Simple Transformers

This library is based on the Transformers library by HuggingFace. Simple Transformers lets you quickly train and evaluate Transformer models. Only 3 lines of code are needed to initialize a model, train the model, and evaluate a model.

Supports

Sequence Classification
Token Classification (NER)
Question Answering
Language Model Fine-Tuning
Language Model Training
Language Generation
T5 Model
Seq2Seq Tasks
Multi-Modal Classification
Conversational AI.
Text Representation Generation.

Documentation Updates

New documentation is now live at simpletransformers.ai

Currently added:

Text classification
NER
Question answering
Language model training
- Language model fine-tuning
- Training language models from scratch

Any feedback will be immensely helpful in improving the documentation! If you have any feedback, please leave a comment in the issue I've opened for this.

Latest models

2020-06-08

Longformer support added

2020-05-05

T5 Model support added

2020-04-27

ELECTRA models can now be used with Language Model Training, Named Entity Recognition (Token Classification), Sequence Classification, and Question Answering.

Install Anaconda or Miniconda Package Manager from here
Create a new virtual environment and install packages. conda create -n transformers python pandas tqdm conda activate transformers If using cuda: conda install pytorch cudatoolkit=10.1 -c pytorch else: conda install pytorch cpuonly -c pytorch
Install Apex if you are using fp16 training. Please follow the instructions here. (Installing Apex from pip has caused issues for several people.)
Install simpletransformers. pip install simpletransformers

Optional

Install Weights and Biases (wandb) for tracking and visualizing training in a web browser. pip install wandb

Usage

Most available hyperparameters are common for all tasks. Any special hyperparameters will be listed in the docs section for the corresponding class. See Default Settings and Args Explained sections for more information.

Example scripts can be found in the examples directory.

See the Changelog for up-to-date changes to the project.

Structure

The file structure has been updated starting with version 0.6.0. This should only affect import statements. The old import paths should still be functional although it is recommended to use the updated paths given below and in the minimal start examples.

simpletransformers.classification - Includes all Classification models.
- ClassificationModel
- MultiLabelClassificationModel
simpletransformers.ner - Includes all Named Entity Recognition models.
- NERModel
simpletransformers.question_answering - Includes all Question Answering models.
- QuestionAnsweringModel

Back to Table of Contents

Text Classification

Supports Binary Classification, Multiclass Classification, and Multilabel Classification.

Supported model types:

ALBERT
BERT
CamemBERT
RoBERTa
DistilBERT
ELECTRA
FlauBERT
XLM
XLM-RoBERTa
XLNet

Task Specific Notes

Set 'sliding_window': True in args to prevent text being truncated. The default stride is 'stride': 0.8 which is 0.8 * max_seq_length. Training text will be split using a sliding window and each window will be assigned the label from the original text. During evaluation and prediction, the mode of the predictions for each window will be the final prediction on each sample. The tie_value (default 1) will be used in the case of a tie. Currently not available for Multilabel Classification

Minimal Start for Binary Classification

from simpletransformers.classification import ClassificationModel
import pandas as pd
import logging


logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)

# Train and Evaluation data needs to be in a Pandas Dataframe of two columns. The first column is the text with type str, and the second column is the label with type int.
train_data = [['Example sentence belonging to class 1', 1], ['Example sentence belonging to class 0', 0]]
train_df = pd.DataFrame(train_data)

eval_data = [['Example eval sentence belonging to class 1', 1], ['Example eval sentence belonging to class 0', 0]]
eval_df = pd.DataFrame(eval_data)

# Create a ClassificationModel
model = ClassificationModel('roberta', 'roberta-base') # You can set class weights by using the optional weight argument

# Train the model
model.train_model(train_df)

# Evaluate the model
result, model_outputs, wrong_predictions = model.eval_model(eval_df)

If you wish to add any custom metrics, simply pass them as additional keyword arguments. The keyword is the name to be given to the metric, and the value is the function that will calculate the metric. Make sure that the function expects two parameters with the first one being the true label, and the second being the predictions. (This is the default for sklearn metrics)

import sklearn


result, model_outputs, wrong_predictions = model.eval_model(eval_df, acc=sklearn.metrics.accuracy_score)

To make predictions on arbitary data, the predict(to_predict) function can be used. For a list of text, it returns the model predictions and the raw model outputs.

predictions, raw_outputs = model.predict(['Some arbitary sentence'])

Minimal Start for Multiclass Classification

For multiclass classification, simply pass in the number of classes to the num_labels optional parameter of ClassificationModel.

from simpletransformers.classification import ClassificationModel
import pandas as pd
import logging


logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)

# Train and Evaluation data needs to be in a Pandas Dataframe containing at least two columns. If the Dataframe has a header, it should contain a 'text' and a 'labels' column. If no header is present, the Dataframe should contain at least two columns, with the first column is the text with type str, and the second column in the label with type int.
train_data = [['Example sentence belonging to class 1', 1], ['Example sentence belonging to class 0', 0], ['Example eval senntence belonging to class 2', 2]]
train_df = pd.DataFrame(train_data)

eval_data = [['Example eval sentence belonging to class 1', 1], ['Example eval sentence belonging to class 0', 0], ['Example eval senntence belonging to class 2', 2]]
eval_df = pd.DataFrame(eval_data)

# Create a ClassificationModel
model = ClassificationModel('bert', 'bert-base-cased', num_labels=3, args={'reprocess_input_data': True, 'overwrite_output_dir': True})
# You can set class weights by using the optional weight argument

# Train the model
model.train_model(train_df)

# Evaluate the model
result, model_outputs, wrong_predictions = model.eval_model(eval_df)

predictions, raw_outputs = model.predict(["Some arbitary sentence"])

Minimal Start for Multilabel Classification

For Multi-Label Classification, the labels should be multi-hot encoded. The number of classes can be specified (default is 2) by passing it to the num_labels optional parameter of MultiLabelClassificationModel.

Warning: Pandas can cause issues when saving and loading lists stored in a column. Check whether your list has been converted to a String!

The default evaluation metric used is Label Ranking Average Precision (LRAP) Score.

from simpletransformers.classification import MultiLabelClassificationModel
import pandas as pd
import logging


logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)

# Train and Evaluation data needs to be in a Pandas Dataframe containing at least two columns, a 'text' and a 'labels' column. The `labels` column should contain multi-hot encoded lists.
train_data = [['Example sentence 1 for multilabel classification.', [1, 1, 1, 1, 0, 1]], ['This is another example sentence. ', [0, 1, 1, 0, 0, 0]]]
train_df = pd.DataFrame(train_data, columns=['text', 'labels'])

eval_data = [['Example eval sentence for multilabel classification.', [1, 1, 1, 1, 0, 1]], ['Example eval senntence belonging to class 2', [0, 1, 1, 0, 0, 0]]]
eval_df = pd.DataFrame(eval_data)

# Create a MultiLabelClassificationModel
model = MultiLabelClassificationModel('roberta', 'roberta-base', num_labels=6, args={'reprocess_input_data': True, 'overwrite_output_dir': True, 'num_train_epochs': 5})
# You can set class weights by using the optional weight argument
print(train_df.head())

# Train the model
model.train_model(train_df)

# Evaluate the model
result, model_outputs, wrong_predictions = model.eval_model(eval_df)
print(result)
print(model_outputs)

predictions, raw_outputs = model.predict(['This thing is entirely different from the other thing. '])
print(predictions)
print(raw_outputs)

Special Attributes

The args dict of MultiLabelClassificationModel has an additional threshold parameter with default value 0.5. The threshold is the value at which a given label flips from 0 to 1 when predicting. The threshold may be a single value or a list of value with the same length as the number of labels. This enables the use of seperate threshold values for each label.
MultiLabelClassificationModel takes in an additional optional argument pos_weight. This should be a list with the same length as the number of labels. This enables using different weights for each label when calculating loss during training and evaluation.

Minimal Start for Sentence Pair Classification

Training and evaluation Dataframes must contain a text_a, text_b, and a labels column.
The predict() function expects a list of lists in the format below. A single sample input should also be a list of lists like [[text_a, text_b]].

[
    [sample_1_text_a, sample_1_text_b],
    [sample_2_text_a, sample_2_text_b],
    [sample_3_text_a, sample_3_text_b],
    # More samples
]

from simpletransformers.classification import ClassificationModel
import pandas as pd
import sklearn
import logging


logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)

train_data = [
    ['Example sentence belonging to class 1', 'Yep, this is 1', 1],
    ['Example sentence belonging to class 0', 'Yep, this is 0', 0],
    ['Example  2 sentence belonging to class 0', 'Yep, this is 0', 0]
]

train_df = pd.DataFrame(train_data, columns=['text_a', 'text_b', 'labels'])

eval_data = [
    ['Example sentence belonging to class 1', 'Yep, this is 1', 1],
    ['Example sentence belonging to class 0', 'Yep, this is 0', 0],
    ['Example  2 sentence belonging to class 0', 'Yep, this is 0', 0]
]

eval_df = pd.DataFrame(eval_data, columns=['text_a', 'text_b', 'labels'])

train_args={
    'reprocess_input_data': True,
    'overwrite_output_dir': True,
    'num_train_epochs': 3,
}

# Create a ClassificationModel
model = ClassificationModel('roberta', 'roberta-base', num_labels=2, use_cuda=True, cuda_device=0, args=train_args)
print(train_df.head())

# Train the model
model.train_model(train_df, eval_df=eval_df)

# Evaluate the model
result, model_outputs, wrong_predictions = model.eval_model(eval_df, acc=sklearn.metrics.accuracy_score)

predictions, raw_outputs = model.predict([["I'd like to puts some CD-ROMS on my iPad, is that possible?'", "Yes, but wouldn't that block the screen?"]])
print(predictions)
print(raw_outputs)

Real Dataset Examples

Yelp Reviews Dataset - Binary Classification
AG News Dataset - Multiclass Classification
Toxic Comments Dataset - Multilabel Classification
Semantic Textual Similarity Benchmark - Sentence Pair
AG News Dataset - BERT (base and distilled), RoBERTa (base and distilled), and XLNet compared
Comparing ELECTRA, BERT, RoBERTa, and XLNET

ClassificationModel

class simpletransformers.classification.ClassificationModel (model_type, model_name, args=None, use_cuda=True) This class is used for Text Classification tasks.

Class attributes

tokenizer: The tokenizer to be used.
model: The model to be used.
model_name: model_name: Default Transformer model name or path to Transformer model file (pytorch_model.bin).
device: The device on which the model will be trained and evaluated.
results: A python dict of past evaluation results for the TransformerModel object.
args: A python dict of arguments used for training and evaluation.

cuda_device: (optional) int - Default = -1. Used to specify which GPU should be used.

Parameters

model_type: (required) str - The type of model to use. Currently, BERT, XLNet, XLM, and RoBERTa models are available.
model_name: (required) str - The exact model to use. Could be a pretrained model name or path to a directory containing a model. See Current Pretrained Models for all available models.
num_labels (optional): The number of labels or classes in the dataset.
weight (optional): A list of length num_labels containing the weights to assign to each label for loss calculation.
args: (optional) python dict - A dictionary containing any settings that should be overwritten from the default values.
use_cuda: (optional) bool - Default = True. Flag used to indicate whether CUDA should be used.

class methods train_model(self, train_df, output_dir=None, show_running_loss=True, args=None, eval_df=None)

Trains the model using 'train_df'

Args:

train_df: Pandas Dataframe containing at least two columns. If the Dataframe has a header, it should contain a 'text' and a 'labels' column. If no header is present, the Dataframe should contain at least two columns, with the first column containing the text, and the second column containing the label. The model will be trained on this Dataframe.
output_dir (optional): The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
args (optional): Optional changes to the args dict of the model. Any changes made will persist for the model.
show_running_loss (optional): Set to False to disable printing running training loss to the terminal.
eval_df (optional): A DataFrame against which evaluation will be performed when evaluate_during_training is enabled. Is required if evaluate_during_training is enabled.
**kwargs: Additional metrics that should be used. Pass in the metrics as keyword arguments (name of metric: function to use). E.g. f1=sklearn.metrics.f1_score. A metric function should take in two parameters. The first parameter will be the true labels, and the second parameter will be the predictions.

