo(FF_&9IMbrSLu2VsYrCXRjS)MQCpOwu-J-&2wtrAoP3whHyRZA?3kN{
z;5Z!?fK<2VL>-o!eEEw*!p;EE2zfq1wOZ6JB+JnUZsQ|j6Rj3p#A{j&S4&6$w632g
zfHeE$+gb&<1<$S(kR3Ot1S@A@wV+L$L>sH3Y}+6f>w%yJX;-zSYPnjB*;T*z=|TWm
z4-QV0riMGa3aS9bVr?y=FwZAK<bwr9F-1&KM6dPRRU=G;2-R)s;4BEvEvc%-U8y$T
zK-zWf(UH`96T`nwmZe7Z|7ZCWRaN;jLxlBk0XRXjMTRalX(IA@lH8z)Ol#5GSIXGh
zfrf;8MGXK|?fMwCtKlNdsy-&jC~A%2Z0@B1iFM~-0nNI?<jQpv)7_20AzUm<Z(p5t
z+akq)A{{eA%Mo>}B0Df82svWvQmB?Ga4W!*aLb~Lo8nf;9J_m}Z3h?WohCsScp3W<
z??!SULS9U<T+Tttb%%3v2&IbUtOz&mOymB@UEV98a&j#h-ok`{U$FRGwgkLZx3tL)
zZh$7eRb@V*;UkTVtexcuYTj!ORHB&dV7hmGV{jONP+HEVZeYkPPtQfN?gIt8b!38U
zcw9=!#S{k!Q7esuWNczhEF}b6OX~H2+tGiQQZb;ED6s)JqHz)y1>@odXH%4!u6JtF
z@3*leE-u0zPpMrN@i)LIOpPif6)g^O#0>DXb|hL_4l$;>rWZ|Wiw3kmkEG^5+5<_g
zXz4)ZGvAxmP<esbwObe}^*28sI6`3f>BuaT;#$dP<hx%K%<Vv%AGG2Isa>BU1yN#F
zi1$~O0K9jYHpzF(P(|+zlEqc={=jLM>O4_m4qPrT5QApy+NhXfxm-Bz!#*gid}<b^
zTac2o)Gk-8NjOE!xm#=p1=V0u*2QXgP931D$vKm94Ii5PFEXHFmpf+bM$9?YPYbIe
z1Aj!=xq2Ha%P*{2tq&A-fg@PBwl5Mj22};FlnNFxS?)k<lr9R1bfQ6npA{N+seI|S
z9c7QkE;_l1qyQA~tcZ9)DNEB8<DfN$gFA`43?8a1LzXdgRSGV`9%ryJFBsIsCcw^e
zSjV+_PJ)nyeA|$G?E-&_?Sw6cu9pm(ra(l1HXwnMdTRko*XkE(dsOG%i9|;tXkB4?
z<vNPVtPYY6^vGLsZU;EpJGyEoxP{;d+EhlZ&pOVrCbmLhCN9Z|uBnAbu1R@pSEr7c
zO{?8$J03zvWDRt7GFk6>Q9uzn8XE!pshWf=pF-<1!A}}xiM>=1#}oX;E+gd<IPCS}
z67H7lq+%4RI<V(j11as)s9HE^a@e$+I>(|ug_H{-5{ZvxjHbJo?AFCkj{6J-kTo;B
z!wRvms~%vg!zcm><tMy(31_O+^o?Til%x_Cvl{XXjTBEs-L)>x5CMkQ+$nA$I8@+z
zQx015nouT?ozl52u9yX!WeU1rROQMnDh*KPIEj@>we<vQw@X{zVXaBSWki6nap!8H
zF5YARt^rjU4q|wOs;viDn+}ld^|>GdSk=<6>&il+r&d$^G{WraO;7Od3Lsw0f-uKZ
zBqHFDeOi6qK|*1WHP1@8R<U9u6(3s`ra1=-nL-yuMJfYrT*HO`qApKYFq1g~%r+-#
z4IZT)+=AC2hN;`|G_HwPQS74OB4Zy}uAp^Y<wv1<1s$Z7G(kNLKdF)J>quG+aXe@f
z<das(l;m1#WNU5+tQt92f>Wqv3XYgtiPw1!)BT$u%D`OtK||V{tLN>gE+zcwcHTbA
zq4Y}r%?)Ict{A0S!61@C(j^tAS4(7&q-Z!mPM)M_pfTVxL>||OyMOXyQn8U_UQU9q
z_eewWqW|n~a<W)S6i1QDkmW#CE=9?(vl2wYjS|)*Olqv<Sy>PzIEc-e62P)Q6Mzb2
z)GS>=^wx6o4qFW!ZT7C|;MmlzR9rV*7)`)c<!?sJgSc9GnkP~_h$jHx^7Z93)esbs
z?itb*3+og@@=l0|1Z~%fDcIp7W5&Yq&m6i&Hh4n8+lBl~=ppk`e}F_2M(CWX76)l4
zLV04hfL*q4ZAEI=N_z|2x>0sfE5TO`Y1S-+4Jim{(K=LvI1^T>{;b-vkWHtrV>6Yo
z!0`r7wi$V|W{f*I+S$c;XAjy+*o|^Ffayrfyb;$*fWU!6+2|*c9L?2O6wy{+DG@JG
zLGcf@hdz4jIabw%NDh~XU<Kqq7E$MjJnQPLIXJfjS84Uz?FLSN0heMdqZ?<^tqAHo
z+?G)Pp5@AL5nU@px`|RZQEP3<w~a|Ou9DJTkF3U0kZ{ecwb9!07-9`lRh@?g4RB@R
zRgKlCoKl4<a;bu0afPgy)WN_ex6E^ec`z%I_q$o5y33nrszhu2%vlYR171|%cC~qo
zD`1M0T1LT<P^c&+v9@=2@&b$SYA;AyqD05|m|%c}B(T0?3lhL0zf`_6quLSlc9(dW
zD#FOJ44SEqG#eQff@`qk6V~_FzNE|WNw@$Kb3@8ax(YiBQge`+pfhWbuB_rnTp3VF
zqV3ZO=Gs-9UAYcY8Q4m?WRrm7jgWZ*fKUf9C*QrX7h}Pp0_%|p;AFm=WdWifJ8B?M
zx!l*&cCDfuB-M9!o!djy)&OL)hNy~_08KQ?-w`aNe%1rgsSdG>VCPm5g|e$lM`^&5
zX9_D`x#t>EEgROEJS!P5X?AQh3Re?l_ud;=4z45l+ynwm@fM*niIln~*dUOq>}UlJ
zz9a=k1&ztxRb<83#5<a_ciB>iJeRYTQhyG><lXvw&~=ZRNvH+@5){_8%8bRxtOc>J
z6RH-8>8K$HW4wc+7N)(p-NNFugd3TC$TCfZ;JPiw`UgjQZ>{4F6bb0{07xuAnY?9z
z#US5$cssqyYUx*(p6cZi358V&TrCSW3sOJO*Z{*?5@?XpMjEc&E3m2yuebu~$`W`6
z0je?|K{mq#sOvk%vnv=)cEP$7BN}A-35(ymDl(}TQ1TckDB96NiN)%`fCda8Gpblc
zTrFW$1d;b*Ctm-gb)f0gJtJ-vqkGmzz%@t=%84#=iWmp#HVw(3C9hGs0;|ZBxaI7{
zJz{NxNG^6_D9WMBW!nOfcG<F$yi0yz(muRW+lGu=vaHR%uMMvVMHb_cC{tXeYpYU=
z!A9{?05m~!-E?8HyKjYp`f{~V8--#=F6!S{QnFKdk*JL21VE@mf`n>;gjy<9X01wH
z*pdTk29Sax2M%~#Masc3WveD)6sD`FBG|Rp1=nJ=?-X?IagpM7av~d-F^d?GL_JBT
zC177|Y8CM)zt3%o^h-S=HaRrNwQ{bOhHTzI1Y}vAra>%SFcy-kvm_2k+IRt(-@v-L
zCN?ST#)?p5Ibd8ZL(;K2ara+WCLU6|T#8YvZRWNqHNtdn-&pv{)QVWsBARLI;Jc1A
z!5~Tig}Wl@VbnL|Aj);<qs%Ab793+_d!K`y>rcovb$;n{S>g2XK2RMxj*?x)^y)s4
zXYp`NS8N0wm3_+!^MeNfpG4k2!gzKav^(hlLdu0&BkdlYVR`ldz{=i=e6)}3>h1*y
zB`5feS?pQ^x&o>aSgoK}OO&exxLO&)zH)_9tb^=ZEl9ftX${UM7|ou-Y=6g(!Fa9U
zIcVG+0Jvn_fAiBQ4}TZTD?lkIR$yH-joK3^O(!JAboYuWep>lmYV2Zf2KbodMG8Xl
zO`B7&UzNDVn&fdeASBL0=C#%!14!}>K#i_qG`)=v{gt1_CtkkW&zp!mzUcGcf^Yez
zufXzfS^KDJjAFKn@BOR)8-DMjuk~LyAzui3%@Kevc;8#_t>5wiERU*sR}(Hxv3niA
z`l~PDZ~S*Z@4YXT^l_fYcl`Tbh0p!GXHg#CuW!U;51)GF5Fh&1|Ng?7eWnOSkzrKi
z7*9r+O$+SpO|XAuh8x#*aqH$D?%dwTom*FM?fO12Er8Vm*}(zk^J8=U?4s;8i0Ftq
zQ-#^?O?>>*$M`GX`D5rFO`D#D{?51E!FPSz2XMT21=Zp{Kw0f#sXEKI{<#hYj;2?T
zkEh_OtRv>0!lj)xVdD)%Tl*w{K#3CQPLfU(iHU@+l@j2{%-b?#it*z=^Q-vS$KSZ{
zjlOqqif{Y24+1nYl3;<X$naBt|6kzOfAbR<^6_iG@d<qUH-9CvYz(a{@4Lum6a3<@
zd;~xGcm5v?{rJ;=>I?9`FMbg?9RpyfQGt8+kMW~F@k`j|BhM6W-rU2BFWkl#eg0eV
zm0$jDeCe0G1A9;1#Pa@qEYFT>zR<*j*ttw~jH5zP6u5Wq9)9d6f91mG@p<ok79aYK
z4}w(TyUtuWaA-z8o?$%OsZ}nOp{bW*$=+>z&;%e^%zt6&K64}p9J40w`Mj;(Q#`eU
z$jQ2)#7mhCC^CTkz3E0bad&5IB*xr)bOtcv%HCw~1@q?JBRsfwgezC3(5tHTUA*$?
z`&(J4MUfdvxv+OR%M`Z#C@YQEUO&KVuOHx-e&wV1z90NQ@Zwu;<1haCKZ<Yu#`ojq
zou_d8=3RhSR+%CQsh8>krc7<tNP2H~V&vr<N{#BOMMVIsI@?MEDpQ#3>_C?L$`p4e
zUWu;t)I{e^Yam;h$f6;h4qDTuL88+0#^I-s#9S<7&edv^ssV1IqbmiVM)iNU^mya$
zF<yD~K2$LR)EH1BkQ&vyTBY&Q%WrP!lN;r3RR81bq0a<w9v}PoXYikX?@!?izU)84
z&-}fAj@>)Ap|Tl3h1uap=3fC)BXh5Z`3~w$u>KrU1wgs^oC4$7K2$am>np28AoDX(
zCN}EK%K;3B1=%2Xs(ouzH6+<Ggd*;h@?9*ObST}e%mxLC2HpWxVHVIfXN@vmdigF;
z<VNf@U$9cZe2JG|xwj>JG_wVB^G`jjv3$wLCtkjbul)-@fbajle}LJ|s~|N7XjCs=
zYQRzW66z*j0Rd3C14aW#WEf3$k&UKcHsGM#fGKk#5vdv^GQ}xOdlG6a*{0np9j7{P
zstHjfq8<kSyc5HKyGz4$^i~(o$3FfBko!&HvJ5yrI>oE6J=mfJ=`*nN;Xl6PyZ;y5
zxqSs+_vin0oV<Aq$l7(=vS^s|rN&}#-J(=$AR84ZCgVC7sG1gAb<ogd;P_<?Vj8t=
zwd$l`G3dxdSePQSt%E!+AlluX+A3CPNkr-hc`ny5J4+9hbkN5~|KRnO(xEbiH}9R`
zjl0L&`{2uo4*<Y7f5(sD;~#ky<Js7_x{Vbn%nm%veI{c;8Y`l*F-FrXU^JFB15s|R
zNbP1ht76?RX-P;0Ssd;BzSfwzloJz;0tdxQO}S$-r%Hn5F?@XNr8j`G^xdCA;nmj<
zuvn}f!WnSTHRaJi4v*&e%isIc$ftSjnotH(^<w2AJq5~*8(o8BG~I(zg>k29x2til
zNC}V#U`VmZ%jS^lHOL|-n}wdJe<JcNNNea~EDjQ+hLEfnv?S^Adwk;MyC_c<P)ek+
z<O-j9<?fb0`Eq|>Kk?JQijVyE%NUO{@05p}K1-BC8S9accaUWhFx6zdcE57V(`u09
zg11SKWDX>dLF|wgL?-Yii|`0)UG*BZGJGSTk<s+Wa3OvAwFkI+?*v&cQd<<YgPmKs
zCK}YIAM&v(EBu3>`z>U%BG^2K`wEzHFy+WcQ{<x=n6g?3Xt)VUK*T_i8#POSL`j7!
zgDlUG<ehr7vs4LvH8TCj2ZvW1B2C(3_;_$|iq~F00OkHpxjl(j$T+b%hnMK_OTY5}
zxv>UY`3f?qe2jdwV-|a1I7i_uM`KM$S`)`X?C&<KJPRy{1c}e7LETm)?}T?S?bNRX
zRKiRNLiCurBQJmY9zORA-;SoX0A+=jK6!V$;D{0S4Ih8`yS^5G{43szPksD#{L(Lf
z1b_QSe*vrV;w?D;;1h44JYAS%K=Rx04DnDoifmjLVrkQFS$mTrg==RRn^CKXL~bIu
zek-$yth?Z%Vhv5o7L1{pzUw3}*ZVgJAm;8~vA5{`8%gc9p)yv|^6!QQ#`Ys<FMaY&
zAXkPf6ZGj4pW0FsJ?Ux<@Oa;Qp2z#Y@{53Kj<5fwuf|us|K0dA-|z#_`eNVV8*d)r
