domain: add core classifier

Signed-off-by: Salman Maqbool [email protected]
Sem-ver: new feature

.travis.yml

@@ -6,7 +6,7 @@ before_install:
 - travis_retry pip install --upgrade pip setuptools
 - travis_retry pip install coveralls
 install:
-- travis_retry pip install -e .[tests]
+- travis_retry pip install --progress-bar off --process-dependency-links -e .[tests]
 script:
 - "./run-tests.sh"
 deploy:

inspire_classifier/__init__.py

@@ -21,5 +21,3 @@
 # or submit itself to any jurisdiction.
 
 """INSPIRE module aimed at automatically classifying the new papers that are added to INSPIRE, such as if they are core or not, or the arXiv category corresponding to each of them."""
-
-from __future__ import absolute_import, division, print_function

inspire_classifier/api.py

@@ -0,0 +1,207 @@
+# -*- coding: utf-8 -*-
+#
+# This file is part of INSPIRE.
+# Copyright (C) 2014-2018 CERN.
+#
+# INSPIRE is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# INSPIRE is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with INSPIRE. If not, see <http://www.gnu.org/licenses/>.
+#
+# In applying this license, CERN does not waive the privileges and immunities
+# granted to it by virtue of its status as an Intergovernmental Organization
+# or submit itself to any jurisdiction.
+#
+# Modified from the fastai library (https://github.com/fastai/fastai).
+
+"""Classifier API."""
+
+
+from flask import current_app
+from inspire_classifier.domain.models import (
+    Classifier,
+    LanguageModel
+)
+from inspire_classifier.domain.preprocessor import (
+    generate_and_save_classifier_tokens,
+    generate_and_save_language_model_tokens,
+    map_and_save_tokens_to_ids_for_classifier,
+    map_and_save_tokens_to_ids_for_language_model,
+    split_and_save_data_for_language_model_and_classifier
+)
+from inspire_classifier.utils import path_for
+import numpy as np
+import requests
+
+
+def create_directories():
+    """Create the project data and model directories"""
+    path_for('classifier_data').mkdir(parents=True, exist_ok=True)
+    path_for('language_model_data').mkdir(parents=True, exist_ok=True)
+    path_for('classifier_model').mkdir(parents=True, exist_ok=True)
+    (path_for('language_model') / 'wikitext_103').mkdir(exist_ok=True, parents=True)
+
+
+def preprocess_and_save_data():
+    """
+    Prepares the data for training.
+    """
+    try:
+        split_and_save_data_for_language_model_and_classifier(
+            dataframe_path=path_for('dataframe'), language_model_data_dir=path_for('language_model_data'),
+            classifier_data_dir=path_for('classifier_data'),
+            val_fraction=current_app.config['CLASSIFIER_VALIDATION_DATA_FRACTION']
+        )
+    except IOError as error:
+        raise IOError('Training dataframe not found.') from error
+
+    try:
+        generate_and_save_language_model_tokens(language_model_data_dir=path_for('language_model_data'))
+    except IOError as error:
+        raise IOError('Language Model data directory does not exist.') from error
+
+    try:
+        map_and_save_tokens_to_ids_for_language_model(
+            language_model_data_dir=path_for('language_model_data'), data_itos_path=path_for('data_itos'),
+            max_vocab_size=current_app.config['CLASSIFIER_MAXIMUM_VOCABULARY_SIZE'],
+            minimum_frequency=current_app.config['CLASSIFIER_MINIMUM_WORD_FREQUENCY']
+        )
+    except IOError as error:
+        raise IOError('Language Model data directory or the data directory do not exist.') from error
+
+    try:
+        generate_and_save_classifier_tokens(classifier_data_dir=path_for('classifier_data'))
+    except IOError as error:
+        raise IOError('Classifier data directory does not exist.') from error
+
+    try:
+        map_and_save_tokens_to_ids_for_classifier(classifier_data_dir=path_for('classifier_data'),
+                                                  data_itos_path=path_for('data_itos'))
+    except IOError as error:
+        raise IOError('Classifier data directory or the data ITOS does not exist.') from error
+
+
+def finetune_and_save_language_model():
+    """
+    Finetunes the pretrained (on wikitext103) language model on our dataset.
+    """
+    try:
+        language_model = LanguageModel(
+            training_data_ids_path=path_for('language_model_data') / 'training_token_ids.npy',
+            validation_data_ids_path=path_for('language_model_data') / 'validation_token_ids.npy',
+            language_model_model_dir=path_for('language_model_data'),
+            data_itos_path=path_for('data_itos'), cuda_device_id=current_app.config['CLASSIFIER_CUDA_DEVICE_ID'],
+            batch_size=current_app.config['CLASSIFIER_LANGUAGE_MODEL_BATCH_SIZE']
+        )
+    except IOError as error:
+        raise IOError('Training files, language model data directory, or data ITOS do not exist.') from error
+
+    if not path_for('pretrained_language_model').exists():
+        wikitext103_language_model_response = requests.get(
+            current_app.config['CLASSIFIER_WIKITEXT103_LANGUAGE_MODEL_URL'], allow_redirects=True)
+        wikitext103_language_model_response.raise_for_status()
+        with open(path_for('pretrained_language_model'), 'wb') as fd:
+            fd.write(wikitext103_language_model_response.content)
+    if not path_for('wikitext103_itos').exists():
+        wikitext103_itos_response = requests.get(current_app.config['CLASSIFIER_WIKITEXT103_ITOS_URL'],
+                                                 allow_redirects=True)
+        wikitext103_itos_response.raise_for_status()
+        with open(path_for('wikitext103_itos'), 'wb') as fd:
+            fd.write(wikitext103_itos_response.content)
+
+    try:
+        language_model.load_pretrained_language_model_weights(
+            pretrained_language_model_path=path_for('pretrained_language_model'),
+            wikitext103_itos_path=path_for('wikitext103_itos')
+        )
+    except IOError as error:
+        raise IOError('Wikitext103 pretrained language model and Wikitext103 ITOS do not exist.') from error
+
+    try:
+        language_model.train(finetuned_language_model_encoder_save_path=path_for('finetuned_language_model_encoder'),
+                             cycle_length=current_app.config['CLASSIFIER_LANGUAGE_MODEL_CYCLE_LENGTH'])
+    except IOError as error:
+        raise IOError('Unable to save the finetuned language model.') from error
+
+
+def train_and_save_classifier():
+    """
+    Trains the classifier on our dataset and save the weights.
+    """
+    try:
+        classifier = Classifier(data_itos_path=path_for('data_itos'),
+                                number_of_classes=3, cuda_device_id=current_app.config['CLASSIFIER_CUDA_DEVICE_ID'])
+    except IOError as error:
+        raise IOError('Data ITOS not found.') from error
+
+    try:
+        classifier.load_training_and_validation_data(
+            training_data_ids_path=path_for('classifier_data') / 'training_token_ids.npy',
+            training_data_labels_path=path_for('classifier_data') / 'training_labels.npy',
+            validation_data_ids_path=path_for('classifier_data') / 'validation_token_ids.npy',
+            validation_data_labels_path=path_for('classifier_data') / 'validation_labels.npy',
+            classifier_data_dir=path_for('classifier_data'),
+            batch_size=current_app.config['CLASSIFIER_CLASSIFIER_BATCH_SIZE']
+        )
+    except IOError as error:
+        raise IOError('Training and Validation data for Classifier not found.') from error
+
+    classifier.initialize_learner()
+
+    try:
+        classifier.load_finetuned_language_model_weights(
+            finetuned_language_model_encoder_path=path_for('finetuned_language_model_encoder')
+        )
+    except IOError as error:
+        raise IOError('Finetuned Language Model Encoder does not exist.') from error
+
+    try:
+        classifier.train(trained_classifier_save_path=path_for('trained_classifier'),
+                         cycle_length=current_app.config['CLASSIFIER_CLASSIFIER_CYCLE_LENGTH'])
+    except IOError as error:
+        raise IOError('Unable to save the trained classifier.') from error
+
+
+def train():
+    """
+    Runs the complete training pipeline.
+    """
+    create_directories()
+    preprocess_and_save_data()
+    finetune_and_save_language_model()
+    train_and_save_classifier()
+
+
+def predict_coreness(title, abstract):
+    """
+    Predicts class-wise probabilities given the title and abstract.
+    """
+    text = title + abstract
+    categories = ['rejected', 'non_core', 'core']
+    try:
+        classifier = Classifier(data_itos_path=path_for('data_itos'),
+                                number_of_classes=3, cuda_device_id=current_app.config['CLASSIFIER_CUDA_DEVICE_ID'])
+    except IOError as error:
+        raise IOError('Data ITOS not found.') from error
+
+    try:
+        classifier.load_trained_classifier_weights(path_for('trained_classifier'))
+    except IOError as error:
+        raise IOError('Could not load the trained classifier weights.') from error
+
+    class_probabilities = classifier.predict(text)
+    assert len(class_probabilities) == 3
+
+    predicted_class = categories[np.argmax(class_probabilities)]
+    output_dict = {'prediction': predicted_class}
+    output_dict['scores'] = dict(zip(categories, class_probabilities))
+
+    return output_dict

inspire_classifier/app.py

@@ -20,16 +20,14 @@
 # granted to it by virtue of its status as an Intergovernmental Organization
 # or submit itself to any jurisdiction.
 