Returns:

None

eval_model(self, eval_df, output_dir=None, verbose=False)

Evaluates the model on eval_df. Saves results to output_dir.

Args:

eval_df: Pandas Dataframe containing at least two columns. If the Dataframe has a header, it should contain a 'text' and a 'labels' column. If no header is present, the Dataframe should contain at least two columns, with the first column containing the text, and the second column containing the label. The model will be evaluated on this Dataframe.
output_dir: The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
verbose: If verbose, results will be printed to the console on completion of evaluation.
silent: If silent, tqdm progress bars will be hidden.
**kwargs: Additional metrics that should be used. Pass in the metrics as keyword arguments (name of metric: function to use). E.g. f1=sklearn.metrics.f1_score. A metric function should take in two parameters. The first parameter will be the true labels, and the second parameter will be the predictions.

Returns:

result: Dictionary containing evaluation results. (Matthews correlation coefficient, tp, tn, fp, fn)
model_outputs: List of model outputs for each row in eval_df
wrong_preds: List of InputExample objects corresponding to each incorrect prediction by the model. The text of the incorrect prediction can be obtained from the InputFeature.text_a attribute. To obtain the true label of the text, use InputFeature.label attribute.

predict(self, to_predict)

Performs predictions on a list of text.

Args:

to_predict: A python list of text (str) to be sent to the model for prediction.

Returns:

preds: A python list of the predictions (0 or 1) for each text.
model_outputs: A python list of the raw model outputs for each text.

If config: {"output_hidden_states": True}, two additional values will be returned.

all_embedding_outputs: Numpy array of shape (batch_size, sequence_length, hidden_size)
all_layer_hidden_states: Numpy array of shape (num_hidden_layers, batch_size, sequence_length, hidden_size)

train(self, train_dataset, output_dir)

Trains the model on train_dataset. Utility function to be used by the train_model() method. Not intended to be used directly.

evaluate(self, eval_df, output_dir, prefix="")

Evaluates the model on eval_df. Utility function to be used by the eval_model() method. Not intended to be used directly

load_and_cache_examples(self, examples, evaluate=False)

Converts a list of InputExample objects to a TensorDataset containing InputFeatures. Caches the InputFeatures. Utility function for train() and eval() methods. Not intended to be used directly

compute_metrics(self, preds, labels, eval_examples, **kwargs):

Computes the evaluation metrics for the model predictions.

Args:

preds: Model predictions
labels: Ground truth labels
eval_examples: List of examples on which evaluation was performed

Returns:

result: Dictionary containing evaluation results. (Matthews correlation coefficient, tp, tn, fp, fn)
wrong: List of InputExample objects corresponding to each incorrect prediction by the model

Back to Table of Contents

Named Entity Recognition

This section describes how to use Simple Transformers for Named Entity Recognition. (If you are updating from a Simple Transformers before 0.5.0, note that seqeval needs to be installed to perform NER.)

This model can also be used for any other NLP task involving token level classification. Make sure you pass in your list of labels to the model if they are different from the defaults.

Supported model types:

BERT
CamemBERT
DistilBERT
ELECTRA
RoBERTa
XLM-RoBERTa

model = NERModel('bert', 'bert-base-cased', labels=["LABEL_1", "LABEL_2", "LABEL_3"])

Minimal Start

from simpletransformers.ner import NERModel
import pandas as pd
import logging


logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)

# Creating train_df  and eval_df for demonstration
train_data = [
    [0, 'Simple', 'B-MISC'], [0, 'Transformers', 'I-MISC'], [0, 'started', 'O'], [1, 'with', 'O'], [0, 'text', 'O'], [0, 'classification', 'B-MISC'],
    [1, 'Simple', 'B-MISC'], [1, 'Transformers', 'I-MISC'], [1, 'can', 'O'], [1, 'now', 'O'], [1, 'perform', 'O'], [1, 'NER', 'B-MISC']
]
train_df = pd.DataFrame(train_data, columns=['sentence_id', 'words', 'labels'])

eval_data = [
    [0, 'Simple', 'B-MISC'], [0, 'Transformers', 'I-MISC'], [0, 'was', 'O'], [1, 'built', 'O'], [1, 'for', 'O'], [0, 'text', 'O'], [0, 'classification', 'B-MISC'],
    [1, 'Simple', 'B-MISC'], [1, 'Transformers', 'I-MISC'], [1, 'then', 'O'], [1, 'expanded', 'O'], [1, 'to', 'O'], [1, 'perform', 'O'], [1, 'NER', 'B-MISC']
]
eval_df = pd.DataFrame(eval_data, columns=['sentence_id', 'words', 'labels'])

# Create a NERModel
model = NERModel('bert', 'bert-base-cased', args={'overwrite_output_dir': True, 'reprocess_input_data': True})

# Train the model
model.train_model(train_df)

# Evaluate the model
result, model_outputs, predictions = model.eval_model(eval_df)

# Predictions on arbitary text strings
predictions, raw_outputs = model.predict(["Some arbitary sentence"])

print(predictions)

Real Dataset Examples

CoNLL Dataset Example

NERModel

class simpletransformers.ner.ner_model.NERModel (model_type, model_name, labels=None, args=None, use_cuda=True) This class is used for Named Entity Recognition.

Class attributes

tokenizer: The tokenizer to be used.
model: The model to be used. model_name: Default Transformer model name or path to Transformer model file (pytorch_model.bin).
device: The device on which the model will be trained and evaluated.
results: A python dict of past evaluation results for the TransformerModel object.
args: A python dict of arguments used for training and evaluation.

cuda_device: (optional) int - Default = -1. Used to specify which GPU should be used.

Parameters

model_type: (required) str - The type of model to use. Currently, BERT, XLNet, XLM, and RoBERTa models are available.
model_name: (required) str - The exact model to use. Could be a pretrained model name or path to a directory containing a model. See Current Pretrained Models for all available models.
labels (optional): A list of all Named Entity labels. If not given, ["O", "B-MISC", "I-MISC", "B-PER", "I-PER", "B-ORG", "I-ORG", "B-LOC", "I-LOC"] will be used.
args: (optional) python dict - A dictionary containing any settings that should be overwritten from the default values.
use_cuda: (optional) bool - Default = True. Flag used to indicate whether CUDA should be used.

class methods train_model(self, train_data, output_dir=None, args=None, eval_df=None)

Trains the model using 'train_data'

Args:

train_data: train_data should be the path to a .txt file containing the training data OR a pandas DataFrame with 3 columns. If a text file is used the data should be in the CoNLL format. i.e. One word per line, with sentences seperated by an empty line. The first word of the line should be a word, and the last should be a Name Entity Tag. If a DataFrame is given, each sentence should be split into words, with each word assigned a tag, and with all words from the same sentence given the same sentence_id.
output_dir: The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
show_running_loss (optional): Set to False to prevent running loss from being printed to console. Defaults to True.
args (optional): Optional changes to the args dict of the model. Any changes made will persist for the model.
eval_df (optional): A DataFrame against which evaluation will be performed when evaluate_during_training is enabled. Is required if evaluate_during_training is enabled.

Returns:

None

eval_model(self, eval_data, output_dir=None, verbose=True)

Evaluates the model on eval_data. Saves results to output_dir.

Args:

eval_data: eval_data should be the path to a .txt file containing the evaluation data or a pandas DataFrame. If a text file is used the data should be in the CoNLL format. I.e. One word per line, with sentences seperated by an empty line. The first word of the line should be a word, and the last should be a Name Entity Tag. If a DataFrame is given, each sentence should be split into words, with each word assigned a tag, and with all words from the same sentence given the same sentence_id.
output_dir: The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
verbose: If verbose, results will be printed to the console on completion of evaluation.

Returns:

result: Dictionary containing evaluation results. (eval_loss, precision, recall, f1_score)
model_outputs: List of raw model outputs
preds_list: List of predicted tags

predict(self, to_predict)

Performs predictions on a list of text.

Args:

to_predict: A python list of text (str) to be sent to the model for prediction.

Returns:

preds: A Python list of lists with dicts containg each word mapped to its NER tag.
model_outputs: A Python list of lists with dicts containing each word mapped to its list with raw model output.

train(self, train_dataset, output_dir)

Trains the model on train_dataset. Utility function to be used by the train_model() method. Not intended to be used directly.

evaluate(self, eval_dataset, output_dir, prefix="")

Evaluates the model on eval_dataset. Utility function to be used by the eval_model() method. Not intended to be used directly

load_and_cache_examples(self, data, evaluate=False, no_cache=False, to_predict=None)

Converts a list of InputExample objects to a TensorDataset containing InputFeatures. Caches the InputFeatures. Utility function for train() and eval() methods. Not intended to be used directly

Back to Table of Contents

Question Answering

Supported model types:

ALBERT
BERT
DistilBERT
ELECTRA
XLM
XLNet

Data format

For question answering tasks, the input data can be in JSON files or in a Python list of dicts in the correct format.

The file should contain a single list of dictionaries. A dictionary represents a single context and its associated questions.

Each such dictionary contains two attributes, the "context" and "qas".

context: The paragraph or text from which the question is asked.
qas: A list of questions and answers.

Questions and answers are represented as dictionaries. Each dictionary in qas has the following format.

id: (string) A unique ID for the question. Should be unique across the entire dataset.
question: (string) A question.
is_impossible: (bool) Indicates whether the question can be answered correctly from the context.
answers: (list) The list of correct answers to the question.

A single answer is represented by a dictionary with the following attributes.

answer: (string) The answer to the question. Must be a substring of the context.
answer_start: (int) Starting index of the answer in the context.

Minimal Example

from simpletransformers.question_answering import QuestionAnsweringModel
import json
import os
import logging


logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)

# Create dummy data to use for training.
train_data = [
    {
        'context': "This is the first context",
        'qas': [
            {
                'id': "00001",
                'is_impossible': False,
                'question': "Which context is this?",
                'answers': [
                    {
                        'text': "the first",
                        'answer_start': 8
                    }
                ]
            }
        ]
    },
    {
        'context': "Other legislation followed, including the Migratory Bird Conservation Act of 1929, a 1937 treaty prohibiting the hunting of right and gray whales,
            and the Bald Eagle Protection Act of 1940. These later laws had a low cost to society—the species were relatively rare—and little opposition was raised",
        'qas': [
            {
                'id': "00002",
                'is_impossible': False,
                'question': "What was the cost to society?",
                'answers': [
                    {
                        'text': "low cost",
                        'answer_start': 225
                    }
                ]
            },
            {
                'id': "00003",
                'is_impossible': False,
                'question': "What was the name of the 1937 treaty?",
                'answers': [
                    {
                        'text': "Bald Eagle Protection Act",
                        'answer_start': 167
                    }
                ]
            }
        ]
    }
]

# Save as a JSON file
os.makedirs('data', exist_ok=True)
with open('data/train.json', 'w') as f:
    json.dump(train_data, f)


# Create the QuestionAnsweringModel
model = QuestionAnsweringModel('distilbert', 'distilbert-base-uncased-distilled-squad', args={'reprocess_input_data': True, 'overwrite_output_dir': True})

# Train the model with JSON file
model.train_model('data/train.json')

# The list can also be used directly
# model.train_model(train_data)

# Evaluate the model. (Being lazy and evaluating on the train data itself)
result, text = model.eval_model('data/train.json')

print(result)
print(text)

print('-------------------')

# Making predictions using the model.
to_predict = [{'context': 'This is the context used for demonstrating predictions.', 'qas': [{'question': 'What is this context?', 'id': '0'}]}]

print(model.predict(to_predict))

Real Dataset Examples

SQuAD 2.0 - Question Answering

QuestionAnsweringModel

class simpletransformers.question_answering.QuestionAnsweringModel (model_type, model_name, args=None, use_cuda=True) This class is used for Question Answering tasks.