z=y-vh*$h6;Ha}lxATOpSm$0nUOPFjodN7%wH+eEIhl;GIf1zgnNx1)58JKNKii^BA
z>)%gSWCqlH!B~}?XcH9!M8rh3Iw=mdZXTY?@$r}5u*rxjQ#d$0#cQ8=uzheOn?YRc
zaeB4{$~ji|7pU$Y;%olwm*De0_l1qrtfP}TR2J*0yrJ>Pa%8!Yk*!))c)vrsSB9Xg
zfUJ|?CvoB~-7Q~mqM(6DgO}xsakgM6*1G;>$*6QtxrfNlq(tZQAPr?u73<;p<F(XV
z8LO}G=9@>jd++2S;3L)me3*x309l52yzPk#A7egW;p}t?HLyZSM4;SMe3KHOC=BN)
z3|K}s{ipuj@eHYQk%^BuA|DAsP}q4#gfa~0E-PR@XRNB)Jmt;iuPb2A8hNfz6wT-j
zik7D$*=m#?Sw7Lm7UYv}LZ2?`G0J%z11ZaDtKdjmlA#V<_ba=$zjNWEEmtMx^A%L4
z;JS@$ZbgXvncz*qQ=Suw@fawSoBz_(aOetHF0ed3#PaMATpokCs%2r*0}R6D3VZ&!
z8W-i7-x+#QR*cn(QB}1!)ocj4b&<@YEEyFuih?o98<W#+9+&+2@f5P~%B%Nr@BRsH
zT)zqw89w=`d)xm|ukGqU2qAG?(@SbP9c^^Tv&9Oq02!7uH9esxGX;#tIC%X4AO4M3
z@p~VA4WInfJ=}Y60?ojc-3gw0Vju5(`;&OjyPn4Fr=A4Pj&XedQ{d%&Ta!eZ+ngk*
zxp$2%7*$1BmW;Bjo7pL4vOx^{&ncLxq0E0*E1=REd66N{&GQuLvM8(vJ+9~gaPR&J
zUVZHVH}6~ri13MROT4kExisXl-G4@-&5}JAi&b4WZl^>I^k5**7}?b^e)FGy3_tik
z`~v>rKl<&po70_JSMb62zYE{|?|vD+_)Fe}<-Jd1e)y6BK<@Vt4n{N(>NZDJMOdx?
zUDgTxibVfkG3#BC@RFnvXk4t-im_NSmSwXcNuT*4BojR(%rP@Q`SM*;(#sDa2J$ej
zYMA@7TGa(qO3`5F&dk7U1f2nY>AQa%U+@+GC4T5f{%J4E{rVe+_@92_m+>Y4#(#}(
z|IQx;#y2s&{tlZJO${dzBweWtgbbx7%$J1u5;f;m1Vx!x*Jh2IYs~%5e@b{v<oSv#
z<9cPxwMFs#$gnfw<GH9l_R?p7Tmh?c%Te_3I~&-=0-}>&FAR`UTNbL+Fz30uGr`fp
z8NTM<{yzNUU;O<GALj>t_@CfoFMS3-^FMz(M!U~qdGv8pd1&HU$e_6{6e~-hs%m++
z_K1|3YUKuu+K!B_A3Av?k%%loX<S(XS-w<NgvE-nT-A)9TWKhd@8hE%dmYFK<;eoC
zy!sGgAVk{%KhdCWTUBY7yf@G$_~zYX{F`6%*ERx2^Z41H{~dhjulyt?x1R$Q`!;&6
zIm5D4SS*26S;tsZ(|p9pzKyjYEBm^cn?&RWlC`^OlCUBW2QqOSuC2;)S;tW-y`4>V
zDAlC51<{OygHycz`oY@Qk4O0cWwOJ%sI-M+m)^~#LBz^m{<S~AFZ}XH2fKHF?Qi}g
z{La7l1ZMlspwtS>lCUTVl_uk2DVv`~wy7a`V5l+TD~gJ|erfG#v^EGx(l;a1tEm22
zYXg=N*rqai(Z}BI1Vz5SO!TFf-vkyV-ne^=`v>b|AkRN{3zJE)*>WD-@D8mq#ae!;
zOP9D^X^p@6w|@cIbRQ*8z+CrbQ<T|cMp7o*SCO?M4k(B;Jh#KlgZ8&@HSls^jBk=$
z<k7li46x>bT;Ji`k7u90j%!zEy`<j7;Tc|k<Dge+#yj8sWG~>Q-JsNv)6HG1Z7*O&
zm#_BcAOGU-;q+*UVzgsRYe^{q1Z7-A(^TC9h6~o(6>n$0#wHyjuq<MJImsD(M6H3g
z+hOAuKXB6zYTe`T@C>iNafnxYxK{x1))#JHILB!_@@@)*ThJFSA%}eJ^+Vjfe~Mx>
z4O}V#2t}su$NxuOBXvhByn?A%-v-}cQ?e`2w(s>OkNv$Vo_pqcuh)L^Q+M&{9t2$}
z!t>AF+92A#B?}UpM{-HElT0am@co~IulbYj!*fqx-}rc^^CeDC7s#^0s}X6e2#Kiz
zGLO(3+hi476G?B(OcRn<)K(%n94s616zQgIg{g10OOM9|p6Fpg0N^8k@H#&A>GfLt
z-tH97JaxT}5o{zh1%rex>%&Iphdboo`o4b$-}tQ`0M3>;IXuIE^xZ#>|Mq|Urx!ln
zYE@yrD50`RJ+z5Yr|@d!#s$`sK}9;fufnpKlhA-fMb>*OX_JD#`$mL(-)w+;teJsQ
z*J&>m*KW)PH%2MK+unM+*K2?0!=LIEMZbAt54Uga2MW&xEXX!UI??d{8wWpZc-ZS#
zckpk0)xV0>XHIeS#u7WD3I3CB`znmb7mxaD%~-BVL&l}G@aNHx64|`uC+Ic=-f^v1
z=W$Q!ZS#_(d_S-Wsi%0=iDX=A@CDamRcXBVmVN;Fe}C%}`1mK*w@rQW&Q)B!IvXJM
zVhbcaxEGhvF5SL$1>^Au%T?V*==7|_^=mtL>dw`Tj$2mRxKx>+2Bl;hDYLG{)MXde
z@IxCapxsf6ky`(96ack`YYwpyl2gbw*bJAe5-&V=v)5~X{Wo3`ZGh5`=bpWZ*=)Qa
zSD=Tqf@v_YH9?R~_l3Q^F~*}DF6)*tofNosb!W2!vz%b81Tf4T5=DMD6Ow~yA+kuC
zY+g^9Xs^OpR}@N3NZ7_rji%e56Cle9cWzz9&dwMoCyUP4J~%k-b<7u@zg73w-PC2F
zZCH?Dfu}VCRh{w5%qU8rtTsc)WL(ri)|GP$@?2qOHro8^?V@j*#~GK38_wbCLRij$
zB(me*tc&H{F!nU9sB7D{DvW5U$RMO}m}8cu#`PP!c=FEvVDNp*^S4n|8{NgNn2vge
zP0iaH?zZV}&-OSR@i#86U{O%vt(4kq(}{sZO1XBZtS^j^q9})Ubs+?gnD|0rJTgmx
zhIq;Pp7|U^YpA^`o_prTV8?v+*;`mFs|`R~54KsY>VH^ilN7k=wbr}+s$#4v1C~K}
z5^p0&-|fv^c;5+0crfmdC_y7A*pGu&h}1D%*)IwuJKTq8cmon?VYm4sFYwmqZx41O
zpL^y8O4*I^!Z$L^>K*F4Bd8%gHoKGZU?6GbF%OkUfUr-3?Cpl2&5ysj!`P8Z-8}O8
zH3c8u<uPn?ciHg$eU+?rkKLUKp1gAv<)%A+kmfMp3fY1+@A4-$1l`b!#^wu>9UDlX
z<Q25S7RknONuadHMLvG%?HVt>FbqKM+}X$ejlDVzd_xXGu@Ui``WQrmMsK%rmwCPC
zve`L@QF9*-O<Fsx8N03eMBqYKiaT7Qp<7c9=6#eEp6@C3u6sQ5)D6&d0<HCii;jYn
z>tQ$1L4g<CXriHd^A2x3C>kscHzZoK1(@*%MCVf^)-6~X?kb0C3t;-WJoV&N?Cowo
z#;pe0XrPrbtjtSM_DO@?52COw?IrS|qYMYFQjI>zuq3m2*N;lzwEHYXj?4Ne@*L2N
zTi5sS<P%pnKHghjxQ%VpTRaX9D_zogCh>D6P^-su4+EeQB;8x|QH;o732)U@4o)_i
zG2WZv`2nQfw>*Cvbzj#3Kq1@o484C2R^~3!F9zP=X&W)ppp{juwuo95CVmd86G`VH
zT)h2VYopc3z$i1j|7MTpp1TR48sMhVHujX^cX;aKWn4->tJk{f`b4$v98+QlW1FDZ
zUix?6tlp%%R%Cek$?I@k1_zRIz>kS2NLxC}HZqU9EGzeVoAO<z;D&jYfjzf?yu<SP
zGuxPK(PrNm0jM^YdT-shg4;Lufl3c@xr4?QVrtGn96cy1z2w8hxV0`ND0m%`c%22|
zejjd9zmh@4OZR=P^u|)}Q%_#QmHio@Hw_X78x!+jeXLqH@-ga{Oaoqf%`4b)(6}M!
zUi2#Y#mWORcYNUzFqbwVwLa6b!c$LN#r~D)g^%&<GdDrg5x_u}tBu$0R#zCUA)37p
z$y)`)kkrOcS!sJgQncywP~A6<JAMwD0w<Aqd3KemWZby6gQuRncHv{Z<%L^7mel~+
zIN3>lciVfEdJ9d<F11y1>yo9r6)LgJ9TY*%Ne?YV>*~XUNB%v88@=g6Go#p@;`wJU
zF7*zTuF^(554u3)t~_5XH#(Y9k7375)`@kBHL5EI79;@<Mi*$D&&(~5T)r=pWKf{U
z@!|`2E_|#Pp1XxX+DLNvy$!!DH}~e~Xui>Lay#foHV(YZpfMg!T^tiwiqK-jHh0$p
zLG`nv!X$CRa#*kox48~`58S!EeFLfYnWwJ<rQX6uFd#UjzF%qFy?4CP;fg#9eD>kb
zbx<P0@YbaROA07$U^Kt0zW0q|=woc3gV_259zd=npwt%*7GB+-;fXs}y)x0E-R8lL
z!={DPVue>ez1c9J@i-HcGq<_v^X85C?7?)L26$f2Iv!#z2uDDXHc1YRjrr2zgf8Q*
zDsF69;rVB8T*x{-apx+oU)zPQ^nj~*dx}Oj%JAUc5kC5{S2sH3Y&tdr4+j(uD(zb(
z0$|$q4CP*O!$VS@kNo@uFH!z|@&UoIZHe1x<iU0P?U8*hHJ-e)kE<7$+xN`V*OBi`
z1O*x3Jb73PJe`gJnxeQi24)ldH-GE@#KGa&g^!nI3OhTa!DBRMXG^ouxXadB<7~d#
zkS}F5LAT)WEeQ+G$AJ-GZkzMG?Mt~iZl*Pz|9oh^nv<c%iVSEy+M}yC23%D!u3z89
z(@$Q*{RiipV*A2#w}4RwoK(XA13b8k^_#!_NnH84e}<2I_+|X;Km2w4{r~;fH#*+V
zY=phNiLoHVxzP^}&#+jO$ny~@)b8M7QR3k6^x|Mv3X^dT)|Cg6+LGwFC9(WDb`u!+
zdh%foKn&tw7@`0Qv))1^Se^?FavZG*qbp;)@Z1gj_J==pu6H+S?lh>C<3SPfJ>UCx
zf$#mhgMIT?_h;DOn}Vxh@9nFv9pLcj40mo%P?Z&mNrr<r7I^J5hZjEPWSnC*8>6a9
z;r8o*3UJ!(LVM3rR(p#1V2vqa85q<7AR44@gM)18k!KgGPk+nvw*bu|D{UY)jqL_n
zH+ej9=PGvgrggWJ0Usx)3;f7W{4?O%1d|(MXkz@0AO0seUemX%>#?^t!QSo!Rkd^$
zM66Txe&;~7ir<606@Djp{u+fNbQ`yd1UN|79j;k#_ihciX}|3)=Z}Rv_sk7oRoTsi
zE-6nR^6}2M-2rA}@M1am$$r;={=eeqf8oQpadQv9`CFgF@BaQ~V#2K(yV&0yV|lW)
zZk6X=!BGU}z?vl7-yDLjkVKN<AS}SaYyd+nfYJO6Ft#Lr4qb{_YP|J@Tj!KIi#kQ6
z)b{(?DDD#y^Q|N+ZvObvFMbEMDewN=|Mk0r9qX+x+(N!nV16(co=v?z+5FBe`Bw)b
zlg~j^lqJRzygX^wCqFFB+2in7X*~DL^>d|A+`hSoTQ~N=<raj9OKL@C(-FS<1MdZ&
ztRf(}d?)=y?|T6t4b}@oyfwg;?`pOFrQECfZ;?1iOe@wt+C&V5?Fm%1bs2Qdyf*(f
zJ7|*@!HwWhR(SHxKCWNe?R|`=p1fZB;+4iA{XD&m=yIauo4(<X<Eghlg>tbn3(fqy
zq=^1gA9yeH@d2Qh9ymB4?Cizu+VZdT4ODwS2!etkSXfBAo=UI<3!+OvF9E%oteWSU
z8JtoAyhK%&xOR01&zvKx_xWdU0Fx28IVTKIFvS|oC78=g=U%=2t#|M}ANn)UM`vcC
zt<3k=C`6mNU87Mx==px#*M0%s`*|<m^ze1FX;%SJ7i*R|P+!xQQF%TwBT@$<V;=&6
z?0~_n!Kg5J90aAiF`Kat^JSD^y@J+Ds97!bo=d^X$ehn=<IRAq6>hHa^6bp-c<~N?
z_NV?s>|fhKIiK5WX7g`|jqpGI)(_%C-}xs8Ip61>xsLDuuf7he!#ALp4~#XaLBd>l