-from __future__ import absolute_import, division, print_function
-
 import datetime
 
 from flask import Flask, jsonify, Response
 from flask_apispec import use_kwargs, marshal_with, FlaskApiSpec
+from inspire_classifier.api import predict_coreness
 from marshmallow import fields
 
 from . import serializers
-from .domain.models import CoreClassifier
 
 
 class JsonResponse(Response):
@@ -43,9 +41,10 @@ def force_type(cls, rv, environ=None):
 
 
 def create_app():
-    classifier = CoreClassifier()
     Flask.response_class = JsonResponse
     app = Flask(__name__)
+    app.config['CLASSIFIER_BASE_PATH'] = app.instance_path
+    app.config.from_object('inspire_classifier.config')
     docs = FlaskApiSpec(app)
 
     @app.route("/api/health")
@@ -62,9 +61,7 @@ def date():
     def core_classifier(**kwargs):
         """Endpoint for the CORE classifier."""
 
-        classifier.predict(kwargs['title'], kwargs['abstract'])
-
-        return classifier
+        return predict_coreness(kwargs['title'], kwargs['abstract'])
 
     docs.register(core_classifier)
 

inspire_classifier/config.py

@@ -0,0 +1,48 @@
+# -*- coding: utf-8 -*-
+#
+# This file is part of INSPIRE.
+# Copyright (C) 2014-2018 CERN.
+#
+# INSPIRE is free software: you can redistribute it and/or modify
+# it under the terms of the GNU General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# INSPIRE is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+# GNU General Public License for more details.
+#
+# You should have received a copy of the GNU General Public License
+# along with INSPIRE. If not, see <http://www.gnu.org/licenses/>.
+#
+# In applying this license, CERN does not waive the privileges and immunities
+# granted to it by virtue of its status as an Intergovernmental Organization
+# or submit itself to any jurisdiction.
+
+"""Classifier Configuration."""
+
+
+CLASSIFIER_MAXIMUM_VOCABULARY_SIZE = 60000
+CLASSIFIER_MINIMUM_WORD_FREQUENCY = 2
+CLASSIFIER_VALIDATION_DATA_FRACTION = 0.1
+CLASSIFIER_LANGUAGE_MODEL_CYCLE_LENGTH = 15
+CLASSIFIER_CLASSIFIER_CYCLE_LENGTH = 14
+CLASSIFIER_LANGUAGE_MODEL_BATCH_SIZE = 32
+CLASSIFIER_CLASSIFIER_BATCH_SIZE = 10
+CLASSIFIER_CUDA_DEVICE_ID = 0
+
+CLASSIFIER_DATA_PATH = 'data'
+CLASSIFIER_LANGUAGE_MODEL_PATH = 'models/language_model'
+CLASSIFIER_CLASSIFIER_MODEL_PATH = 'models/classifier_model'
+CLASSIFIER_LANGUAGE_MODEL_DATA_PATH = 'data/language_model_data'
+CLASSIFIER_CLASSIFIER_DATA_PATH = 'data/classifier_data'
+CLASSIFIER_DATAFRAME_PATH = 'data/inspire_data.df'
+CLASSIFIER_PRETRAINED_LANGUAGE_MODEL_PATH = 'models/language_model/wikitext_103/fwd_wt103.h5'
+CLASSIFIER_FINETUNED_LANGUAGE_MODEL_ENCODER_PATH = 'models/language_model/finetuned_language_model_encoder.h5'
+CLASSIFIER_TRAINED_CLASSIFIER_PATH = 'models/classifier_model/trained_classifier_model.h5'
+CLASSIFIER_WIKITEXT103_ITOS_PATH = 'models/language_model/wikitext_103/itos_wt103.pkl'
+CLASSIFIER_DATA_ITOS_PATH = 'data/inspire_data_itos.pkl'
+
+CLASSIFIER_WIKITEXT103_LANGUAGE_MODEL_URL = 'http://files.fast.ai/models/wt103/fwd_wt103.h5'
+CLASSIFIER_WIKITEXT103_ITOS_URL = 'http://files.fast.ai/models/wt103/itos_wt103.pkl'

inspire_classifier/domain/__init__.py

@@ -19,5 +19,3 @@
 # In applying this license, CERN does not waive the privileges and immunities
 # granted to it by virtue of its status as an Intergovernmental Organization
 # or submit itself to any jurisdiction.
-
-from __future__ import absolute_import, division, print_function