Class attributes

tokenizer: The tokenizer to be used.
model: The model to be used. model_name: Default Transformer model name or path to Transformer model file (pytorch_model.bin).
device: The device on which the model will be trained and evaluated.
results: A python dict of past evaluation results for the TransformerModel object.
args: A python dict of arguments used for training and evaluation.
cuda_device: (optional) int - Default = -1. Used to specify which GPU should be used.

Parameters

model_type: (required) str - The type of model to use.
model_name: (required) str - The exact model to use. Could be a pretrained model name or path to a directory containing a model. See Current Pretrained Models for all available models.
args: (optional) python dict - A dictionary containing any settings that should be overwritten from the default values.
use_cuda: (optional) bool - Default = True. Flag used to indicate whether CUDA should be used.

class methods train_model(self, train_df, output_dir=None, args=None, eval_df=None)

Trains the model using 'train_file'

Args:

train_df: Path to JSON file containing training data. The model will be trained on this file. output_dir: The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
output_dir (optional): The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
show_running_loss (Optional): Set to False to prevent training loss being printed.
args (optional): Optional changes to the args dict of the model. Any changes made will persist for the model.
eval_file (optional): Path to JSON file containing evaluation data against which evaluation will be performed when evaluate_during_training is enabled. Is required if evaluate_during_training is enabled.
**kwargs: Additional metrics that should be used. Pass in the metrics as keyword arguments (name of metric: function to use). A metric function should take in two parameters. The first parameter will be the true labels, and the second parameter will be the predictions.

Returns:

None

eval_model(self, eval_df, output_dir=None, verbose=False)

Evaluates the model on eval_file. Saves results to output_dir.

Args:

eval_file: Path to JSON file containing evaluation data. The model will be evaluated on this file.
output_dir: The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
verbose: If verbose, results will be printed to the console on completion of evaluation.
**kwargs: Additional metrics that should be used. Pass in the metrics as keyword arguments (name of metric: function to use). A metric function should take in two parameters. The first parameter will be the true labels, and the second parameter will be the predictions.

Returns:

result: Dictionary containing evaluation results. (correct, similar, incorrect)
text: A dictionary containing the 3 dictionaries correct_text, similar_text (the predicted answer is a substring of the correct answer or vise versa), incorrect_text.

predict(self, to_predict)

Performs predictions on a list of text.

Args:

to_predict: A python list of python dicts containing contexts and questions to be sent to the model for prediction.

E.g: predict([
    {
        'context': "Some context as a demo",
        'qas': [
            {'id': '0', 'question': 'What is the context here?'},
            {'id': '1', 'question': 'What is this for?'}
        ]
    }
])

n_best_size (Optional): Number of predictions to return. args['n_best_size'] will be used if not specified.

Returns:

preds: A python list containg the predicted answer, and id for each question in to_predict.

train(self, train_dataset, output_dir, show_running_loss=True, eval_file=None)

Trains the model on train_dataset. Utility function to be used by the train_model() method. Not intended to be used directly.

evaluate(self, eval_df, output_dir, , verbose=False)

Evaluates the model on eval_df. Utility function to be used by the eval_model() method. Not intended to be used directly

load_and_cache_examples(self, examples, evaluate=False, no_cache=False, output_examples=False)

Converts a list of InputExample objects to a TensorDataset containing InputFeatures. Caches the InputFeatures. Utility function for train() and eval() methods. Not intended to be used directly

Additional attributes for Question Answering tasks

QuestionAnsweringModel has a few additional attributes in its args dictionary, given below with their default values.

  'doc_stride': 384,
  'max_query_length': 64,
  'n_best_size': 20,
  'max_answer_length': 100,
  'null_score_diff_threshold': 0.0

doc_stride: int

When splitting up a long document into chunks, how much stride to take between chunks.

max_query_length: int

Maximum token length for questions. Any questions longer than this will be truncated to this length.

n_best_size: int

The number of predictions given per question.

max_answer_length: int

The maximum token length of an answer that can be generated.

null_score_diff_threshold: float

If null_score - best_non_null is greater than the threshold predict null.

Back to Table of Contents

Language Model Training

Supported model types:

BERT
CamemBERT
DistilBERT
ELECTRA
GPT-2
OpenAI-GPT
RoBERTa

Data format

The data should simply be placed in a text file. E.g.: WikiText-2

Minimal Example For Language Model Fine Tuning

The minimal example given below assumes that you have downloaded the WikiText-2 dataset.

from simpletransformers.language_modeling import LanguageModelingModel
import logging


logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)

train_args = {
    "reprocess_input_data": True,
    "overwrite_output_dir": True,
}

model = LanguageModelingModel('bert', 'bert-base-cased', args=train_args)

model.train_model("wikitext-2/wiki.train.tokens", eval_file="wikitext-2/wiki.test.tokens")

model.eval_model("wikitext-2/wiki.test.tokens")

Example (Medium Article)

Language Model Fine-tuning

Minimal Example For Language Model Training From Scratch

You can use any text file/files for training a new language model. Setting model_name to None will indicate that the language model should be trained from scratch.

Required for Language Model Training From Scratch:

train_files must be specifief when creating the LanguagModelingModel. This may be a path to a single file or a list of paths to multiple files.
vocab_size (in args dictionary)

from simpletransformers.language_modeling import LanguageModelingModel
import logging


logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)

train_args = {
    "reprocess_input_data": True,
    "overwrite_output_dir": True,
    "vocab_size": 52000,
}

model = LanguageModelingModel('roberta', None, args=train_args)

model.train_model("wikitext-2/wiki.train.tokens", eval_file="wikitext-2/wiki.test.tokens")

model.eval_model("wikitext-2/wiki.test.tokens")

Minimal Example For Language Model Training With ELECTRA

ELECTRA is a new approach to pretraining Transformer Language Models. This method is comparatively less compute-intensive.

You can use the save_discriminator() and save_generator() methods to extract the pretrained models. The two models will be saved to <output_dir>/discriminator_model and <output_dir>/generator_model by default.

from simpletransformers.language_modeling import LanguageModelingModel
import logging


logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)

train_args = {
    "reprocess_input_data": True,
    "overwrite_output_dir": True,
    "vocab_size": 52000,
}

model = LanguageModelingModel('electra', None, args=train_args, train_files="wikitext-2/wiki.train.tokens")

# Mixing standard ELECTRA architectures example
# model = LanguageModelingModel(
#     "electra",
#     None,
#     generator_name="google/electra-small-generator",
#     discriminator_name="google/electra-large-discriminator",
#     args=train_args,
#     train_files="wikitext-2/wiki.train.tokens",
# )

model.train_model("wikitext-2/wiki.train.tokens", eval_file="wikitext-2/wiki.test.tokens")

model.eval_model("wikitext-2/wiki.test.tokens")

Real Dataset Example For Training a Language Model

Esparanto Model trained with ELECTRA

LanguageModelingModel

class simpletransformers.language_modeling.LanguageModelingModel (model_type, model_name, generator_name=None, discriminator_name=None, args=None, use_cuda=True, cuda_device=-1) This class is used for language modeling tasks.

Class attributes

tokenizer: The tokenizer to be used.
model: The model to be used. model_name: Default Transformer model name or path to Transformer model file (pytorch_model.bin).
device: The device on which the model will be trained and evaluated.
results: A python dict of past evaluation results for the TransformerModel object.
args: A python dict of arguments used for training and evaluation.
cuda_device: (optional) int - Default = -1. Used to specify which GPU should be used.

Parameters

model_type: (required) str - The type of model to use.
model_name: (required) str - The exact model to use. Could be a pretrained model name or path to a directory containing a model. See Current Pretrained Models for all available models. Set to None for language model training from scratch.
generator_name: (optional) A pretrained model name or path to a directory containing an ELECTRA generator model.
discriminator_name: (optional) A pretrained model name or path to a directory containing an ELECTRA discriminator model.
args: (optional) python dict - A dictionary containing any settings that should be overwritten from the default values.
train_files: (optional) List of files to be used when training the tokenizer.
use_cuda: (optional) bool - Default = True. Flag used to indicate whether CUDA should be used.
cuda_device: (optional) Specific GPU that should be used. Will use the first available GPU by default.

class methods train_model(self, train_file, output_dir=None, show_running_loss=True, args=None, eval_file=None, verbose=True,)

Trains the model using 'train_file'

Args:

train_file: Path to text file containing the text to train the language model on.
output_dir (optional): The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
show_running_loss (Optional): Set to False to prevent training loss being printed.
args (optional): Optional changes to the args dict of the model. Any changes made will persist for the model.
eval_file (optional): Path to eval file containing the text to evaluate the language model on. Is required if evaluate_during_training is enabled.

Returns:

None

eval_model(self, eval_file, output_dir=None, verbose=True, silent=False,)

Evaluates the model on eval_file. Saves results to output_dir.

Args:

eval_file: Path to eval file containing the text to evaluate the language model on.
output_dir (optional): The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
verbose: If verbose, results will be printed to the console on completion of evaluation.
silent: If silent, tqdm progress bars will be hidden.

Returns:

result: Dictionary containing evaluation results. (correct, similar, incorrect)
text: A dictionary containing the 3 dictionaries correct_text, similar_text (the predicted answer is a substring of the correct answer or vise versa), incorrect_text.

**train_tokenizer(self, train_files, tokenizer_name=None, output_dir=None, use_trained_tokenizer=True)

Train a new tokenizer on train_files.

Args:

train_files: List of files to be used when training the tokenizer.
tokenizer_name: Name of a pretrained tokenizer or a path to a directory containing a tokenizer.
output_dir (optional): The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
use_trained_tokenizer (optional): Load the trained tokenizer once training completes.

Returns: None

train(self, train_dataset, output_dir, show_running_loss=True, eval_file=None)

Trains the model on train_dataset. Utility function to be used by the train_model() method. Not intended to be used directly.

evaluate(self, eval_dataset, output_dir, , verbose=False)

Evaluates the model on eval_dataset. Utility function to be used by the eval_model() method. Not intended to be used directly

load_and_cache_examples(self, examples, evaluate=False, no_cache=False, output_examples=False)

Reads a text file from file_path and creates training features. Utility function for train() and eval() methods. Not intended to be used directly

Additional attributes for Language Modeling tasks

LanguageModelingModel has a few additional attributes in its args dictionary, given below with their default values.

    "dataset_type": "None",
    "dataset_class": None,
    "custom_tokenizer": None,
    "block_size": 512,
    "mlm": True,
    "mlm_probability": 0.15,
    "max_steps": -1,
    "config_name": None,
    "tokenizer_name": None,
    "min_frequency": 2,
    "special_tokens": ["<s>", "<pad>", "</s>", "<unk>", "<mask>"],
    "sliding_window": False,
    "stride": 0.8
    "config": {},
    "generator_config": {},
    "discriminator_config": {},

dataset_type: str

Used to specify the Dataset type to be used. The choices are given below.

simple (or None) - Each line in the train files are considered to be a single, separate sample. sliding_window can be set to True to automatically split longer sequences into samples of length max_seq_length. Uses multiprocessing for significantly improved performance on multicore systems.
line_by_line - Treats each line in the train files as a seperate sample.
text - Treats each file in train_files as a seperate sample.