zN-N6^o?`)Uf9>aSf0Aoeio_eKiHB$qE1_UfeIkJ2Bm)s?!0fD*(-Jl=zo>66YX#&A
zFP=lmdg}|f0bPOh%6wnr4bR2Qz0)gmT?HO>CNu%~#;^ZU{MG;B&*A#bT`V8mHydK;
zS!~{0?2rACx8s9<_DgVdvcO;ep`YIf96$Sa{sY{;HO1j89|2%22P&apRdyryLlPSC
zbSmL}kjl)P8(BB>i)*q&PzH5W6Cn_YEBDPZtOeOb%Lahcvt@lyZFbhI{`KVaV#lrD
zM`jnSYJjuGTb}P<e*4_hH-NHs(He_eZBU3*mZjNr%(b$Ht{%;h`M1A#2Os>3_uw!7
zg)hbDfAQOa<6|uEziCv;K)^9x#3pf(OewtgUC#gyj`028{|$KWyPw5h`ktS}8+VV+
zb)2vM6Ys_kec#vPnWy(~`03vQ_{?}hr3H!OA(i7E4HuQki!4co`#GN<-a|2&Asfxy
zN|*5v79eS+eVVvOGcCe-X%#dR-v3A6iD#d_7IY<HB+WqZ)E9o<TcB4}%|Unt5J8uW
zKlVr7f&IN{`i+Ul7k}Z4_3?}k%>eY-9AEgp7x5>)>~lJ7m}U8R$J?HSK05{LWsTi(
zh5fx5zVgf7-EIlqLz}*H`x<!JxN88HC3bgb_{uMT&v_Vr_tT=G-)NL$Iw`QbGr{$1
zJ9zSmt9aqLn|SBjp2Ry|yo2%0UEu5#t2bVCR+2fYV(PLo*Z!{FrzQXNue{?4Jn_Ue
zERPR>6T&zCdtZSMy#GD;!5{v4{N&&Jway7&S*GwMU-VXd%Qt>G{@kDb*TBmI9KP}!
zfIc<nTsqgvx>r@3he!(|lN!Y+LLvqF{U0qj%L|Nm_A%bQhGIMgTssESW-hCzhEnE!
zMh2+*_hwK#)CAey=VIr|b;oWf1WBrAsm!?5I6FQBmnKb{6sWk4>F$2b17hz9sTGOO
zDmCUOhd_DH0iR|VO>aR>N0H-&Bej6)bb)ey7l5&1YJzNZ1NqK0cuS;bQf#rDEwNfW
z0MHaOjiHKtb39%z4aZhwm<<tT%hnTN%qeN0Tmg$Y^!yanY5_(WX>P#n=V%HwzKMKi
z(h?ndfm#ELqa&a?GK)0Bcy<NY+r!!2Q~b_{U%_wx?#uXtPrQM959)<~^~wxyfAJ1J
z|Gm%S-S2t|$Sa(F=94JruN#iafG&;bQCXmXE3j_3NL33y4QJMs;U$$5pG*fFU9GS>
zI{|(4$EMtv9LfufrdKf8yNY7615jqtOtKmP<iOC(pl<$VXHaJ&P`+2UoNs{B{LM`K
zv7NUmm=i8KHLza#&9F&<)JTk;ru9Q5$SoA+l>;c%9Dr0|u7f-<_-(k!T1RjhOfW4#
zDuZl{9B(Y#?(A3l!)5Td+-yRzu7H%@s5ud{ViEi`q?D!VD1&L<5gQ2~fomoAxg4bO
z;G$xiLYTQm=GZYXX2^>P^4SiM7l2as^#B9q5`21u)02Cs76$;E%}UqzhNb1;<~i+K
z6kWCABQ;nI&n7}SKf&39H?UgHK_C6G3EL)Z%}{iep_uMsvUd%m$u6KWQ%&1g4yyla
zKm&7A>&4czoX-T1ly7OEOcP`Y6tuuX-t#PWdGbe42Mdy=$YAdoiQ|x+`ArmbrV4H_
zQDTu&aIA1(W2c3SeU|JXOUch<PQ_9XxI#@TOsru+P;D12EJB>XGPPZ1dwcv%&b^l0
zpE=!0Jyey-vJ_x1Yl)p#W`VPfzttj<S^NSZ{?C%ES3=dQtp^dny2AYEJ{Cv!p`k&f
zBH6sE-U(M{Cs-}cFq+IT-noWix(iZSP2@EIWe7V=Zjw@tx{wk#z-Z>AZPFtVgISpq
z<Msk&5*s7=i^(i%4S(WQ%lB?Jz2^AL78Ld~s2&_o!SUryXKz;&=gJiH`7~$4qN^IO
z>1)Zs36mIZAG02vRj>Y{rp^u??%puE`g5f-UF*ve*U5fRwl+8vfFtES#*+pKD`?Ay
zL~^XwR;`KyLV&=&uM!qi<`mXkqnw}M^x#cY%QG{_P$81;E=;o-w!T`-Q7&FbHa@_3
z=PHWn9w;9f#j7$?MB#F`Oe}?zan0-nH!ZW3$#AJ%t$NC6u&CskX%-;69ocDZB&Y_(
z7Y~bdDnTL^HjCri&5+y2Yc(ohVRb=v!K?*tYn{t$r*?x~?wax&EG|ZxJ9zP}1UL{j
zTGI(@6}v`h;R5T=iSv(|#9Coe#LzN&@8wig!h&co($5@mp=|1jb1g{`SZkae-o^ar
z0RUaQ%;KE<WJR}<T=uT%WKpe_I6Zg+*~t+`vnv?w>>|s@_2pY~W|%oKSaUsjm$g6b
z$NQPwnBH)qn{(lC!3ALHt7NgasGA$>FGgayE4GPugD)S9!o_ONS-2K0jUs5;+9vbl
zxh7E{buBJ733eRm-qCWa!oMNmjv?6LBb!KqH?e1RBeyGY4b9I=F?6}I9ZGDV4?Z}!
zu^P?uHn1JWKpF#}uv(np^ue2`7H73VX!udQqEHaprZ|dqz2g&GKc1@<=12FiI6KC8
zx{J|l4|y?hVSBhH3-v`Q@kJA;2`+jwfMW|k;bM?#0jQTNBMno_4GMNc=_W+hBvA^m
zA?O+xtv#WlEkuPpECzlGECxB%-YlxD@pEl}x9!`v1AnpZnL6SANU$=}orM#>%hphF
zLS@!XeGX*HD$JtS#$5sn6-xUNb8B%b6W?=|gNVSo!r9S%%#ZGabp=u?>Hp>a=^5XZ
zf*KGTK?hhF6-t+wpB!L0KS43x!FaZZd_1WE#g*rD+CZa}k6D468F8GLh}308#NTZR
zi`@O~boi9QEP>g44vAwyC@9q4^t}0Y)V53r&nNKPlagEDYIZC-=2&WWpgw6%xL9p)
zaICtK(M|&=3DlU|DTujzUUsnOR0aoHsZL4(Rkad%fnlwO1W=d)<tzsgp<JBe^x!V4
z#Th_sHallQxm7S_{e}S90wj$^VPcThrk{*aqF7@&KgMctiefy&Xts-dGy@^4fkL`=
zqgY#+xBjxUaJk&BM4>fk2)hMS+saq~F$*3Lu>h8j1iQWmMrsUTu#0ocwPHet1&p05
z;sYK9|7kafq7KoD+xcd(woHgv4d$qIXcK$?bJzn-Y={9lFm#Fw@dnZyQQnA*uv3+y
z8C;NjD6B@aDaWkF;`qQ=4f1<s#rl46q$a(wBGDRjt!v7_zPTl%S}hkRSS`<xk0uyR
z_fU+dAj-|j%0<?EajLdVp0C{y?{%?&gq7hgUxvCt5}Zdet{Ey`%@q)}{c`q$kgY>9
ziS>MeK?rscg#eKEC{oVA9C34WQWcB4g^l9pNO;u!>bxhR#W0&^@YzZcs<IM*ke0%M
z#;Q~%1Dq70oS$KS_+}l~Fk*|7dgCc3zKhkviU7KfZ7@TVhO$;bT3!VrM!B4$TrH93
zBNWpe6yq5v%MH2(nKxTJCwEQZs_1ZGw_nFmb%=vFi2w=5xP%{9iI&xJd#o^fE?=Zg
zpxCVAZWeV01`~>d4UYR+sNwDABneIfCk2j(YsGCl7PTH-5_RM=EnhNntWtn3OMgwW
zRJYi(8k#FC4j*7~`~aG@^?3Z0cLZHQ{f#4!fB%w;X!D&YP#pj<i&eYLL)5ObR)+Lw
zl;r}|@e0ec6BOeaipdmNUYJZvv5=^?CZ1<v6(w>qXQ(o*b7kZ@<~1M@pXDEh+SG;7
zcBd?{Ct*Yq10*>C4#N8Q2e=(;4AiIi^skglm2j&%El^+%I1<oJSrX<BwVLBW&qjdA
z2)b-|M(sv%02IeH&Tw{g59Q*_4SZs0abgy94M~?PtI2RA1f<;%ErvGqf^96sx?UP~
zkRk!#YKg_^3d_X_M&qdg$_VXdU7<U~R<&ZpQ$Rc-EUh31F_7ra-vA&cGy86lQ@}zO
ztD9hR(i%|QG9_OYszwaOJ{9J!Et@$ZuW}-vD`q)3PU>Qpz%j6-01{aT9M1<+icnRR
zE5#6X=QY>Rm>(ZtesTcT#%dt!u&5*AL_CKr5A@3xNXV)WY+@QR-l5$$qHEsbCQB<v
zpqr_qdWFT=3aj}E^3epvWQM#LnatqQb93Czty@7U>uM$VNGORiiC$J5A?6hF7fQGe
zND%nQd)LaI9m$$W>mid(IXF#Tlfr?xMUnTvCFh=J;7qfrc3DYI2Rg0t7;_gLKT?^3
zF1b!Xlu>sgl*=>BkM5&fo<a66V`isr1=#`&?ug5TBs~NP!|)ANe}GbgKq;}AYL6td
zjC-I?E3pIUO@Y;Fae{JrhI}+aF`gkCjX|Vrj=lv6$Nh4vYUw7_LJ<l~BDNE_J5rKm
z8NjD_(jB2KB*YeWXa$H8q?8iGmDU#)HeM<Bm~jzx&v;FmM~ri%I0FR>jUo~*CFDCT
z#EI1W4i^(?E;q|Ftjx@lns95(j}Ngpt*r){vWym?XUkQhZezsV%G>f~!ji}XSZ+cd
z;pQfCENVsAM&y-C$uNnC42{)tj&iwxDhd?iDT>j=0YzF5N~CWzAn~?SLaY?UonItb
z(8|e9GH&w*3O4o#Py%3u8RxBZnlA9fMOJI0*MlNvTR5T`^Z^IQKaMEjqQ#>A6foU{
z?AuD44+e57f>L$7ty;}-dUzkza&BjxJBtx%U>zgm9TqZ=Yy1Kvtihx_R8zBUH(Z2O
zU<%3QUzsnEHH*wyCDbfaZ?RgSS}n26Yo3yi#!xCV$B%Fb6NjJ_QTGB<<b%Y1=M<=S
zp;W2cR-y%zCT_Au(rg(9>IIH=yhG-rr7S_0Rq2{Dc^=Cl*A+)C2PJq1Q>cdNfVYHT
z)ILWcLRMrrzV{{;r$;r{P-MDZYQIGl3|M5N^|j!Lageo^q_Gla|6=WP5gQ;<!u{7G
zQqBSxY!)iJj9Mc#aJ9nxWQA;=)jVZ924#f{f_XfZ1Ro(+BOfk8VvqQ{_i#Y&5l@z6
z4y=O)w4Z7phU|!Y$emq2RDz~Zcu&?Icm0r<I4}~0Sz-JmRGeCzgj)(}8w1=OU6(E#
z)uK<eSU|6q;Hs+Qa9UIc+LBuB<yLX5J#Gg$A}Uq_B-V!)xjfXELYZ_}gF@L#M0pGn
zu01<h*GAHtD0&SRt*bghzdS=e8lxCZkmb3tmK=|yM;92i6Awe8&k>nbhy`s0E3g19
zw~kzX6!j$OsYseIaH<m!)XQav`S-j|@k;O+PR`eeTLSJB5<0-4n>D0SE*7X(3xI*q
zYzM35nGFiE1c&V5#@fqQ>lQA0%Dn9ZHEtCP8=wRRv;~T8+%P7O924o<S}*{r3@99c
z-0rVj>V>YbnxCRvoFOa5D8^%CdEpa!kc-GSJ2wSg*qQuQ&W-9uPAZMO#>v{D(iM|D
zBda>)!&A>qznPWSiqeM<oxaftSBX=KQV+7v>HL@IvABaHt&@L=s4jtBoIzKmTTe=%
zm`t%aKD1=LJ#tO=RP@lrK~it<fl~7Ua;1>O-a4q)jBvL^tptd`nnWfqyzW9a>>+$o
zhH`EgfYtI0<!XT}FEAR9krgBH#gNN7Zn9jx^z@WFwZ`T02)u}DgvobanhHxh65s95
z;c_WTja*nWH;~g$^s^eAIBb{Y0K%qv>%oz@HsiglmJ3wNg^SAx3pN_husS<|u2un`
za+Q%TMZP~U)&QgfC<2AdZ1S+$R^H>&*4mieo<$ifH|g|3;j%7|_DFS;d0CZ9oK-7i
zS&m{nMs9)P7O~fqzDM5(28(Jq4pb?2RdW<1eM^LQe*Gi_u72p?1{~A_3>~ge=Won?
zCn85I2dCpM)CG=pt`%(OnAKu#tcFkcZ6KQzM$;Xf9p1BDcWix<?V7014UC=uX<s;#
zGD(aoNP7-|&{&X$e^8SPBdm-U7?Q6Q@efJjMJhF+PJymYS5R4wq8Onl#<ncjPb*_l
zRx5)Zz8l0WTr=i?zc6AIg6uU%==$opV?1}5G0axuSJWp+*KIBOxl`yQ|D2p-i8$$H
z^5S7_IXMwoH`iM(7APC5L1;(5o1b-sVm!gB7@;&$uMB+hhJBPT&>&T0TtXQuB3&u%
zXV}_w5%%0Jm*P<YB#vW2<iDo|3DBey`^r~|MOC6)&TF1B8bPTn5I5PgLe9)_B&djh
zn?zKbASmcru8{DbYjwMyl_K<NG&LkJdVwQF;={~X?ue(f7jDfpN~{)hR1MdVooOO*
zoMuxV&2~{P&uqzev>VF>9_sw~5OQnei{hZSF9#PR{U2s-OL=mtRXAokP+;V1>LIC&
z)#40iCr4PEo!T$iZ*IrwG>4!_h+|o>iTFXRd%pdS`216(G8H6Ab`GOzKgk!h8iRnN
zJ8;<oq1pD!<s8fT3A8SQ5{QUvCzB~z*GlJjT1V3DrEWfL;DRl-g)3l8DT%@#o5N-6
z1lV>GvA%wjPQpz*1deE%_gd0UjG&W`3qS_1vBYk*Kv}Mk<vEI?R=!}Otkzm}sa*0)
z_Hw^CBJ1*c>6ZKZz)BW}0n-pH@AWqw>1X>K=Uilb2yn2l8WmQ{d0h|{kl9wWog+en
zd2Y0`ht=ZL1Pq56QITI1A95kXq?8b$1&De|m6UTue=A#a;sDe1{%va6SrTCjpl*<`
z;1T6&g|aM>=Q;9Xj4aO$p?5dPlhlBS+vR}b4e{h4DCXxG_aNjH4F)P(BYgqtao))K
zI$&%I98#fLEv%X-Jrh56!NH~Z?TnZiSyo^)-NoYQz!b^g5dlQ_>g}XGnZFfg)XFSj
z2a+b^w{2BK<eMG8sYyKd0Ex>9F3T0FvV_WV<i!YiQCOfvtw{nX_C^LyKy-<SWWIQ#
zMUJ>u**Pu5c|g$(jB~kE8-PPv4P8}Eif8V&w9FB(igRF%@pKo<vlHlY)o#=DPyrHg
zxttX-yQ;J@UWg&}H`m~enS~TALBcjaED4h4cA4i%bRyAJiDgxyTxH0MkqJ5`693i#
z#Xc2FN9hR4am<1+vMjrR%gpCXH|Y(GepaJ5IK*SMTA*4kg+di@PK5lS+qIJ5A(de~
z+r`<zy)~UJw?)#|5O(GW(dF>VT}Rrr_|mxjS+U`Z4F4G7d6pf#8;xC8SkBKdKRv;6
zF^7&$vEl@g4;BLl&DA(Iha2hRgQ1jKv*SfZ<8dE1in~~QkiZw*qh8?PUf>`j$u1XX
zC>M)BIP3|Ih-(#Ijn);4=?=2dc#wHDn_G}3Pb=8XU7&!20HV)F+U}Kbm16FgEt6$m
zsU;ST+gQY0xb2pS6nbWj)nb8iRTKKW7$M8DH5MhZV^X4pBCmIl>hx_w{{+Nyu#&zO
zW1ZDVf&*)=)rlrm*;oy6|2X2F_KLQL*25{LL<*CgD>!-Z#^rLcoMk{;)oSq(LBtt7
zl&zGz)#w6LHyLX_Q;G(Hm)O3wUY5H_gj+39m34$ZFGf&#9<R&@3WeH=6G!+rbx69|
z0MpmjyHUddqYH@a#Wkw3M7dhjTqF1%IOwbrCFLB}+9TOm!v0xn<dZ3i$qcKrQ_~~i
zQ3AwUwC>`&`AD;%+Y+s>eDwo~1y%qON2gr`*Ao{&9`UZbir30lmgmUx!YW^>1<A3b
zl2~Lrf{`rChXKag?31ost%1SOC{zn$HNdTnCJK%Yfq)|}x*srXTb1$76_ktlWw0P|
zRV$Q!{Y5R~Cm{fuc0lS}mc+9@&EhBaqx~&wB8#Pbg-VwzEXoop%aG@VRleFmM+!JR
zM+DtAXjG=op|tS1|HSB*w%!LA-7Zx;u2CyglunZm!NCy@6yqK1z+vu9R*W#3?qG3p
zbYUwp>@42E)e5EGXsn~dO4yW=FwlrVBKGEoKv*A{s*<uM0w_cR%%Y~=Dql<FMS(0Y
zOspcci&KP(1mlI3I#;}7V^EwYSa~jR1Xg2NSB6@y5o9hfC+fd11P<<KSW$yyw6llR
z{0ywC3#nRaXnb%mN|%G)Ml3IsW#i>Yav5vOQRElL!3+yuxoSG>i8hwrYc*Qqe&u<z
z&d-GaYZYd0m9M(3OzeA)sZCA|BOCANCt7u<nK&h?(_(P+mq?u9Jc*IcYH(FyIX|rv
z3<Y4;frC51kxtL)nX<zSrE-jBdmRlf9;yXF+w}?KBQiM094K1}cZ;a*jM9vwZjvNz
z-8;KX+@#Cpra?Jt;?WRq(2L*jjlE`YXq2zzvZZ`+=K#;90a|fEwTjWx2AnIkWgRFu
zRmX1bVYytQTrGTsDr_8a0dQ~#mY{DhcbifhJRD7Tu$rHsx<K$yVH=NRd_<-!>$XIf
zv?OviL=z`5qZMtDek4{RW?{swiAZ>44lD}0_=#r$*ruH@80Bh-#n}nwXQ!yj(t#@)
z3#o&F=a2CA*fCocrblw*T8qJv8gkc%s<s+uSS{zi#D6Yu1n1BRj&9zOx@1xc<DDxP
z%FW8Ltw$t2(pI*b`RjcwNf0OLBJUBdk+LSv^iyzL7q@XE&0;}}`gT_33RUT9Q@orn
z9*B3W-Sl~rH}_D-q}5oUlw1S2AUHxE(ghCP4UW{6XpLe#Lq49?tw_&7*tZ7|b3qKX
z9gvVu`dPMDNV;5XN4|9fMS53G8n#b!I>u2|D^%4AN@X=rjPgaGy6MEAaX-HJfqDRg
zlM2<MuJV%vEYftX9U%`~rE`HJCFl0P{>jc2953gW#DZAys;z8AV<hP2B#@v0S`m1N
zCUs;Bvs`sJ1)V~gZtN5p4CAF$MY_AS!6GtwZ9B0Fi}@L7xj<2jFxlCS^s*XaK@WwI
zG%%b(wKN??`dW<u92_6d&<7mMe6F}pOwr0lV~nOdSe_pDwjx`$AV|b3WVgm<(Q&6G
zvGpND!EPejz1fmoVRh^fAQCAm!tIL1@OlEpp1w$(kO9cLad!n8!DN>zkji@GTn#EK
z_WLZ3Tc-<IjZSc218{8i6)`i$vnyCF&gwo|J-FcZ0g?iVEMZN#TV_#u!A8imcDA#_
zV=kBNqbRLOkjv#;6xmxbXfw*~?pvW776OKuQH<)Q652FRrm6%f1(juf!;2wJ#5uEj
zqi$Pot8osc>LQloTn)5)$i8`j@oW$CqX)gjl7|D31W4E{bZUVlNHh#$jvTmAEnXt6
ziA-|{<XnnPT!^_|a?zuOt%!9o1w46y(RkwRfNV&k*186Y$_#NQT!>TfIbXjH7z0_2
z!N9=-fdkI<9Zh$!Iy<RLve$s*;Q^$>l6WP%@KaE{dJ(%8Q@q6N@*X#F8@%s!^#h8W
z%I-JPW^vzq4pN712;<4r@4_B-PpT?Zx`L+Kq9`wTID}iV4iw>Q>g#I9HTqbM6dAuf
z;E=NKXy*#f4(_gXvmP!$x;RNRn21*Xl*db2E*B!!1ZlS`2^1kL5Osqk0vKk-WVTl~
z-{v)n3)jIc3lx<>W!VKR%6UMs&T5z{zgAqMkJUgTjU-GTc{Jc))U!p3$rP*k#Iy=h
zmk1!|1PNO8ALLb{MP06_HEGb(2PiTg;+dZ|SV9iMjL~F@yeQ65u+zL(qpCD2T_RIi
zP3V>CSR`wfexF!L6W3TRQF>K=>GIcpRwD=kir#6D0vsHfPcquM;!<NS6+q4b60aLO
zizLI8%N0-p_BAEHGrR{B1Unr0@AJM}2^J?KGc)p{z<4sP8!oTMM10X|Mx|E(>P%vl
z<pEH5Jyd$6nlw$qSIY%-S-OG0sb0S4u^RHZaCRaiDiJdeNXV1GfZdGp(b&|CoFrx6
zM+cA)BuE7lr3_5#T`ubxIf=JXy1aGS0*d@w+p4t7IsAEn8g-<6G~2bEg`-$F7pGuU
zWmywt<9aEjynM3rg2Gp*Wace8EJnmtO8Rhpz!3@z!-IqSw76i#cxMmg;_MPykPsyB
zTLRS}k?lSM@*c!g2|>C#;HAY|^4$VT;Cc~UjYzAtsx$!Yt#&-!fg<Iaml16q*=XjG
zRwcBqXJOP<CG#(+=clylHClbUII>O?Ia?eVH63-S(nQ@K;Q8>tA!XmuY`5+ZAF(2j
z9v~JZb!IMkZj_9J1iU2R(w;T3+yqg#%Lj@ZQfkK{h&(QeC|98y#>|-P>>@9MMgap8
z%({G_M5RmUGIKzo{@vosWu1i#Sq9w&jM#N>M{Rh^Y9PrsTGMU$Aps+>BBR+ZR`XME
z)vg)2900Ljpmq^&N~fmS<r$HJ5pa{Z+tmS-g!|Q+m~MlC8pUXW(Rf-H<+OY%8u*K9
zM%o#PxFo17hsyJI4F(6%aIZETEpYJK65PbhkEGQ=B#qsH1CJOS&H2@>g7$EBc)t}F
zxjX>r%S#Yc35NZrStwmCw@XCB<+x#4q)UO@LevIJqg*kdvJ8{i9@bBmy&z_I{SH{I
zmUV(;mP08?=uKQ>CE^+(lj?HOn3Fssl|IrQj}QSzR~`Dp1V{6Jtx-&Ou$-TwS}i=;
z_woTG%1gpU99fSZs6;t!3@D!KCCGf(LNo~$_t}%3D}Ykac7}WpGz{6n#9BfF6x3Z=
zvZAQ#RhEmoP)H0Z-53}fT8%buc{IQXg^Tgd6`TgGf-WaOECGzAtV?SWlY@N{@5fFY
zAx{HET9p)74DZ;zf-EnLYZaL%C9*MxaJ2e}R77>k{n_yYROJfAc!I1LTWPaj$HXpR
z#9|jXCpdWH&4LHcc4n>3(AXJPbJr^9@&csG<pS1QlYqoGETaDZw&DfP{ptn_v&LvL
z!)UT2NO?+j9YD$IX@W>8RI4TC$A`5`1wy%)BP&M8CsQ$ri+g|~1`J<~N&*9Mu5nph
zs`S^%7|r%jikgwj4-f$osmdm;i6HSYe&Sh_fMc|+iX=9$;K=d<lbtJKXe2svdU)$!
zPqXM3Cr9<%aG68J_1ukeIj=`SjAzJ4V~|oVm>C(R76nGqVnnW!J6x&_zM8?n!Erv9
z6(bar9W2j|K%^cMfQTEh)|z;etoi~4tr=<@WfJwmo9^wq;i0KfV)C6d^lm5e0~8sn
zae#8QwAG99d>Z$#TCGqnAJk>RlPQYv)V43;W>j+0mkTnP*A3EHN5uJ2t)qUjZ@D;y
z*41MKke=40<#wfjg0_O?8I6F;u=(<4SFRz;M|Cie(ojlCqeEH^0CB6)p%m547D_CR
z53o8ruKn)GtWIZ<ZJJt&Rm^J09VHb5*=9b~kl=9Fp|T9cbQklZ`;Q?&lGY?_>myqW
zQQP&B0Ek!>0G<oRT4OZZLouDXPVhDlORn5LH*vFWiXPW6*{Bp#rjo3_zAjx=SezbV
zd3J(gTmvP`N4`Z#T8xN$*5yt;O4sVVle)A<F_}V_tH%@|DWxmb|H_+g>jR};(_uuJ
zJW|iBkr!i3cCWZJlGLqZ9{|*$Y@PcC8?erg4?U~F34-1JuG(%|vxzLvPOzMx+62D5
z7@HJK-7|6?8HI2eEXM`AzYr-*_O3qW0ExO?VL5FID8UV|W7}kak^x=s*wZW5f?c{F
zF;B1C^Gv@bV>K3-A041vEd<v9QVE3VndY2q=d+rhVl_WSUKpT^riRdKv?JgM=Dcjq
z0k+RK&UXTz9RLa3uBfNe0!oUf2)Wl3Kke*YN0#M2DeoexM1!pqA_WFxadLpg$q|^f
zi1dbB;~aoMteJw#&nlO5l#8=EnQyv-Vr+mC4!YU$>c4mj$uP+L>;cGHpu}7+QG#d}
zzH`A?YfSd`QA}pSwMvaz!+PNK1ya+~i)yQJcJu({a$$m3kyQ62Kj;C9h*Zl$x~^7B
zoGlkv<T*x@9Td|UD9h{bgEw8Y56f~~{DVLH0Fnesz)=9NPZg1TAK4)+lihvOktCR6
z8*^>woFtq%X*B@G{P+Nilfzao<Mm)z2M8dZ#T!d>tglCxCC-lSV{v+jVmw1J+pR}y
zGhetd^j53~4~xprUVy9vN-P3z4Ty%ya?Gw=_xcMr=7=Mx@VWOz6BG^CST<I}C{%=A
zWw+b~JWfr9T_08Fpdatz<PgiVV-%AaM$=tnqw(NCxeVa=mjfUnP%I&BcXxN57-719
z11isb)hbY^Br3QA5q*U#W{t)10p_PiiB9QV#QOq(=-?`xss$mk@s{Q3F;-_M$j1{D
zlO5!vafjRZ__$ReK#uV*+oQv(ux1|3cJtA6mT6l+lSq+Zip)HxbFB?gkj3-k`>XlU
zeWolUBGrFKPAOS(KW23Dj<vvuox6#`8;O`FyZbcW-Op)#G>9LUrSks+KK%XX*+xJr
P00000NkvXXu0mjf66+?u