Using simple is recommended.

dataset_class: Subclass of Pytorch Dataset

A custom dataset class to use.

block_size: int

Optional input sequence length after tokenization. The training dataset will be truncated in block of this size for training. Default to the model max input length for single sentence inputs (take into account special tokens).

mlm: bool

Train with masked-language modeling loss instead of language modeling

mlm_probability: float

Ratio of tokens to mask for masked language modeling loss

max_steps: int

If > 0: set total number of training steps to perform. Override num_train_epochs.

config_name: str

Name of pretrained config or path to a directory containing a config.json file.

tokenizer_name: str

Name of pretrained tokenizer or path to a directory containing tokenizer files.

min_frequencey: int

Minimum frequency required for a word to be added to the vocabulary.

special_tokens: list

List of special tokens to be used when training a new tokenizer.

sliding_window: bool

Whether sliding window technique should be used when preparing data. Only works with SimpleDataset.

stride: float

A fraction of the max_seq_length to use as the stride when using a sliding window

config: dict

Key-values given here will override the default values used in a model Config.

generator_config: dict

Key-values given here will override the default values used in an Electra generator model Config.

discriminator_config: dict

Key-values given here will override the default values used in an Electra discriminator model Config.

Back to Table of Contents

Language Generation

This section describes how to use Simple Transformers for Langauge Generation.

Supported model types:

CTRL
GPT-2
OpenAI-GPT
Transformer-XL
XLM
XLNet

Minimal Start

import logging
from simpletransformers.language_generation import LanguageGenerationModel


logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)

model = LanguageGenerationModel("gpt2", "gpt2")
model.generate("Let's give a minimal start to the model like")

Real Dataset Examples

Writing Scientific Paper Abstracts with GPT-2

LanguageGenerationModel

class simpletransformers.language_generation.language_generation_model.LanguageGenerationModel (self, model_type, model_name, args=None, use_cuda=True, cuda_device=-1, **kwargs) This class is used for Language Generation.

Class attributes

tokenizer: The tokenizer to be used.
model: The model to be used. model_name: Default Transformer model name or path to Transformer model file (pytorch_model.bin).
device: The device on which the model will be trained and evaluated.
args: A python dict of arguments used for training and evaluation.

cuda_device: (optional) int - Default = -1. Used to specify which GPU should be used.

Parameters

model_type: (required) str - The type of model to use.
model_name: (required) str - The exact model to use. Could be a pretrained model name or path to a directory containing a model. See Current Pretrained Models for all available models.
args: (optional) python dict - A dictionary containing any settings that should be overwritten from the default values.
use_cuda: (optional) bool - Default = True. Flag used to indicate whether CUDA should be used.
cuda_device (optional): Specific GPU that should be used. Will use the first available GPU by default.
**kwargs (optional): For providing proxies, force_download, resume_download, cache_dir and other options specific to the 'from_pretrained' implementation where this will be supplied.

class methods

generate(self, prompt=None, args=None, verbose=True)

Generate text using a LanguageGenerationModel

Args:

prompt (optional): A prompt text for the model. If given, will override args["prompt"]
args (optional): Optional changes to the args dict of the model. Any changes made will persist for the model.
verbose (optional): If verbose, generated text will be logged to the console.

Returns:

generated_sequences: Sequences of text generated by the model.

Additional attributes for Language Generation tasks

LanguageGenerationModel has a few additional attributes in its args dictionary, given below with their default values.

    "do_sample": True,
    "prompt": "",
    "length": 20,
    "stop_token": None,
    "temperature": 1.0,
    "repetition_penalty": 1.0,
    "k": 0,
    "p": 0.9,
    "padding_text": "",
    "xlm_language": "",
    "num_return_sequences": 1,
    "config_name": None,
    "tokenizer_name": None,

do_sample: bool

If set to False greedy decoding is used. Otherwise sampling is used. Defaults to False as defined in configuration_utils.PretrainedConfig.

prompt: str

A prompt text for the model.

max_length: int

Length of the text to generate

stop_token: str

Token at which text generation is stopped

temperature: float

Temperature of 1.0 is the default. Lowering this makes the sampling greedier

repetition_penalty: float

Primarily useful for CTRL model; in that case, use 1.2

k: int

k value for top-k sampling

p: float

p value for top-p (nucleus) sampling

padding_text: str

Padding text for Transfo-XL and XLNet.

xlm_language: str

Optional language when used with the XLM model.

num_return_sequences: int

The number of samples to generate.

config: dict

Key-values given here will override the default values used in a model Config.

Back to Table of Contents

T5 Transformer

T5 model seems to be working fine, but please open an issue if you run across any problems

The T5 Transformer is an Encoder-Decoder architecture where both the input and targets are text sequences. The task that should be performed on the input is defined by a prefix. This means that the same T5 model can perform multiple tasks.

You can train the T5 model on a completely new task by simply using a new prefix.

Data Format

The input to a T5 model has the following pattern;

"<prefix>: <input_text> </s>"

The label sequence has the following pattern;

"<target_sequence> </s>"

Train and evaluation input formats

The inputs to both the train_model() and eval_model() methods should be a Pandas DataFrame containing the 3 columns - prefix, input_text, target_text.

prefix: A string indicating the task to perform. (E.g. "question", "stsb")
input_text: The input text sequence. prefix is automatically prepended to form the full input. (<prefix>: <input_text>)
target_text: The target sequence

If preprocess_inputs is set to True in the model args, then the < /s> tokens (including preceeding space) and the : (prefix separator including trailing separator) between prefix and input_text are automatically added. Otherwise, the input DataFrames must contain the < /s> tokens (including preceeding space) and the : (prefix separator including trailing separator).

Prediction data format

The prediction data should be a list of strings with the prefix and the : (prefix separator) included.

If preprocess_inputs is set to True in the model args, then the < /s> token (including preceeding space) is automatically added to each string in the list. Otherwise, the strings must have the < /s> (including preceeding space) must be included.

Minimal Start

import logging

import pandas as pd
from simpletransformers.t5 import T5Model

logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)


train_data = [
    ["convert", "one", "1"],
    ["convert", "two", "2"],
]

train_df = pd.DataFrame(train_data, columns=["prefix", "input_text", "target_text"])

eval_data = [
    ["convert", "three", "3"],
    ["convert", "four", "4"],
]

eval_df = pd.DataFrame(eval_data, columns=["prefix", "input_text", "target_text"])

eval_df = train_df.copy()

model_args = {
    "reprocess_input_data": True,
    "overwrite_output_dir": True,
    "max_seq_length": 10,
    "train_batch_size": 2,
    "num_train_epochs": 200,
}

# Create T5 Model
model = T5Model("t5-base", args=model_args)

# Train T5 Model on new task
model.train_model(train_df)

# Evaluate T5 Model on new task
results = model.eval_model(eval_df)

# Predict with trained T5 model
print(model.predict(["convert: four"]))

Evaluating with custom metrics

You can evaluate the models' generated sequences using custom metric functions (including evaluation during training). However, due to the way T5 outputs are generated, this may be significantly slower than evaluation with other models.

Note, you must set evaluate_generated_text to True to evaluate generated sequences.

import logging

import pandas as pd
from simpletransformers.t5 import T5Model

logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)


train_data = [
    ["convert", "one", "1"],
    ["convert", "two", "2"],
]

train_df = pd.DataFrame(train_data, columns=["prefix", "input_text", "target_text"])

eval_data = [
    ["convert", "three", "3"],
    ["convert", "four", "4"],
]

eval_df = pd.DataFrame(eval_data, columns=["prefix", "input_text", "target_text"])

eval_df = train_df.copy()

model_args = {
    "reprocess_input_data": True,
    "overwrite_output_dir": True,
    "max_seq_length": 10,
    "train_batch_size": 2,
    "num_train_epochs": 200,
    "save_eval_checkpoints": False,
    "save_model_every_epoch": False,
    # "silent": True,
    "evaluate_generated_text": True,
    "evaluate_during_training": True,
    "evaluate_during_training_verbose": True,
}

model = T5Model("t5-base", args=model_args)


def count_matches(labels, preds):
    print(labels)
    print(preds)
    return sum([1 if label == pred else 0 for label, pred in zip(labels, preds)])


model.train_model(train_df, eval_data=eval_df, matches=count_matches)

print(model.eval_model(eval_df, matches=count_matches))

Training a T5 Model on a new task

Question Generation With T5
The Guide to Multi-Tasking with the T5 Transformer

T5Model

class simpletransformers.t5.t5_model.T5Model (self, model_name, args=None, use_cuda=True, cuda_device=-1, **kwargs) This class is used for the T5 Transformer.

Class attributes

tokenizer: The tokenizer to be used.
model: The model to be used. model_name: Default Transformer model name or path to Transformer model file (pytorch_model.bin).
device: The device on which the model will be trained and evaluated.
args: A python dict of arguments used for training and evaluation.

cuda_device: (optional) int - Default = -1. Used to specify which GPU should be used.

Parameters

model_name: (required) str - The exact model to use. Could be a pretrained model name or path to a directory containing a model. See Current Pretrained Models for all available models.
args: (optional) python dict - A dictionary containing any settings that should be overwritten from the default values.
cuda_device: (optional) Specific GPU that should be used. Will use the first available GPU by default.
use_cuda: (optional) bool - Default = True. Flag used to indicate whether CUDA should be used.
**kwargs: (optional) For providing proxies, force_download, resume_download, cache_dir and other options specific to the 'from_pretrained' implementation where this will be supplied.

class methods train_model(self, train_data, output_dir=None, show_running_loss=True, args=None, eval_df=None)

Trains the model using 'train_data'

Args:

train_data: Pandas DataFrame containing the 3 columns - prefix, input_text, target_text. - prefix: A string indicating the task to perform. (E.g. "question", "stsb") - input_text: The input text sequence. prefix is automatically prepended to form the full input. (<prefix>: <input_text>) - target_text: The target sequence
output_dir (optional): The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
args (optional): Optional changes to the args dict of the model. Any changes made will persist for the model.
show_running_loss (optional): Set to False to disable printing running training loss to the terminal.
eval_data (optional): A DataFrame against which evaluation will be performed when evaluate_during_training is enabled. Is required if evaluate_during_training is enabled.
**kwargs: Additional metrics that should be used. Pass in the metrics as keyword arguments (name of metric: function to use). A metric function should take in two parameters. The first parameter will be the true labels, and the second parameter will be the predictions. Both inputs will be lists of strings. Note that this will slow down training significantly as the predicted sequences need to be generated.

Returns:

None

eval_model(self, eval_data, output_dir=None, verbose=True, silent=False)

Evaluates the model on eval_data. Saves results to output_dir.

Args:

eval_data: Pandas DataFrame containing the 3 columns - prefix, input_text, target_text.
- prefix: A string indicating the task to perform. (E.g. "question", "stsb")
- input_text: The input text sequence. prefix is automatically prepended to form the full input. (<prefix>: <input_text>)
- target_text: The target sequence
output_dir: The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
verbose: If verbose, results will be printed to the console on completion of evaluation.
silent: If silent, tqdm progress bars will be hidden.
**kwargs: Additional metrics that should be used. Pass in the metrics as keyword arguments (name of metric: function to use). A metric function should take in two parameters. The first parameter will be the true labels, and the second parameter will be the predictions. Both inputs will be lists of strings. Note that this will slow down evaluation significantly as the predicted sequences need to be generated.

Returns:

result: Dictionary containing evaluation results.

predict(self, to_predict)

Performs predictions on a list of text.

Args:

to_predict: A python list of text (str) to be sent to the model for prediction. Note that the prefix should be prepended to the text.

Returns:

preds: A python list of the generated sequences.

train(self, train_dataset, output_dir)

Trains the model on train_dataset. Utility function to be used by the train_model() method. Not intended to be used directly.

evaluate(self, eval_dataset, output_dir, prefix="")

Evaluates the model on eval_dataset. Utility function to be used by the eval_model() method. Not intended to be used directly

load_and_cache_examples(self, examples, evaluate=False)

Creates a T5Dataset from data.

Utility function for train() and eval() methods. Not intended to be used directly

compute_metrics(self, preds, labels, **kwargs):

Computes the evaluation metrics for the model predictions.