diff --git a/examples/tfidf_svm_example.py b/examples/tfidf_svm_example.py
deleted file mode 100644
index aa69a129..00000000
--- a/examples/tfidf_svm_example.py
+++ /dev/null
@@ -1,63 +0,0 @@
-"""
-Example how to use tfidf with svm on simple SST dataset.
-"""
-from sklearn.metrics import accuracy_score
-
-from podium import Iterator
-from podium.datasets import SST
-from podium.experimental.models import AbstractSupervisedModel, FeatureTransformer
-from podium.experimental.models.impl import ScikitLinearSVCModel
-from podium.experimental.models.impl.simple_trainers import SimpleTrainer
-from podium.storage import LargeResource
-from podium.vectorizers import TfIdfVectorizer
-
-
-def tfidf_svm_example_main():
-    """
-    Function obtains sst dataset and then trains scikit svc linear model by
-    using tfidf as input.
-    """
-    train_set, test_set, _ = SST.get_dataset_splits()
-
-    train_iter = Iterator(batch_size=len(train_set))
-
-    tfidf_vectorizer = TfIdfVectorizer()
-    tfidf_vectorizer.fit(dataset=train_set, field=train_set.field_dict["text"])
-
-    def feature_extraction_fn(x_batch):
-        return tfidf_vectorizer.transform(x_batch.text)
-
-    def label_extraction_fn(y_batch):
-        return y_batch.label.ravel()
-
-    feature_transformer = FeatureTransformer(feature_extraction_fn)
-
-    model = ScikitLinearSVCModel()
-    trainer = SimpleTrainer()
-
-    trainer.train(
-        model=model,
-        dataset=train_set,
-        iterator=train_iter,
-        feature_transformer=feature_transformer,
-        label_transform_fun=label_extraction_fn,
-        **{trainer.MAX_EPOCH_KEY: 1},
-    )
-
-    x_batch, y_batch = train_set.batch()
-    x_train = feature_transformer.transform(x_batch)
-    y_train = label_extraction_fn(y_batch)
-    prediction_train = model.predict(X=x_train)[AbstractSupervisedModel.PREDICTION_KEY]
-    print(x_train.shape, y_train.shape, prediction_train.shape)
-    print(accuracy_score(y_true=y_train, y_pred=prediction_train))
-
-    x_batch, y_batch = test_set.batch()
-    x_test = feature_transformer.transform(x_batch)
-    y_test = label_extraction_fn(y_batch)
-    prediction_test = model.predict(X=x_test)[AbstractSupervisedModel.PREDICTION_KEY]
-    print(x_test.shape, y_test.shape, prediction_test.shape)
-    print("Accuracy:", accuracy_score(y_true=y_test, y_pred=prediction_test))
-
-
-LargeResource.BASE_RESOURCE_DIR = "downloaded_datasets"
-tfidf_svm_example_main()

From 6ff3314dbd776a13d8a2436a1c123c14b66113be Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Thu, 1 Apr 2021 20:27:35 +0200
Subject: [PATCH 20/26] Remove examples dir from commands

---
 Makefile | 14 +++++++-------
 make.bat | 28 ++++++++++++++--------------
 2 files changed, 21 insertions(+), 21 deletions(-)

diff --git a/Makefile b/Makefile
index d3dfb810..a61593b5 100644
--- a/Makefile
+++ b/Makefile
@@ -3,19 +3,19 @@
 # Check code quality
 quality:
 	@echo Checking code and doc quality.
-	black --check --line-length 90 --target-version py36 podium tests examples
-	isort --check-only podium tests examples
-	docformatter podium tests examples --check --recursive \
+	black --check --line-length 90 --target-version py36 podium tests
+	isort --check-only podium tests
+	docformatter podium tests --check --recursive \
     	--wrap-descriptions 80 --wrap-summaries 80 \
         --pre-summary-newline --make-summary-multi-line
-	flake8 podium tests examples
+	flake8 podium tests
 
 # Enforce code quality in source 
 style:
 	@echo Applying code and doc style changes.
-	black --line-length 90 --target-version py36 podium tests examples
-	isort podium tests examples
-	docformatter podium tests examples -i --recursive \
+	black --line-length 90 --target-version py36 podium tests
+	isort podium tests
+	docformatter podium tests -i --recursive \
     	--wrap-descriptions 80 --wrap-summaries 80 \
         --pre-summary-newline --make-summary-multi-line
 
diff --git a/make.bat b/make.bat
index 0ce1f11f..37d63d73 100644
--- a/make.bat
+++ b/make.bat
@@ -8,24 +8,24 @@ if "%1"!="" goto error
 
 :quality
     echo Checking code quality.
-    echo black --check --line-length 90 --target-version py36 podium tests examples
-    black --check --line-length 90 --target-version py36 podium tests examples
-    echo isort --check-only podium tests examples
-    isort --check-only podium tests examples
-    echo docformatter podium tests examples --check --recursive --wrap-descriptions 80 --wrap-summaries 80 --pre-summary-newline --make-summary-multi-line
-    docformatter podium tests examples --check --recursive --wrap-descriptions 80 --wrap-summaries 80 --pre-summary-newline --make-summary-multi-line
-    echo flake8 podium tests examples
-    flake8 podium tests examples
+    echo black --check --line-length 90 --target-version py36 podium tests
+    black --check --line-length 90 --target-version py36 podium tests
+    echo isort --check-only podium tests
+    isort --check-only podium tests
+    echo docformatter podium tests --check --recursive --wrap-descriptions 80 --wrap-summaries 80 --pre-summary-newline --make-summary-multi-line
+    docformatter podium tests --check --recursive --wrap-descriptions 80 --wrap-summaries 80 --pre-summary-newline --make-summary-multi-line
+    echo flake8 podium tests
+    flake8 podium tests
     goto :EOF
 
 :style
     echo Applying code style changes.
-    echo black --line-length 90 --target-version py36 podium tests examples
-    black --line-length 90 --target-version py36 podium tests examples
-    echo isort podium tests examples
-    isort podium tests examples
-    echo docformatter podium tests examples -i --recursive --wrap-descriptions 80 --wrap-summaries 80 --pre-summary-newline --make-summary-multi-line
-    docformatter podium tests examples -i --recursive --wrap-descriptions 80 --wrap-summaries 80 --pre-summary-newline --make-summary-multi-line
+    echo black --line-length 90 --target-version py36 podium tests
+    black --line-length 90 --target-version py36 podium tests
+    echo isort podium tests
+    isort podium tests
+    echo docformatter podium tests -i --recursive --wrap-descriptions 80 --wrap-summaries 80 --pre-summary-newline --make-summary-multi-line
+    docformatter podium tests -i --recursive --wrap-descriptions 80 --wrap-summaries 80 --pre-summary-newline --make-summary-multi-line
     goto :EOF
 
 :test

From 715859db2a1fb858f1f05f5ce7e5b7d1f755c06a Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Thu, 1 Apr 2021 20:30:00 +0200
Subject: [PATCH 21/26] Remove examples dir from action

---
 .github/workflows/ci.yml | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index a12bc935..8121a045 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -38,18 +38,18 @@ jobs:
         pip install .[quality]
     - name: Check black compliance
       run: |
-        black --check --line-length 90 --target-version py36 podium tests examples
+        black --check --line-length 90 --target-version py36 podium tests
     - name: Check isort compliance
       run: |
-        isort --check-only podium tests examples
+        isort --check-only podium tests
     - name: Check docformatter compliance
       run: |
-        docformatter podium tests examples --check --recursive \
+        docformatter podium tests --check --recursive \
           --wrap-descriptions 80 --wrap-summaries 80 \
           --pre-summary-newline --make-summary-multi-line
     - name: Check flake8 compliance
       run: |
-        flake8 podium tests examples
+        flake8 podium tests
   
   build_and_test:
     runs-on: ${{ matrix.os }}

From cf26d20c334542864c579634ba720cef1f8f3b4d Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Thu, 1 Apr 2021 21:04:04 +0200
Subject: [PATCH 22/26] Update readme outputs

---
 README.md | 25 ++++++++++++++++---------
 1 file changed, 16 insertions(+), 9 deletions(-)

diff --git a/README.md b/README.md
index 41019095..c6cfcc22 100644
--- a/README.md
+++ b/README.md
@@ -56,13 +56,13 @@ SST({
             name: text,
             keep_raw: False,
             is_target: False,
-            vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, finalized: True, size: 16284})
+            vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 16284})
         }),
         LabelField({
             name: label,
             keep_raw: False,
             is_target: True,
-            vocab: Vocab({specials: (), eager: False, finalized: True, size: 2})
+            vocab: Vocab({specials: (), eager: False, is_finalized: True, size: 2})
         })
     ]
 })
@@ -94,7 +94,7 @@ HFDatasetConverter({
             name: 'text',
             keep_raw: False,
             is_target: False,
-            vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, finalized: True, size: 280619})
+            vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 280619})
         }),
         LabelField({
             name: 'label',
@@ -105,7 +105,7 @@ HFDatasetConverter({
 })
 ```
 
-Load your own dataset from a standardized tabular format (e.g. `csv`, `tsv`, `jsonl`):
+Load your own dataset from a standardized tabular format (e.g. `csv`, `tsv`, `jsonl`, ...):
 