Args:

labels: List of target sequences
preds: List of model generated outputs
**kwargs: Custom metrics that should be used. Pass in the metrics as keyword arguments (name of metric: function to use). A metric function should take in two parameters. The first parameter will be the true labels, and the second parameter will be the predictions. Both inputs will be lists of strings. Note that this will slow down evaluation significantly as the predicted sequences need to be generated.

Returns:

result: Dictionary containing evaluation results.

Additional attributes for T5 Model

T5Model has a few additional attributes in its args dictionary, given below with their default values.

{
    "dataset_class": None,
    "do_sample": False,
    "max_steps": -1,
    "evaluate_generated_text": False,
    "num_beams": 1,
    "max_length": 20,
    "repetition_penalty": 1.0,
    "length_penalty": 2.0,
    "top_k": None,
    "top_p": None,
    "num_return_sequences": 1,
    "early_stopping": True,
    "preprocess_inputs": True,
}

dataset_class: Subclass of Pytorch Dataset

A custom dataset class to use.

do_sample: bool

If set to False greedy decoding is used. Otherwise sampling is used. Defaults to False as defined in configuration_utils.PretrainedConfig.

max_steps: int

Maximum number of training steps. Will override the effect of num_train_epochs.

evaluate_generated_text: bool

Generate sequences for evaluation.

num_beams: int

Number of beams for beam search. Must be between 1 and infinity. 1 means no beam search. Default to 1.

num_return_sequences: int

The number of samples to generate.

max_lemgth: int

The max length of the sequence to be generated. Between min_length and infinity. Default to 20.

repetition_penalty: float

The parameter for repetition penalty. Between 1.0 and infinity. 1.0 means no penalty. Default to 1.0.

length_penalty: float

Exponential penalty to the length. Default to 1.

top_k: int

Filter top-k tokens before sampling (<=0: no filtering)

top_p: float

Nucleus filtering (top-p) before sampling (<=0.0: no filtering)

early_stopping: bool

if set to True beam search is stopped when at least num_beams sentences finished per batch.

preprocess_inputs: bool

Automatically add : and < /s> tokens to train_model() and eval_model() inputs. Automatically add < /s> to each string in to_predict in predict().

Back to Table of Contents

Sequence-to-Sequence Models

These models are Sequence-to-Sequence models (Seq2SeqModel) where both the input and targets are text sequences. For example, translation and summarization are sequence-to-sequence tasks.

Currently, three main types of Sequence-to-Sequence models are available.

BART (Summarization)
Marian (Translation)
Encoder-Decoder (Generic)

Note that these models are not restricted to the specifed task. The task is merely given as a starting point.

Types of Sequence-to-Sequence Models

BART

Commonly used for summarization tasks.

Marian

Commonly used for translation tasks.

Encoder-Decoder

Encoder-Decoder is a generic type of Sequence-to-Sequence model and it can be configured with different Encoder-Decoder combinations.

There is a known issue with loading saved Encoder-Decoder models. The loaded model seems to underperform compared to the model that was saved.

The following rules currently apply to Encoder-Decoder models:

The decoder must be a bert model.
The encoder can be one of [bert, roberta, distilbert, camembert, electra].
The encodr and the decoder must be of the same "size". (E.g. roberta-base encoder and a bert-base-uncased decoder)

Data Format

Train and evaluation input formats

The inputs to both the train_model() and eval_model() methods should be a Pandas DataFrame containing the 2 columns - input_text and target_text.

input_text: The input text sequence.
target_text: The target text sequence.

Prediction data format

The prediction data should be a list of strings.

Minimal Starts For Sequence-to-Sequence Tasks

BART minimal start

The Seq2SeqModel must be initialized with encoder_decoder_type="bart" and encoder_decoder_name set to a pre-trained model name or the path to a saved model directory.

import logging

import pandas as pd
from simpletransformers.seq2seq import Seq2SeqModel

logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)


train_data = [
    ["one", "1"],
    ["two", "2"],
]

train_df = pd.DataFrame(train_data, columns=["input_text", "target_text"])

eval_data = [
    ["three", "3"],
    ["four", "4"],
]

eval_df = pd.DataFrame(eval_data, columns=["input_text", "target_text"])

model_args = {
    "reprocess_input_data": True,
    "overwrite_output_dir": True,
    "max_seq_length": 10,
    "train_batch_size": 2,
    "num_train_epochs": 10,
    "save_eval_checkpoints": False,
    "save_model_every_epoch": False,
    "evaluate_during_training": True,
    "evaluate_generated_text": True,
    "evaluate_during_training_verbose": True,
    "use_multiprocessing": False,
    "max_length": 15,
    "manual_seed": 4,
}

# Initialize model
model = Seq2SeqModel(
    encoder_decoder_type="bart",
    encoder_decoder_name="bart-large",
    args=model_args,
)

# Train the model
model.train_model(train_df)

# Evaluate the model
results = model.eval_model(eval_df)

# Use the model for prediction
print(model.predict(["five"]))

# Load a saved model
model1 = Seq2SeqModel(
    encoder_decoder_type="bart",
    encoder_decoder_name="outputs",
    args=model_args,
)
print(model1.predict(["five"]))

Marian minimal start

The Seq2SeqModel must be initialized with encoder_decoder_type="marian" and encoder_decoder_name set to a pre-trained model name or the path to a saved model directory.

Everything else is identical to the Bart model usage.

import logging

import pandas as pd
from simpletransformers.seq2seq import Seq2SeqModel

logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)

model_args = {
    "reprocess_input_data": True,
    "overwrite_output_dir": True,
    "max_seq_length": 50,
    "train_batch_size": 2,
    "num_train_epochs": 10,
    "save_eval_checkpoints": False,
    "save_model_every_epoch": False,
    "evaluate_generated_text": True,
    "evaluate_during_training_verbose": True,
    "use_multiprocessing": False,
    "max_length": 50,
    "manual_seed": 4,
}

model = Seq2SeqModel(
    encoder_decoder_type="marian",
    encoder_decoder_name="Helsinki-NLP/opus-mt-en-de",
    args=model_args,
)

src = [
    "People say nothing is impossible, but I do nothing every day.",
    "My opinions may have changed, but not the fact that I'm right.",
    "He who laughs last didn't get the joke.",
]

predictions = model.predict(src)

for en, de in zip(src, predictions):
    print("-------------")
    print(en)
    print(de)
    print()

Generic Encoder-Decoder minimal start

import logging

import pandas as pd
from simpletransformers.seq2seq import Seq2SeqModel

logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)


train_data = [
    ["one", "1"],
    ["two", "2"],
]

train_df = pd.DataFrame(train_data, columns=["input_text", "target_text"])

eval_data = [
    ["three", "3"],
    ["four", "4"],
]

eval_df = pd.DataFrame(eval_data, columns=["input_text", "target_text"])

model_args = {
    "reprocess_input_data": True,
    "overwrite_output_dir": True,
    "max_seq_length": 10,
    "train_batch_size": 2,
    "num_train_epochs": 10,
    "save_eval_checkpoints": False,
    "save_model_every_epoch": False,
    "evaluate_generated_text": True,
    "evaluate_during_training_verbose": True,
    "use_multiprocessing": False,
    "max_length": 15,
    "manual_seed": 4,
}

encoder_type = "roberta"

model = Seq2SeqModel(
    encoder_type,
    "roberta-base",
    "bert-base-cased",
    args=model_args,
    use_cuda=True,
)

model.train_model(train_df)

results = model.eval_model(eval_df)

print(model.predict(["five"]))


model1 = Seq2SeqModel(
    encoder_type,
    encoder_decoder_name="outputs",
    args=model_args,
    use_cuda=True,
)
print(model1.predict(["five"]))

Evaluating with custom metrics

You can evaluate the models' generated sequences using custom metric functions (including evaluation during training). However, this may be significantly slower than evaluation with other models.

Note, you must set evaluate_generated_text to True to evaluate generated sequences.

import logging

import pandas as pd
from simpletransformers.t5 import T5Model

logging.basicConfig(level=logging.INFO)
transformers_logger = logging.getLogger("transformers")
transformers_logger.setLevel(logging.WARNING)


train_data = [
    ["convert", "one", "1"],
    ["convert", "two", "2"],
]

train_df = pd.DataFrame(train_data, columns=["prefix", "input_text", "target_text"])

eval_data = [
    ["convert", "three", "3"],
    ["convert", "four", "4"],
]

eval_df = pd.DataFrame(eval_data, columns=["prefix", "input_text", "target_text"])

eval_df = train_df.copy()

model_args = {
    "reprocess_input_data": True,
    "overwrite_output_dir": True,
    "max_seq_length": 10,
    "train_batch_size": 2,
    "num_train_epochs": 200,
    "save_eval_checkpoints": False,
    "save_model_every_epoch": False,
    "evaluate_generated_text": True,
    "evaluate_during_training": True,
    "evaluate_during_training_verbose": True,
}

# Initialize model
model = Seq2SeqModel(
    encoder_decoder_type="bart",
    encoder_decoder_name="bart-large",
    args=model_args,
    use_cuda=True,
)


def count_matches(labels, preds):
    print(labels)
    print(preds)
    return sum([1 if label == pred else 0 for label, pred in zip(labels, preds)])


# Train the model
model.train_model(train_df, eval_data=eval_df, matches=count_matches)

Seq2SeqModel

class simpletransformers.seq2seq.seq2seq_model.Seq2SeqModel (self, encoder_type=None, encoder_name=None, decoder_name=None, encoder_decoder_type=None, encoder_decoder_name=None, config=None, args=None, use_cuda=True, cuda_device=-1, **kwargs) This class is used for the T5 Transformer.

Class attributes

encoder_tokenizer: The tokenizer to be used with the encoder model.
decoder_tokenizer: The tokenizer to be used with the decoder model.
model: The model to be used.
device: The device on which the model will be trained and evaluated.
args: A python dict of arguments used for training and evaluation.
cuda_device: (optional) int - Default = -1. Used to specify which GPU should be used.

Parameters

encoder_type: (optional) str - The type of model to use as the encoder.
encoder_name: (optional) str - The exact model to use as the encoder. Could be a pretrained model name or path to a directory containing a model. See Current Pretrained Models for all available models.
decoder_name: (optional) str - The exact model to use as the decoder. Could be a pretrained model name or path to a directory containing a model. See Current Pretrained Models for all available models.
encoder_decoder_type: (optional) str - The type of encoder-decoder model. (E.g. bart)
encoder_decoder_name: (optional) str - The path to a directory containing the saved encoder and decoder of a Seq2SeqModel. (E.g. "outputs/") OR a valid BART model OR a valid Marian model.
args: (optional) python dict - A dictionary containing any settings that should be overwritten from the default values.
cuda_device: (optional) Specific GPU that should be used. Will use the first available GPU by default.
use_cuda: (optional) bool - Default = True. Flag used to indicate whether CUDA should be used.
**kwargs: (optional) For providing proxies, force_download, resume_download, cache_dir and other options specific to the 'from_pretrained' implementation where this will be supplied.

class methods train_model(self, train_data, output_dir=None, show_running_loss=True, args=None, eval_df=None)

Trains the model using 'train_data'

Args:

train_data: Pandas DataFrame containing the 2 columns - input_text, target_text.
- input_text: The input text sequence.
- target_text: The target sequence
output_dir (optional): The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
args (optional): Optional changes to the args dict of the model. Any changes made will persist for the model.
show_running_loss (optional): Set to False to disable printing running training loss to the terminal.
eval_data (optional): A DataFrame against which evaluation will be performed when evaluate_during_training is enabled. Is required if evaluate_during_training is enabled.
**kwargs: Additional metrics that should be used. Pass in the metrics as keyword arguments (name of metric: function to use). A metric function should take in two parameters. The first parameter will be the true labels, and the second parameter will be the predictions. Both inputs will be lists of strings. Note that this will slow down training significantly as the predicted sequences need to be generated.