 ```python
 >>> from podium.datasets import TabularDataset
@@ -121,24 +121,27 @@ TabularDataset({
     fields: [
         Field({
             name: 'premise',
+            keep_raw: False,
             is_target: False, 
-            vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, finalized: True, size: 19})
+            vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 15})
         }),
         Field({
             name: 'hypothesis',
+            keep_raw: False,
             is_target: False, 
-            vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, finalized: True, size: 19})
+            vocab: Vocab({specials: ('<UNK>', '<PAD>'), eager: False, is_finalized: True, size: 6})
         }),
         LabelField({
             name: 'label',
+            keep_raw: False,
             is_target: True, 
-            vocab: Vocab({specials: (), eager: False, finalized: True, size: 1})
+            vocab: Vocab({specials: (), eager: False, is_finalized: True, size: 1})
         })
     ]
 })
 ```
 
-Or define your own `Dataset` subclass (tutorial coming soon).
+Also check our documentation to see how you can load a dataset from [Pandas](https://pandas.pydata.org/), the CoNLL format, or define your own `Dataset` subclass (tutorial coming soon).
 
 ### Define your preprocessing
 
@@ -151,6 +154,7 @@ We wrap dataset pre-processing in customizable `Field` classes. Each `Field` has
 >>> label = LabelField(name='label')
 >>> fields = {'text': text, 'label': label}
 >>> sst_train, sst_dev, sst_test = SST.get_dataset_splits(fields=fields)
+>>> sst_train.finalize_fields()
 >>> print(vocab)
 Vocab({specials: ('<UNK>', '<PAD>'), eager: True, finalized: True, size: 5000})
 ```
@@ -175,6 +179,7 @@ You could decide to lowercase all the characters and filter out all non-alphanum
 >>> text.add_posttokenize_hook(filter_alnum)
 >>> fields = {'text': text, 'label': label}
 >>> sst_train, sst_dev, sst_test = SST.get_dataset_splits(fields=fields)
+>>> sst_train.finalize_fields()
 >>> print(sst_train[222])
 Example({
     text: (None, ['a', 'slick', 'engrossing', 'melodrama']),
@@ -201,9 +206,11 @@ A common use-case is to incorporate existing components of pretrained language m
 ...                       numericalizer=tokenizer.convert_tokens_to_ids)
 >>> fields = {'text': subword_field, 'label': label}
 >>> sst_train, sst_dev, sst_test = SST.get_dataset_splits(fields=fields)
+>>> # No need to finalize since we're not using a vocab!
 >>> print(sst_train[222])
 Example({
-    subword: (None, ['a', 'slick', ',', 'eng', '##ross', '##ing', 'mel', '##od', '##rama', '.']),label: (None, 'positive')
+    subword: (None, ['a', 'slick', ',', 'eng', '##ross', '##ing', 'mel', '##od', '##rama', '.']),
+    label: (None, 'positive')
 })
 ```
 

From d5a37e7d215026861681bc537ca88b34cee9980b Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Fri, 2 Apr 2021 11:29:51 +0200
Subject: [PATCH 23/26] Add roadmap

---
 README.md  |  6 +++---
 Roadmap.md | 46 ++++++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 49 insertions(+), 3 deletions(-)
 create mode 100644 Roadmap.md

diff --git a/README.md b/README.md
index c6cfcc22..a3c85505 100644
--- a/README.md
+++ b/README.md
@@ -11,7 +11,7 @@ The main source of inspiration for Podium is an old version of [torchtext](https
 ### Contents
 
 - [Installation](#installation)
-- [Usage examples](#usage-examples)
+- [Usage examples](#usage)
 - [Contributing](#contributing)
 - [Versioning](#versioning)
 - [Authors](#authors)
@@ -216,11 +216,11 @@ Example({
 
 For a more interactive introduction, check out the quickstart on Google Colab: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/takelab/podium/blob/master/docs/source/notebooks/quickstart.ipynb)
 
-More complex examples can be found in our [examples folder](./examples).
+Full usage examples can be found in our [docs](https://takelab.fer.hr/podium/examples).
 
 ## Contributing
 
-We welcome contributions! To learn more about making a contribution to Podium, please see our [Contribution page](CONTRIBUTING.md).
+We welcome contributions! To learn more about making a contribution to Podium, please see our [Contribution page](CONTRIBUTING.md) and our [Roadmap](ROADMAP.md).
 
 ## Versioning
 
diff --git a/Roadmap.md b/Roadmap.md
new file mode 100644
index 00000000..27adff82
--- /dev/null
+++ b/Roadmap.md
@@ -0,0 +1,46 @@
+# Roadmap
+
+If you are interested in contributing to Podium, some of our 
+Order does not reflect importance.
+
+## Major changes
+
+- Dynamic application of Fields
+  - Right now, for every change in Fields the dataset needs to be reloaded. The goal of this change would be to allow users to replace or update a Field in a Dataset. The Dataset should be aware of this change (e.g. by keeping a hash of the Field object) and if it happens, recompute all the necessary data for that Field.
+- Parallelization
+  - For data preprocessing (apply Fields in parallel)
+  - For data loading
+- Conditional processing in Fields
+  - Handle cases where the values computed in one Field are dependent on values computed in another Field
+- Experimental pipeline
+  - `podium.experimental`, wrappers for model framework agnostic training & serving
+  - Low priority 
+
+## Minor changes
+
+- Populate hooks & preprocessing utilities
+  - Lowercase, truncate, extract POS, ...
+- Improve Dataset coverage
+  - Data wrappers / abstract loaders for other source libraries and input formats
+- BucketIterator modifications
+  - Simplify setting the sort key (e.g., in the basic case  where the batch should be sorted according to the length of a single Field, accept a Field name as the argument)
+- Improve HF/datasets integration
+  - Better automatic Field inference from features
+  - Cover additional feature datatypes (e.g., image data)
+- Centralized and intuitive download script
+  - Low priority as most data loading is delegated to hf/datasets
+- Add a Mask token for MLM (can be handled with posttokenization hooks right now, but not ideal)
+- Populate pretrained vectors
+  - Word2vec
+
+## Documentation
+
+- Examples
+  - Language modeling
+  - Tensorflow model
+  - Various task types
+- Chapters
+  - Handling datasets with missing tokens
+  - Loading data from pandas / porting data to pandas
+  - Loading CoNLL datasets
+  - Implementing your own dataset subclass

From 7436d30da335de7e49d478155a1337fbf41536a9 Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Fri, 2 Apr 2021 11:31:02 +0200
Subject: [PATCH 24/26] Add roadmap

---
 README.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/README.md b/README.md
index a3c85505..adf48c20 100644
--- a/README.md
+++ b/README.md
@@ -220,7 +220,7 @@ Full usage examples can be found in our [docs](https://takelab.fer.hr/podium/exa
 
 ## Contributing
 
-We welcome contributions! To learn more about making a contribution to Podium, please see our [Contribution page](CONTRIBUTING.md) and our [Roadmap](ROADMAP.md).
+We welcome contributions! To learn more about making a contribution to Podium, please see our [Contribution page](CONTRIBUTING.md) and our [Roadmap](Roadmap.md).
 
 ## Versioning
 

From a3cd42eceab46efe67d6443b3d2a55cfa51b42e2 Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Fri, 2 Apr 2021 11:40:50 +0200
Subject: [PATCH 25/26] Add roadmap

---
 Roadmap.md | 26 ++++++++++++++++++++++----
 1 file changed, 22 insertions(+), 4 deletions(-)

diff --git a/Roadmap.md b/Roadmap.md
index 27adff82..5fe9bb4d 100644
--- a/Roadmap.md
+++ b/Roadmap.md
@@ -1,17 +1,34 @@
 # Roadmap
 
-If you are interested in contributing to Podium, some of our 
+If you are interested in making a contribution to Podium, this page outlines some changes we are planning to focus on in the near future. Feel free to propose improvements and moficiations either via [discussions](https://github.com/TakeLab/podium/discussions) or by raising an [issue](https://github.com/TakeLab/podium/issues).
+
 Order does not reflect importance.
 
 ## Major changes
 
 - Dynamic application of Fields
   - Right now, for every change in Fields the dataset needs to be reloaded. The goal of this change would be to allow users to replace or update a Field in a Dataset. The Dataset should be aware of this change (e.g. by keeping a hash of the Field object) and if it happens, recompute all the necessary data for that Field.
+
+  The current pattern is:
+  ```python
+    # Load a dataset
+    fields = {'text':text, 'label':label}
+    dataset = load_dataset(fields=fields)
+
+    # Decide to change something with one of the Fields
+    text = Field(..., tokenizer=some_different_tokenizer)
+    # Potentially expensive dataset loading is required again
+    dataset = load_dataset(fields=fields)
+  ```
+  Dataset instances should instead detect changes in a Field and recompute values (Vocabs) for the ones that changed.
+
 - Parallelization
   - For data preprocessing (apply Fields in parallel)
   - For data loading
+
 - Conditional processing in Fields
-  - Handle cases where the values computed in one Field are dependent on values computed in another Field
+  - Handle cases where the values computed in one Field are dependent on values computed in another Field 
+
 - Experimental pipeline
   - `podium.experimental`, wrappers for model framework agnostic training & serving
   - Low priority 
@@ -20,6 +37,8 @@ Order does not reflect importance.
 
 - Populate hooks & preprocessing utilities
   - Lowercase, truncate, extract POS, ...
+- Populate pretrained vectors
+  - Word2vec
 - Improve Dataset coverage
   - Data wrappers / abstract loaders for other source libraries and input formats
 - BucketIterator modifications
@@ -27,11 +46,10 @@ Order does not reflect importance.
 - Improve HF/datasets integration
   - Better automatic Field inference from features
   - Cover additional feature datatypes (e.g., image data)
+  - Cleaner API?
 - Centralized and intuitive download script
   - Low priority as most data loading is delegated to hf/datasets
 - Add a Mask token for MLM (can be handled with posttokenization hooks right now, but not ideal)
-- Populate pretrained vectors
-  - Word2vec
 
 ## Documentation
 

From 10bf1cfb0968b18b3f13765f28a4834e9efb5efc Mon Sep 17 00:00:00 2001
From: Martin Tutek <martin.tutek@gmail.com>
Date: Fri, 2 Apr 2021 13:01:24 +0200
Subject: [PATCH 26/26] Polish, comments, rename BasicVectorStorage to
 WordVectors

---
 Roadmap.md                           |  1 +
 podium/datasets/dataset.py           |  2 +-
 podium/datasets/iterator.py          |  2 +-
 podium/field.py                      |  2 +-
 podium/vectorizers/__init__.py       |  2 +-
 podium/vectorizers/impl/glove.py     |  4 +--
 podium/vectorizers/impl/nlpl.py      |  4 +--
 podium/vectorizers/vectorizer.py     |  2 +-
 tests/vectorizers/test_vectorizer.py | 50 ++++++++++++----------------
 9 files changed, 31 insertions(+), 38 deletions(-)

diff --git a/Roadmap.md b/Roadmap.md
index 5fe9bb4d..caba2749 100644
--- a/Roadmap.md
+++ b/Roadmap.md
@@ -39,6 +39,7 @@ Order does not reflect importance.
   - Lowercase, truncate, extract POS, ...
 - Populate pretrained vectors
   - Word2vec
+  - Interface with e.g. gensim
 - Improve Dataset coverage
   - Data wrappers / abstract loaders for other source libraries and input formats
 - BucketIterator modifications
diff --git a/podium/datasets/dataset.py b/podium/datasets/dataset.py
index f26b3a6e..ab681adf 100644
--- a/podium/datasets/dataset.py
+++ b/podium/datasets/dataset.py
@@ -278,7 +278,7 @@ def shuffled(self) -> "DatasetBase":
         return self[shuffled_indices]
 
     def __repr__(self):
-        fields_str = ",\n".join(textwrap.indent(repr(f), " " * 8) for f in self.fields)
+        fields_str = ",\n".join(textwrap.indent(repr(f), " " * 4) for f in self.fields)
         fields_str = f"[\n{fields_str}\n    \n]"
         attrs = {"size": len(self), "fields": fields_str}
         return repr_type_and_attrs(self, attrs, with_newlines=True, repr_values=False)
diff --git a/podium/datasets/iterator.py b/podium/datasets/iterator.py
index bd8781aa..794df5a0 100644
--- a/podium/datasets/iterator.py
+++ b/podium/datasets/iterator.py
@@ -494,7 +494,7 @@ def __init__(self, dataset: DatasetBase = None, shuffle=True, add_padding=True):
             returned as python lists of ``matrix_class`` instances.
         """
 
-        batch_size = len(dataset) if dataset else 0
+        batch_size = len(dataset) if dataset else None
 
         super().__init__(
             dataset=dataset,
diff --git a/podium/field.py b/podium/field.py
index cd946577..fc554412 100644
--- a/podium/field.py
+++ b/podium/field.py
@@ -954,7 +954,7 @@ def remove_pretokenize_hooks(self):
 
     def __repr__(self):
         fields_str = ",\n".join(
-            textwrap.indent(repr(f), " " * 8) for f in self._output_fields
+            textwrap.indent(repr(f), " " * 4) for f in self._output_fields
         )
         fields_str = f"[\n{fields_str}\n    \n]"
         attrs = {"fields": fields_str}
diff --git a/podium/vectorizers/__init__.py b/podium/vectorizers/__init__.py
index 3342dc08..60807585 100644
--- a/podium/vectorizers/__init__.py
+++ b/podium/vectorizers/__init__.py
@@ -5,8 +5,8 @@
 from .impl import GloVe, NlplVectorizer
 from .tfidf import TfIdfVectorizer
 from .vectorizer import (
-    BasicVectorStorage,
     VectorStorage,
+    WordVectors,
     random_normal_default_vector,
     zeros_default_vector,
 )
diff --git a/podium/vectorizers/impl/glove.py b/podium/vectorizers/impl/glove.py
index ddd3b2cc..dd676d2d 100644
--- a/podium/vectorizers/impl/glove.py
+++ b/podium/vectorizers/impl/glove.py
@@ -1,10 +1,10 @@
 import os
 
 from podium.storage import LargeResource
-from podium.vectorizers.vectorizer import BasicVectorStorage, random_normal_default_vector
+from podium.vectorizers.vectorizer import WordVectors, random_normal_default_vector
 
 
-class GloVe(BasicVectorStorage):
+class GloVe(WordVectors):
     """
     Class represents concrete vector storage for GloVe vectors described in
     https://nlp.stanford.edu/projects/glove/ . Class contains a Large resource
diff --git a/podium/vectorizers/impl/nlpl.py b/podium/vectorizers/impl/nlpl.py
index 77e96c8a..e8398721 100644
--- a/podium/vectorizers/impl/nlpl.py
+++ b/podium/vectorizers/impl/nlpl.py
@@ -1,10 +1,10 @@
 import os
 
 from podium.storage import LargeResource
-from podium.vectorizers.vectorizer import BasicVectorStorage, zeros_default_vector
+from podium.vectorizers.vectorizer import WordVectors, zeros_default_vector
 
 
-class NlplVectorizer(BasicVectorStorage):
+class NlplVectorizer(WordVectors):
     NAME = "nlpl_vectors"
     URL = "http://vectors.nlpl.eu/repository/11/36.zip"
     ARCHIVE_TYPE = "zip"
diff --git a/podium/vectorizers/vectorizer.py b/podium/vectorizers/vectorizer.py
index 2f7b1e1a..42946568 100644
--- a/podium/vectorizers/vectorizer.py
+++ b/podium/vectorizers/vectorizer.py
@@ -227,7 +227,7 @@ def __repr__(self):
         return repr_type_and_attrs(self, attrs)
 
 
-class BasicVectorStorage(VectorStorage):
+class WordVectors(VectorStorage):
     """
     Basic implementation of VectorStorage that handles loading vectors from
     system storage.
diff --git a/tests/vectorizers/test_vectorizer.py b/tests/vectorizers/test_vectorizer.py
index 38454b65..e005d60d 100644
--- a/tests/vectorizers/test_vectorizer.py
+++ b/tests/vectorizers/test_vectorizer.py
@@ -48,7 +48,7 @@ def test_basic_not_initialized():
     with create_temp_vect_file(
         vect_file_name="vect1", file_data=BASIC_VECT_DATA
     ) as vect_file_path:
-        vect = vectorizer.BasicVectorStorage(path=vect_file_path)
+        vect = vectorizer.WordVectors(path=vect_file_path)
         with pytest.raises(RuntimeError):
             vect["."]
         with pytest.raises(RuntimeError):
@@ -59,7 +59,7 @@ def test_basic_load_all_vectors():
     with create_temp_vect_file(
         vect_file_name="vect1", file_data=BASIC_VECT_DATA
     ) as vect_file_path:
-        vect = vectorizer.BasicVectorStorage(path=vect_file_path)
+        vect = vectorizer.WordVectors(path=vect_file_path)
         vect.load_all()
         assert len(vect._vectors) == 4
         assert vect["."].shape == (3,)
@@ -74,7 +74,7 @@ def test_get_vector_dimension():
     with create_temp_vect_file(
         vect_file_name="vect1", file_data=BASIC_VECT_DATA
     ) as vect_file_path:
-        vect = vectorizer.BasicVectorStorage(path=vect_file_path)
+        vect = vectorizer.WordVectors(path=vect_file_path)
         vect.load_all()
         assert vect.get_vector_dim() == vect["."].shape[0]
         assert vect.get_vector_dim() == 3
@@ -84,7 +84,7 @@ def test_get_vector_dim_not_initialized_vector_storage():
     with create_temp_vect_file(
         vect_file_name="vect1", file_data=BASIC_VECT_DATA
     ) as vect_file_path:
-        vect = vectorizer.BasicVectorStorage(path=vect_file_path)
+        vect = vectorizer.WordVectors(path=vect_file_path)
         with pytest.raises(RuntimeError):
             vect.get_vector_dim()
 
@@ -93,7 +93,7 @@ def test_basic_load_with_header():
     with create_temp_vect_file(
         vect_file_name="vect1", file_header=BASIC_VECT_HEADING, file_data=BASIC_VECT_DATA
     ) as vect_file_path:
-        vect = vectorizer.BasicVectorStorage(path=vect_file_path)
+        vect = vectorizer.WordVectors(path=vect_file_path)
         vect.load_all()
         assert len(vect._vectors) == 4
         assert vect["."].shape == (3,)
@@ -108,9 +108,7 @@ def test_basic_no_token():
     with create_temp_vect_file(
         vect_file_name="vect1", file_data=BASIC_VECT_DATA
     ) as vect_file_path:
-        vect = vectorizer.BasicVectorStorage(
-            path=vect_file_path, default_vector_function=None
-        )
+        vect = vectorizer.WordVectors(path=vect_file_path, default_vector_function=None)
         vect.load_all()
         with pytest.raises(KeyError):
             print(vect["a"])
@@ -122,9 +120,7 @@ def test_basic_token_none():
     with create_temp_vect_file(
         vect_file_name="vect1", file_data=BASIC_VECT_DATA
     ) as vect_file_path:
-        vect = vectorizer.BasicVectorStorage(
-            path=vect_file_path, default_vector_function=None
-        )
+        vect = vectorizer.WordVectors(path=vect_file_path, default_vector_function=None)
         vect.load_all()
         with pytest.raises(ValueError):
             vect[None]
@@ -136,7 +132,7 @@ def test_basic_token_default():
     with create_temp_vect_file(
         vect_file_name="vect1", file_data=BASIC_VECT_DATA
     ) as vect_file_path:
-        vect = vectorizer.BasicVectorStorage(
+        vect = vectorizer.WordVectors(
             path=vect_file_path, default_vector_function=vectorizer.zeros_default_vector
         )
         vect.load_all()
@@ -151,9 +147,7 @@ def test_basic_load_vocab():
     with create_temp_vect_file(
         vect_file_name="vect1", file_data=BASIC_VECT_DATA
     ) as vect_file_path:
-        vect = vectorizer.BasicVectorStorage(
-            path=vect_file_path, default_vector_function=None
-        )
+        vect = vectorizer.WordVectors(path=vect_file_path, default_vector_function=None)
         vocab = [".", ":"]
         vect.load_vocab(vocab=vocab)
         assert len(vect._vectors) == 2
@@ -174,7 +168,7 @@ def test_basic_load_vocab_none():
     with create_temp_vect_file(
         vect_file_name="vect1", file_data=BASIC_VECT_DATA
     ) as vect_file_path:
-        vect = vectorizer.BasicVectorStorage(path=vect_file_path)
+        vect = vectorizer.WordVectors(path=vect_file_path)
         with pytest.raises(ValueError):
             vect.load_vocab(vocab=None)
 
@@ -201,7 +195,7 @@ def test_get_embedding_matrix(tokens, expected_matrix, expected_shape):
     with create_temp_vect_file(
         vect_file_name="vect1", file_data=BASIC_VECT_DATA
     ) as vect_file_path:
-        vect = vectorizer.BasicVectorStorage(path=vect_file_path)
+        vect = vectorizer.WordVectors(path=vect_file_path)
         vect.load_all()
 
         embedding_matrix = vect.get_embedding_matrix(vocab=tokens)
@@ -213,7 +207,7 @@ def test_basic_diff_dimensions():
     with create_temp_vect_file(
         vect_file_name="vect1", file_data=DIFF_DIM_VECT_DATA
     ) as vect_file_path:
-        vect = vectorizer.BasicVectorStorage(path=vect_file_path)
+        vect = vectorizer.WordVectors(path=vect_file_path)
         with pytest.raises(RuntimeError):
             vect.load_all()
 
@@ -236,7 +230,7 @@ def test_basic_max_vectors_less_than_num_lines():
     with create_temp_vect_file(
         vect_file_name="vect1", file_data=BASIC_VECT_DATA
     ) as vect_file_path:
-        vect = vectorizer.BasicVectorStorage(path=vect_file_path, max_vectors=2)
+        vect = vectorizer.WordVectors(path=vect_file_path, max_vectors=2)
         vect.load_all()
         assert len(vect._vectors) == 2
         contained_elements = [".", "'"]
@@ -249,7 +243,7 @@ def test_basic_max_vectors_vocab():
     with create_temp_vect_file(
         vect_file_name="vect1", file_data=BASIC_VECT_DATA
     ) as vect_file_path:
-        vect = vectorizer.BasicVectorStorage(path=vect_file_path, max_vectors=2)
+        vect = vectorizer.WordVectors(path=vect_file_path, max_vectors=2)
         vocab = [".", ":", ","]
         vect.load_vocab(vocab)
         assert len(vect._vectors) == 2
@@ -263,7 +257,7 @@ def test_basic_max_vectors_bigger_than_num_lines():
     with create_temp_vect_file(
         vect_file_name="vect1", file_data=BASIC_VECT_DATA
     ) as vect_file_path:
-        vect = vectorizer.BasicVectorStorage(path=vect_file_path, max_vectors=20)
+        vect = vectorizer.WordVectors(path=vect_file_path, max_vectors=20)
         vect.load_all()
         assert len(vect._vectors) == 4
         contained_elements = [".", "'", ":", ","]
@@ -271,7 +265,7 @@ def test_basic_max_vectors_bigger_than_num_lines():
 
 
 def test_basic_both_paths_none():
-    vect = vectorizer.BasicVectorStorage(path=None, cache_path=None)
+    vect = vectorizer.WordVectors(path=None, cache_path=None)
     with pytest.raises(ValueError):
         vect.load_all()
 
@@ -284,7 +278,7 @@ def test_basic_both_paths_doesnt_exist(tmpdir):
     cache_path = os.path.join(base, "cache.t")
     assert not os.path.exists(cache_path)
 
-    vect = vectorizer.BasicVectorStorage(path=file_path, cache_path=cache_path)
+    vect = vectorizer.WordVectors(path=file_path, cache_path=cache_path)
     with pytest.raises(ValueError):
         vect.load_all()
 
@@ -295,7 +289,7 @@ def test_basic_path_none_cache_doesnt_exist(tmpdir):
     cache_path = os.path.join(base, "cache.t")
     assert not os.path.exists(cache_path)
 
-    vect = vectorizer.BasicVectorStorage(path=None, cache_path=cache_path)
+    vect = vectorizer.WordVectors(path=None, cache_path=cache_path)
     with pytest.raises(ValueError):
         vect.load_all()
 
@@ -307,7 +301,7 @@ def test_basic_cache_max_vectors(tmpdir):
         assert os.path.exists(vect_file_path)
         cache_path = os.path.join(tmpdir, "cache.t")
         assert not os.path.exists(cache_path)
-        vect = vectorizer.BasicVectorStorage(
+        vect = vectorizer.WordVectors(
             path=vect_file_path, max_vectors=2, cache_path=cache_path
         )
         vect.load_all()
@@ -329,9 +323,7 @@ def test_basic_cache_vocab():
             assert os.path.exists(vect_file_path)
             cache_path = os.path.join(base, "cache.t")
             assert not os.path.exists(cache_path)
-            vect = vectorizer.BasicVectorStorage(
-                path=vect_file_path, cache_path=cache_path
-            )
+            vect = vectorizer.WordVectors(path=vect_file_path, cache_path=cache_path)
 
             vocab = [".", ":", ","]
             vect.load_vocab(vocab)
@@ -351,7 +343,7 @@ def test_load_plain_text():
             assert os.path.exists(file_path)
             file.writelines(BASIC_VECT_DATA_PLAIN)
 
-        vec_storage = vectorizer.BasicVectorStorage(file_path, binary=False)
+        vec_storage = vectorizer.WordVectors(file_path, binary=False)
         vec_storage.load_all()
 
     assert len(vec_storage) == 4