Returns:

None

eval_model(self, eval_data, output_dir=None, verbose=True, silent=False)

Evaluates the model on eval_data. Saves results to output_dir.

Args:

eval_data: Pandas DataFrame containing the 2 columns - input_text, target_text.
- input_text: The input text sequence.
- target_text: The target sequence
output_dir: The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
verbose: If verbose, results will be printed to the console on completion of evaluation.
silent: If silent, tqdm progress bars will be hidden.
**kwargs: Additional metrics that should be used. Pass in the metrics as keyword arguments (name of metric: function to use). A metric function should take in two parameters. The first parameter will be the true labels, and the second parameter will be the predictions. Both inputs will be lists of strings. Note that this will slow down evaluation significantly as the predicted sequences need to be generated.

Returns:

result: Dictionary containing evaluation results.

predict(self, to_predict)

Performs predictions on a list of text.

Args:

to_predict: A python list of text (str) to be sent to the model for prediction. Note that the prefix should be prepended to the text.

Returns:

preds: A python list of the generated sequences.

train(self, train_dataset, output_dir)

Trains the model on train_dataset. Utility function to be used by the train_model() method. Not intended to be used directly.

evaluate(self, eval_dataset, output_dir, prefix="")

Evaluates the model on eval_dataset. Utility function to be used by the eval_model() method. Not intended to be used directly

load_and_cache_examples(self, examples, evaluate=False)

Creates a T5Dataset from data.

Utility function for train() and eval() methods. Not intended to be used directly

compute_metrics(self, preds, labels, **kwargs):

Computes the evaluation metrics for the model predictions.

Args:

labels: List of target sequences
preds: List of model generated outputs
**kwargs: Custom metrics that should be used. Pass in the metrics as keyword arguments (name of metric: function to use). A metric function should take in two parameters. The first parameter will be the true labels, and the second parameter will be the predictions. Both inputs will be lists of strings. Note that this will slow down evaluation significantly as the predicted sequences need to be generated.

Returns:

result: Dictionary containing evaluation results.

Additional attributes for `Seq2SeqModel`

Seq2SeqModel has a few additional attributes in its args dictionary, given below with their default values.

{
    "dataset_class": None,
    "do_sample": False,
    "max_steps": -1,
    "evaluate_generated_text": False,
    "num_beams": 1,
    "max_length": 20,
    "repetition_penalty": 1.0,
    "length_penalty": 2.0,
    "early_stopping": True,
}

dataset_class: Subclass of Pytorch Dataset

A custom dataset class to use.

do_sample: bool

If set to False greedy decoding is used. Otherwise sampling is used. Defaults to False as defined in configuration_utils.PretrainedConfig.

max_steps: int

Maximum number of training steps. Will override the effect of num_train_epochs.

evaluate_generated_text: bool

Generate sequences for evaluation.

num_beams: int

Number of beams for beam search. Must be between 1 and infinity. 1 means no beam search. Default to 1.

num_return_sequences: int

The number of samples to generate.

max_lemgth: int

The max length of the sequence to be generated. Between min_length and infinity. Default to 20.

repetition_penalty: float

The parameter for repetition penalty. Between 1.0 and infinity. 1.0 means no penalty. Default to 1.0.

length_penalty: float

Exponential penalty to the length. Default to 1.

top_k: int

Filter top-k tokens before sampling (<=0: no filtering)

top_p: float

Nucleus filtering (top-p) before sampling (<=0.0: no filtering)

early_stopping: bool

if set to True beam search is stopped when at least num_beams sentences finished per batch.

Back to Table of Contents

Conversational AI

Chatbot creation based on the Hugging Face State-of-the-Art Conversational AI.

Supported model types:

GPT
GPT2

Data format

Data format follows the Facebook Persona-Chat format. A JSON formatted version by Hugging Face is found here. The JSON file is directly compatible with this library (and it will be automatically downloaded and used if no dataset is specified).

Each entry in personachat is a dict with two keys personality and utterances, the dataset is a list of entries.

personality: list of strings containing the personality of the agent
utterances: list of dictionaries, each of which has two keys which are lists of strings.
- candidates: [next_utterance_candidate_1, ..., next_utterance_candidate_19] The last candidate is the ground truth response observed in the conversational data
- history: [dialog_turn_0, ... dialog_turn N], where N is an odd number since the other user starts every conversation.

Preprocessing:

Spaces before periods at end of sentences
everything lowercase

Example train data:

[
    {
        "personality": [
            "i like computers .",
            "i like reading books .",
            "i like talking to chatbots .",
            "i love listening to classical music ."
        ],
        "utterances": [
            {
                "candidates": [
                    "i try to wear all black every day . it makes me feel comfortable .",
                    "well nursing stresses you out so i wish luck with sister",
                    "yeah just want to pick up nba nfl getting old",
                    "i really like celine dion . what about you ?",
                    "no . i live near farms .",
                    "mother taught me to cook ! we are looking for an exterminator .",
                    "i enjoy romantic movie . what is your favorite season ? mine is summer .",
                    "editing photos takes a lot of work .",
                    "you must be very fast . hunting is one of my favorite hobbies .",
                    "hi there . i'm feeling great! how about you ?"
                ],
                "history": [
                    "hi , how are you ?"
                ]
            },
            {
                "candidates": [
                    "i have trouble getting along with family .",
                    "i live in texas , what kind of stuff do you do in ",
                    "toronto ?",
                    "that's so unique ! veganism and line dancing usually don't mix !",
                    "no , it isn't that big . do you travel a lot",
                    "that's because they are real ; what do you do for work ?",
                    "i am lazy all day lol . my mom wants me to get a job and move out",
                    "i was born on arbor day , so plant a tree in my name",
                    "okay , i should not tell you , its against the rules ",
                    "i like to talk to chatbots too ! do you know why ? ."
                ],
                "history": [
                    "hi , how are you ?",
                    "hi there . i'm feeling great! how about you ?",
                    "not bad ! i am trying out this chatbot ."
                ]
            },
            {
                "candidates": [
                    "ll something like that . do you play games ?",
                    "does anything give you relief ? i hate taking medicine for mine .",
                    "i decorate cakes at a local bakery ! and you ?",
                    "do you eat lots of meat",
                    "i am so weird that i like to collect people and cats",
                    "how are your typing skills ?",
                    "yeah . i am headed to the gym in a bit to weight lift .",
                    "yeah you have plenty of time",
                    "metal is my favorite , but i can accept that people listen to country . haha",
                    "that's why you desire to be controlled . let me control you person one .",
                    "two dogs they are the best , how about you ?",
                    "you do art ? what kind of art do you do ?",
                    "i love watching baseball outdoors on sunny days .",
                    "oh i see . do you ever think about moving ? i do , it is what i want .",
                    "because i am a chatbot too, silly !"
                ],
                "history": [
                    "hi , how are you ?",
                    "hi there . i'm feeling great! how about you ?",
                    "not bad ! i am trying out this chatbot .",
                    "i like to talk to chatbots too ! do you know why ? .",
                    "no clue, why don't you tell me ?"
                ]
            }
        ]
    }
]

Minimal Example

You can download the pretrained (OpenAI GPT based) Conversation AI model open-sourced by Hugging Face here.

For the minimal example given below, you can download the model and extract it to gpt_personachat_cache. Note that you can use any of the other GPT or GPT-2 models but they will require more training.

You will also need to create the JSON file given in the Data Format section and save it as data/minimal_train.json.

from simpletransformers.conv_ai import ConvAIModel


train_args = {
    "num_train_epochs": 50,
    "save_model_every_epoch": False,
}

# Create a ConvAIModel
model = ConvAIModel("gpt", "gpt_personachat_cache", use_cuda=True, args=train_args)

# Train the model
model.train_model("data/minimal_train.json")

# Evaluate the model
model.eval_model()

# Interact with the trained model.
model.interact()

The interact() method can be given a list of Strings which will be used to build a personality. If a list of Strings is not given, a random personality will be chosen from PERSONA-CHAT instead.

Real Dataset Example

Persona-Chat Conversational AI

ConvAIModel

class simpletransformers.conv_ai.ConvAIModel ( model_type, model_name, args=None, use_cuda=True, cuda_device=-1, **kwargs) This class is used to build Conversational AI.

Class attributes

tokenizer: The tokenizer to be used.
model: The model to be used. model_name: Default Transformer model name or path to Transformer model file (pytorch_model.bin).
device: The device on which the model will be trained and evaluated.
results: A python dict of past evaluation results for the TransformerModel object.
args: A python dict of arguments used for training and evaluation.

Parameters

model_type: (required) str - The type of model to use.
model_name: (required) str - The exact model to use. Could be a pretrained model name or path to a directory containing a model. See Current Pretrained Models for all available models.
args: (optional) python dict - A dictionary containing any settings that should be overwritten from the default values.
use_cuda: (optional) bool - Default = True. Flag used to indicate whether CUDA should be used.
cuda_device: (optional) int - Default = -1. Used to specify which GPU should be used.

class methods train_model(self, train_file=None, output_dir=None, show_running_loss=True, args=None, eval_file=None)

Trains the model using 'train_file'

Args:

train_df: ath to JSON file containing training data. The model will be trained on this file. output_dir: The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
output_dir (optional): The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
show_running_loss (Optional): Set to False to prevent training loss being printed.
args (optional): Optional changes to the args dict of the model. Any changes made will persist for the model.
eval_file (optional): Evaluation data against which evaluation will be performed when evaluate_during_training is enabled. If not given when evaluate_during_training is enabled, the evaluation data from PERSONA-CHAT will be used.

Returns:

None

eval_model(self, eval_file, output_dir=None, verbose=True, silent=False)

Evaluates the model on eval_file. Saves results to output_dir.

Args:

eval_file: Path to JSON file containing evaluation data. The model will be evaluated on this file. If not given, eval dataset from PERSONA-CHAT will be used.
output_dir: The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
verbose: If verbose, results will be printed to the console on completion of evaluation.
silent: If silent, tqdm progress bars will be hidden.

Returns:

result: Dictionary containing evaluation results. (correct, similar, incorrect)
text: A dictionary containing the 3 dictionaries correct_text, similar_text (the predicted answer is a substring of the correct answer or vise versa), incorrect_text.

interact(self, personality=None)

Interact with a model in the terminal.

Args:

personality (optional): A list of sentences that the model will use to build a personality. If not given, a random personality from PERSONA-CHAT will be picked.

model.interact(
    personality=[
        "i like computers .",
        "i like reading books .",
        "i love classical music .",
        "i am very social ."
    ]
)

Returns:

None

train(self, train_dataloader, output_dir, show_running_loss=True, eval_dataloader=None, verbose=verbose)

Trains the model on train_dataset. Utility function to be used by the train_model() method. Not intended to be used directly.

evaluate(self, eval_file, output_dir, verbose=True, silent=False)

Evaluates the model on eval_file. Utility function to be used by the eval_model() method. Not intended to be used directly

load_and_cache_examples(self, dataset_path=None, evaluate=False, no_cache=False, verbose=True, silent=False)

Loads, tokenizes, and prepares data for training and/or evaluation. Utility function for train() and eval() methods. Not intended to be used directly

Additional attributes for Conversational AI

ConvAIModel has a few additional attributes in its args dictionary, given below with their default values.

    "num_candidates": 2,
    "personality_permutations": 1,
    "max_history": 2,
    "lm_coef": 2.0,
    "mc_coef": 1.0,
    "no_sample": False,
    "max_length": 20,
    "min_length": 1,
    "temperature": 0.7,
    "top_k": 0,
    "top_p": 0.9,

num_candidates: int

Number of candidates for training

personality_permutations: int

Number of permutations of personality sentences".

max_history: int

Number of previous exchanges to keep in history

lm_coef: int

LM loss coefficient

mc_coef: int

Multiple-choice loss coefficient

no_sample: bool

Set to use greedy decoding instead of sampling

max_length: int

Maximum length of the output utterances

min_length: int

Minimum length of the output utterances

temperature: float

Sampling softmax temperature

top_k: int

Filter top-k tokens before sampling (<=0: no filtering)

top_p: float

Nucleus filtering (top-p) before sampling (<=0.0: no filtering)

Back to Table of Contents

Multi-Modal Classification

Multi-Modal Classification fuses text and image data. This is performed using multi-modal bitransformer models introduced in the paper Supervised Multimodal Bitransformers for Classifying Images and Text.

Supported model types:

BERT

Data format

There are several possible input formats you may use. The input formats are inspired by the MM-IMDb format. Note that several options for data preprocessing have been added for convenience and flexibility when dealing with complex datasets which can be found after the input format definitions.

1 - Directory based

Each subset of data (E.g: train and test) should be in its own directory. The path to the directory can then be given directly to either train_model() or eval_model().

Each data sample should have a text portion and an image associated with it (and a label/labels for training and evaluation data). The text for each sample should be in a separate JSON file. The JSON file may contain other fields in addition to the text itself but they will be ignored. The image associated with each sample should be in the same directory and both the text and the image must have the same identifier except for the file extension (E.g: 000001.json and 000001.jpg).

2 - Directory and file list

All data (including both train and test data) should be in the same directory. The path to this directory should be given to both train_model() and eval_model(). A second argument, files_list specifies which files should be taken from the directory. files_list can be a Python list or the path to a JSON file containing the list of files.

Each data sample should have a text portion and an image associated with it (and a label/labels for training and evaluation data). The text for each sample should be in a separate JSON file. The JSON file may contain other fields in addition to the text itself but they will be ignored. The image associated with each sample should be in the same directory and both the text and the image must have the same identifier except for the file extension (E.g: 000001.json and 000001.jpg).

3 - Pandas DataFrame

Data can also be given in a Pandas DataFrame. When using this format, the image_path argument must be specified and it should be a String of the path to the directory containing the images. The DataFrame should contain at least 3 columns as detailed below.

text (str) - The text associated with the sample.
images (str) - The relative path to the image file from image_path directory.
labels (str) - The label (or list of labels for multilabel tasks) associated with the sample.

Using custom names for column names or fields in JSON files

By default, Simple Transformers will look for column/field names text, images, and labels. However, you can define your own names to use in place of these names. This behaviour is controlled using the three attributes text_label, labels_label, and images_label in the args dictionary.

You can set your custom names when creating the model by assigning the custom name to the corresponding attribute in the args dictionary.

You can also change these values at training and/or evaluation time (but not with the predict() method) by passing the names to the arguments text_label, labels_label, and images_label. Note that the change will persist even after the method call terminates. That is, the args dictionary of the model itself will be modified.

Specifying the file type extension for image and text files

By default, Simple Transformers will assume that any paths will also include the file type extension (E.g: .json or .jpg). Alternatively, you can specify the extensions using the image_type_extension and data_type_extension attributes (for image file extensions and text file extensions respectively) in the args dictionary.

This too can be done when creating the model or when running the train_model() or eval_model() methods. The changes will persist in the args dictionary when using these methods.

The image_type_extension can be specified when using the predict() method but the change WILL NOT persist.

Label formats

With Multi-Modal Classification, labels are always given as strings. You may specify a list of labels by passing in the list to label_list argument when creating the model. If label_list is given, num_labels is not required.

If label_list is not given, num_labels is required and the labels should be Strings starting from "0" up to "<num_labels>".

Creating a Model

Create a MultiModalClassificationModel.

from simpletransformers.classification.multi_modal_classification_model import MultiModalClassificationModel


model = MultiModalClassificationModel("bert", "bert-base-uncased")

Available arguments:

"""
Args:
    model_type: The type of model (bert, xlnet, xlm, roberta, distilbert, albert)
    model_name: Default Transformer model name or path to a directory containing Transformer model file (pytorch_model.bin).
    multi_label (optional): Set to True for multi label tasks.
    label_list (optional) : A list of all the labels (str) in the dataset.
    num_labels (optional): The number of labels or classes in the dataset.
    pos_weight (optional): A list of length num_labels containing the weights to assign to each label for loss calculation.
    args (optional): Default args will be used if this parameter is not provided. If provided, it should be a dict containing the args that should be changed in the default args.
    use_cuda (optional): Use GPU if available. Setting to False will force model to use CPU only.
    cuda_device (optional): Specific GPU that should be used. Will use the first available GPU by default.
    **kwargs (optional): For providing proxies, force_download, resume_download, cache_dir and other options specific to the 'from_pretrained' implementation where this will be supplied.
"""

Training a Model

Use the train_model() method to train. You can use the auto_weights feature to balance out unbalanced datasets.

Available arguments:

"""
Args:
    data: Path to data directory containing text files (JSON) and image files OR a Pandas DataFrame.
        If a DataFrame is given, it should contain the columns [text, labels, images]. When using a DataFrame,
        image_path MUST be specified. The image column of the DataFrame should contain the relative path from
        image_path to the image.
        E.g:
            For an image file 1.jpeg located in "data/train/";
                image_path = "data/train/"
                images = "1.jpeg"
    files_list (optional): If given, only the files specified in this list will be taken from data directory.
        files_list can be a Python list or the path (str) to a JSON file containing a list of files.
    image_path (optional): Must be specified when using DataFrame as input. Path to the directory containing the
        images.
    text_label (optional): Column name to look for instead of the default "text"
    labels_label (optional): Column name to look for instead of the default "labels"
    images_label (optional): Column name to look for instead of the default "images"
    image_type_extension (optional): If given, this will be added to the end of each value in "images".
    data_type_extension (optional): If given, this will be added to the end of each value in "files_list".
    auto_weights (optional): If True, weights will be used to balance the classes.
    output_dir: The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
    show_running_loss (optional): Set to False to prevent running loss from being printed to console. Defaults to True.
    args (optional): Optional changes to the args dict of the model. Any changes made will persist for the model.
    eval_data (optional): A DataFrame against which evaluation will be performed when evaluate_during_training is enabled. Is required if evaluate_during_training is enabled.
    **kwargs: Additional metrics that should be used. Pass in the metrics as keyword arguments (name of metric: function to use). E.g. f1=sklearn.metrics.f1_score.
                A metric function should take in two parameters. The first parameter will be the true labels, and the second parameter will be the predictions.
"""

Evaluating a Model

Use the eval_model() method to evaluate. You can load a saved model by giving the path to the model directory as model_name. Note that you need to provide the same arguments when loading a saved model as you did when creating the original model.

model = MultiModalClassificationModel("bert", "outputs")
results, _ = model.eval_model("data/dataset/", "data/dev.json")

Available arguments:

"""
Args:
    data: Path to data directory containing text files (JSON) and image files OR a Pandas DataFrame.
        If a DataFrame is given, it should contain the columns [text, labels, images]. When using a DataFrame,
        image_path MUST be specified. The image column of the DataFrame should contain the relative path from
        image_path to the image.
        E.g:
            For an image file 1.jpeg located in "data/train/";
                image_path = "data/train/"
                images = "1.jpeg"
    files_list (optional): If given, only the files specified in this list will be taken from data directory.
        files_list can be a Python list or the path (str) to a JSON file containing a list of files.
    image_path (optional): Must be specified when using DataFrame as input. Path to the directory containing the
        images.
    text_label (optional): Column name to look for instead of the default "text"
    labels_label (optional): Column name to look for instead of the default "labels"
    images_label (optional): Column name to look for instead of the default "images"
    image_type_extension (optional): If given, this will be added to the end of each value in "images".
    data_type_extension (optional): If given, this will be added to the end of each value in "files_list".
    output_dir: The directory where model files will be saved. If not given, self.args['output_dir'] will be used.
    verbose: If verbose, results will be printed to the console on completion of evaluation.
    silent: If silent, tqdm progress bars will be hidden.
    **kwargs: Additional metrics that should be used. Pass in the metrics as keyword arguments (name of metric: function to use). E.g. f1=sklearn.metrics.f1_score.
                A metric function should take in two parameters. The first parameter will be the true labels, and the second parameter will be the predictions.

"""

Predicting from a trained Model

Use the predict() method to make predictions. You can load a saved model by giving the path to the model directory as model_name. Note that you need to provide the same arguments when loading a saved model as you did when creating the original model.

model = MultiModalClassificationModel("bert", "outputs")
model.predict(
    {
        "text": [
            "A lawyer is forced to defend a guilty judge, while defending other innocent clients, and trying to find punishment for the guilty and provide justice for the innocent."
        ],
        "labels": ["Crime"],
        "images": ["0078718"]
    },
    image_path="data/dataset",
    image_type_extension=".jpeg"
)

Back to Table of Contents

Text Representation Generation

Use transformers language models to generate contextual word or sentence representations from text that you can then feed to any down-stream tasks of your preference.
For more complete examples of how to use this component with downstream tasks refer to: https://github.com/ThilinaRajapakse/simpletransformers/tree/master/examples/language_representation

Minimal example for generating word embeddings

Generate a list of contextual word embeddings for every sentence in a list

from simpletransformers.language_representation import RepresentationModel
        
sentences = ["Example sentence 1", "Example sentence 2"]
model = RepresentationModel(
        model_type="bert",
        model_name="bert-base-uncased",
        use_cuda=False
    )
word_vectors = model.encode_sentences(sentences, combine_strategy=None)
assert word_vectors.shape === (2, 5, 768) # token vector for every token in each sentence, bert based models add 2 tokens per sentence by default([CLS] & [SEP])

Minimal example for generating sentence embeddings

Same code as for generating word embeddings, the only differennce is that we pass combine_strategy="mean" parameter to combine_strategy="mean"

from simpletransformers.language_representation import RepresentationModel
sentences = ["Example sentence 1", "Example sentence 2"]
model = RepresentationModel(
        model_type="bert",
        model_name="bert-base-uncased",
        use_cuda=False
    )
word_vectors = model.encode_sentences(sentences, combine_strategy="mean")
assert word_vectors.shape === (2, 768) # one sentence embedding per sentence

Back to Table of Contents

Regression

Regression tasks also use the ClassificationModel with 2 caveats.

num_labels should be 1.
regression should be True in args dict.

Regression can be used with either single sentence or sentence pair tasks.

Minimal Start for Regression

from simpletransformers.classification import ClassificationModel
import pandas as pd


train_data = [
    ['Example sentence belonging to class 1', 'Yep, this is 1', 1.8],
    ['Example sentence belonging to class 0', 'Yep, this is 0', 0.2],
    ['Example  2 sentence belonging to class 0', 'Yep, this is 0', 4.5]
]

train_df = pd.DataFrame(train_data, columns=['text_a', 'text_b', 'labels'])

eval_data = [
    ['Example sentence belonging to class 1', 'Yep, this is 1', 1.9],
    ['Example sentence belonging to class 0', 'Yep, this is 0', 0.1],
    ['Example  2 sentence belonging to class 0', 'Yep, this is 0', 5]
]

eval_df = pd.DataFrame(eval_data, columns=['text_a', 'text_b', 'labels'])

train_args={
    'reprocess_input_data': True,
    'overwrite_output_dir': True,
    'num_train_epochs': 3,

    'regression': True,
}

# Create a ClassificationModel
model = ClassificationModel('roberta', 'roberta-base', num_labels=1, use_cuda=True, cuda_device=0, args=train_args)
print(train_df.head())

# Train the model
model.train_model(train_df, eval_df=eval_df)

# Evaluate the model
result, model_outputs, wrong_predictions = model.eval_model(eval_df)

predictions, raw_outputs = model.predict([["I'd like to puts some CD-ROMS on my iPad, is that possible?'", "Yes, but wouldn't that block the screen?"]])
print(predictions)
print(raw_outputs)

Visualization Support

The Weights & Biases framework is supported for visualizing model training.

To use this, simply set a project name for W&B in the wandb_project attribute of the args dictionary. This will log all hyperparameter values, training losses, and evaluation metrics to the given project.

model = ClassificationModel('roberta', 'roberta-base', args={'wandb_project': 'project-name'})

For a complete example, see here.

Back to Table of Contents

Experimental Features

To use experimental features, import from simpletransformers.experimental.X

from simpletransformers.experimental.classification import ClassificationModel

Sliding Window For Long Sequences

Normally, sequences longer than max_seq_length are unceremoniously truncated.

This experimental feature moves a sliding window over each sequence and generates sub-sequences with length max_seq_length. The model output for each sub-sequence is averaged into a single output before being sent to the linear classifier.

Currently available on binary and multiclass classification models of the following types:

BERT
DistilBERT
RoBERTa
AlBERT
XLNet
CamemBERT

Set sliding_window to True for the ClassificationModel to enable this feature.

from simpletransformers.classification import ClassificationModel
import pandas as pd
import sklearn

# Train and Evaluation data needs to be in a Pandas Dataframe of two columns. The first column is the text with type str, and the second column in the label with type int.
train_data = [['Example sentence belonging to class 1' * 50, 1], ['Example sentence belonging to class 0', 0], ['Example  2 sentence belonging to class 0', 0]] + [['Example sentence belonging to class 0', 0] for i in range(12)]
train_df = pd.DataFrame(train_data, columns=['text', 'labels'])


eval_data = [['Example eval sentence belonging to class 1', 1], ['Example eval sentence belonging to class 0', 0]]
eval_df = pd.DataFrame(eval_data)

train_args={
    'sliding_window': True,
    'reprocess_input_data': True,
    'overwrite_output_dir': True,
    'evaluate_during_training': True,
    'logging_steps': 5,
    'stride': 0.8,
    'max_seq_length': 128
}

# Create a TransformerModel
model = ClassificationModel('camembert', 'camembert-base', args=train_args, use_cuda=False)
print(train_df.head())

# Train the model
model.train_model(train_df, eval_df=eval_df)

# Evaluate the model
result, model_outputs, wrong_predictions = model.eval_model(eval_df, acc=sklearn.metrics.accuracy_score)

predictions, raw_outputs = model.predict(["I'd like to puts some CD-ROMS on my iPad, is that possible?' — Yes, but wouldn't that block the screen?" * 25])
print(predictions)
print(raw_outputs)

Back to Table of Contents

Loading Saved Models

To load a saved model, provide the path to the directory containing the saved model as the model_name. Note that you will need to specify the correct (usually the same used in training) args when loading the model

model = ClassificationModel('roberta', 'outputs/', args={})

model = NERModel('bert', 'outputs/', args={})

Back to Table of Contents

Default Settings

The default args used are given below. Any of these can be overridden by passing a dict containing the corresponding key: value pairs to the the init method of a Model class.

self.args = {
    "output_dir": "outputs/",
    "cache_dir": "cache/",
    "best_model_dir": "outputs/best_model/",

    "fp16": True,
    "fp16_opt_level": "O1",
    "max_seq_length": 128,
    "train_batch_size": 8,
    "eval_batch_size": 8,
    "gradient_accumulation_steps": 1,
    "num_train_epochs": 1,
    "weight_decay": 0,
    "learning_rate": 4e-5,
    "adam_epsilon": 1e-8,
    "warmup_ratio": 0.06,
    "warmup_steps": 0,
    "max_grad_norm": 1.0,
    "do_lower_case": False,

    "logging_steps": 50,
    "evaluate_during_training": False,
    "evaluate_during_training_steps": 2000,
    "evaluate_during_training_verbose": False,
    "use_cached_eval_features": False,
    "save_eval_checkpoints": True
    "save_steps": 2000,
    "no_cache": False,
    "save_model_every_epoch": True,
    "tensorboard_dir": None,

    "overwrite_output_dir": False,
    "reprocess_input_data": True,

    "process_count": cpu_count() - 2 if cpu_count() > 2 else 1
    "n_gpu": 1,
    "silent": False,
    "use_multiprocessing": True,

    "wandb_project": None,
    "wandb_kwargs": {},

    "use_early_stopping": True,
    "early_stopping_patience": 3,
    "early_stopping_delta": 0,
    "early_stopping_metric": "eval_loss",
    "early_stopping_metric_minimize": True,

    "manual_seed": None,
    "encoding": None,
    "config": {},
}

Args Explained

output_dir: str

The directory where all outputs will be stored. This includes model checkpoints and evaluation results.

cache_dir: str

The directory where cached files will be saved.

best_model_dir: str

The directory where the best model (model checkpoints) will be saved if evaluate_during_training is enabled and the training loop achieves a lowest evaluation loss calculated after every evaluate_during_training_steps, or an epoch.

fp16: bool

Whether or not fp16 mode should be used. Requires NVidia Apex library.

fp16_opt_level: str

Can be '01', '02', '03'. See the Apex docs for an explanation of the different optimization levels (opt_levels).

max_seq_length: int

Maximum sequence level the model will support.

train_batch_size: int

The training batch size.

gradient_accumulation_steps: int

The number of training steps to execute before performing a optimizer.step(). Effectively increases the training batch size while sacrificing training time to lower memory consumption.

eval_batch_size: int

The evaluation batch size.

num_train_epochs: int

The number of epochs the model will be trained for.

weight_decay: float

Adds L2 penalty.

learning_rate: float

The learning rate for training.

adam_epsilon: float

Epsilon hyperparameter used in AdamOptimizer.

max_grad_norm: float

Maximum gradient clipping.

do_lower_case: bool

Set to True when using uncased models.

evaluate_during_training

Set to True to perform evaluation while training models. Make sure eval_df is passed to the training method if enabled.

evaluate_during_training_steps

Perform evaluation at every specified number of steps. A checkpoint model and the evaluation results will be saved.

evaluate_during_training_verbose

Print results from evaluation during training.

use_cached_eval_features

Evaluation during training uses cached features. Setting this to False will cause features to be recomputed at every evaluation step.

save_eval_checkpoints

Save a model checkpoint for every evaluation performed.

logging_steps: int

Log training loss and learning at every specified number of steps.

save_steps: int

Save a model checkpoint at every specified number of steps.

no_cache: bool

Cache features to disk.

save_model_every_epoch: bool

Save a model at the end of every epoch.

tensorboard_dir: str

The directory where Tensorboard events will be stored during training. By default, Tensorboard events will be saved in a subfolder inside runs/ like runs/Dec02_09-32-58_36d9e58955b0/.

overwrite_output_dir: bool

If True, the trained model will be saved to the ouput_dir and will overwrite existing saved models in the same directory.

reprocess_input_data: bool

If True, the input data will be reprocessed even if a cached file of the input data exists in the cache_dir.

process_count: int

Number of cpu cores (processes) to use when converting examples to features. Default is (number of cores - 2) or 1 if (number of cores <= 2)

n_gpu: int

Number of GPUs to use.

silent: bool

Disables progress bars.

use_multiprocessing: bool

If True, multiprocessing will be used when converting data into features. Disabling can reduce memory usage, but may substantially slow down processing.

wandb_project: str

Name of W&B project. This will log all hyperparameter values, training losses, and evaluation metrics to the given project.

wandb_kwargs: dict

Dictionary of keyword arguments to be passed to the W&B project.

use_early_stopping

Use early stopping to stop training when early_stopping_metric doesn't improve (based on early_stopping_patience, and early_stopping_delta)

early_stopping_patience

Terminate training after this many evaluations without an improvement in eval_loss greater then early_stopping_delta.

early_stopping_delta

The improvement over best_eval_loss necessary to count as a better checkpoint.

early_stopping_metric

The metric that should be used with early stopping. (Should be computed during eval_during_training).

early_stopping_metric_minimize

Whether early_stopping_metric should be minimized (or maximized).

manual_seed

Set a manual seed if necessary for reproducible results.

encoding

Specify an encoding to be used when reading text files.

config

A dictionary containing configuration options that should be overriden in a model's config.

Current Pretrained Models

For a list of pretrained models, see Hugging Face docs.

The model_types available for each task can be found under their respective section. Any pretrained model of that type found in the Hugging Face docs should work. To use any of them set the correct model_type and model_name in the args dictionary.

Back to Table of Contents

Acknowledgements

None of this would have been possible without the hard work by the HuggingFace team in developing the Pytorch-Transformers library.

Icon for the Social Media Preview made by Freepik from www.flaticon.com

Contributors ✨

Thanks goes to these wonderful people (emoji key):

_hawktang 💻	_{Mabu Manaileng} 💻	_{Ali Hamdi Ali Fadel} 💻	_{Tovly Deutsch} 💻	_hlo-world 💻	_huntertl 💻	_{Yann Defretin} 💻 📖 💬 🤔
_Manuel 📖 💻	_{Gilles Jacobs} 📖	_shasha79 💻	_{Mercedes Garcia} 💻	_{Hammad Hassan Tarar} 💻 📖	_{Todd Cook} 💻	_{Knut O. Hellan} 💻 📖
_nagenshukla 💻	_flaviussn 💻 📖	_{Marc Torrellas} 🚧	_{Adrien Renaud} 💻	_jacky18008 💻	_{Matteo Senese} 💻	_sarthakTUM 📖 💻
_djstrong 💻	_{Hyeongchan Kim} 📖	_Pradhy729 💻 🚧	_{Iknoor Singh} 📖	_{Gabriel Altay} 💻	_flozi00 📖 💻 🚧	_{alexysdussier} 💻
_{Jean-Louis Queguiner} 📖	_aced125 💻	_Laksh1997 💻	_{Changlin_NLP} 💻	_jpotoniec 💻	_fcggamou 💻 📖	_guy-mor 🐛 💻
_{Cahya Wirawan} 💻	_{BjarkePedersen} 💻	_tekkkon 💻	_{Amit Garg} 💻	_caprone 🐛	_{Ather Fawaz} 💻	_{Santiago Castro} 📖
_taranais 💻	_{Pablo N. Marino} 💻

This project follows the all-contributors specification. Contributions of any kind welcome!

If you should be on this list but you aren't, or you are on the list but don't want to be, please don't hesitate to contact me!

Files

README.md

Latest commit

History

README.md

File metadata and controls

Simple Transformers

Documentation Updates

Latest models

2020-06-08

2020-05-05

2020-04-27

Table of contents

Setup

With Conda

Optional

Usage

Structure

Text Classification

Task Specific Notes

Minimal Start for Binary Classification

Minimal Start for Multiclass Classification

Minimal Start for Multilabel Classification

Special Attributes

Minimal Start for Sentence Pair Classification

Real Dataset Examples

ClassificationModel

Named Entity Recognition

Minimal Start

Real Dataset Examples

NERModel

Question Answering

Data format

Minimal Example

Real Dataset Examples

QuestionAnsweringModel

Additional attributes for Question Answering tasks

doc_stride: int

max_query_length: int

n_best_size: int

max_answer_length: int

null_score_diff_threshold: float

Language Model Training

Data format

Minimal Example For Language Model Fine Tuning

Example (Medium Article)

Minimal Example For Language Model Training From Scratch

Minimal Example For Language Model Training With ELECTRA

Real Dataset Example For Training a Language Model

LanguageModelingModel

Additional attributes for Language Modeling tasks

dataset_type: str

dataset_class: Subclass of Pytorch Dataset

block_size: int

mlm: bool

mlm_probability: float

max_steps: int

config_name: str

tokenizer_name: str

min_frequencey: int

special_tokens: list

sliding_window: bool

stride: float

config: dict

generator_config: dict

discriminator_config: dict

Language Generation

Minimal Start

Real Dataset Examples

LanguageGenerationModel

Additional attributes for Language Generation tasks

do_sample: bool

prompt: str

max_length: int

stop_token: str

temperature: float

repetition_penalty: float

k: int

p: float

padding_text: str

Additional attributes for `Seq2SeqModel`