simpletransformers 0.22.0

An easy-to-use wrapper library for the Transformers library.

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. Currently supports Sequence Classification, Token Classification (NER), Question Answering, Multi-Modal Classification, and Conversational AI.

Table of contents

• Simple Transformers
• 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
    • LanguageModelingModel
    • Additional attributes for Question Answering tasks
      • line_by_line: bool
      • block_size: int
      • mlm: bool
      • mlm_probability: float
      • max_steps: int
  • Conversational AI
    • Data format
    • Minimal Example
    • Real Dataset Example
    • ConvAIModel
    • Additional attributes for Conversational AI
      • num_candidates: int
      • personality_permutations: int
      • max_history: int
      • lm_coef: int
      • mc_coef: int
      • no_sample: bool
      • max_length: int
      • min_length: int
      • temperature: float
      • top_k: float
      • top_p: float
  • Multi-Modal Classification
    • Data format
      • 1 - Directory based
      • 2 - Directory and file list
      • 3 - Pandas DataFrame
    • Using custom names for column names or fields in JSON files
    • Specifying the file type extension for image and text files
    • Label formats
    • Creating a Model
    • Training a Model
    • Evaluating a Model
    • Predicting from a trained Model
  • Regression
    • Minimal Start for Regression
  • Visualization Support
  • Experimental Features
    • Sliding Window For Long Sequences
  • Loading Saved Models
  • Default Settings
    • Args Explained
      • output_dir: str
      • cache_dir: str
      • best_model_dir: str
      • fp16: bool
      • fp16_opt_level: str
      • max_seq_length: int
      • train_batch_size: int
      • gradient_accumulation_steps: int
      • eval_batch_size: int
      • num_train_epochs: int
      • weight_decay: float
      • learning_rate: float
      • adam_epsilon: float
      • max_grad_norm: float
      • do_lower_case: bool
      • evaluate_during_training
      • evaluate_during_training_steps
      • evaluate_during_training_verbose
      • use_cached_eval_features
      • save_eval_checkpoints
      • logging_steps: int
      • save_steps: int
      • no_cache: bool
      • save_model_every_epoch: bool
      • tensorboard_dir: str
      • overwrite_output_dir: bool
      • reprocess_input_data: bool
      • process_count: int
      • n_gpu: int
      • silent: bool
      • use_multiprocessing: bool
      • wandb_project: str
      • wandb_kwargs: dict
      • use_early_stopping
      • early_stopping_patience
      • early_stopping_delta
      • early_stopping_metric
      • early_stopping_metric_minimize
      • manual_seed
  • Current Pretrained Models
  • Acknowledgements
  • Contributors ✨

Setup

With Conda

1. Install Anaconda or Miniconda Package Manager from here

2. 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

3. Install Apex if you are using fp16 training. Please follow the instructions here. (Installing Apex from pip has caused issues for several people.)

4. Install simpletransformers.
   pip install simpletransformers

Optional

1. 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:

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

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

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_nodel.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.

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.

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

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.

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
• RoBERTa
• DistilBERT
• CamemBERT
• 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_nodel.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: 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.

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 the raw model outputs for each text.

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:

• BERT
• XLNet
• XLM
• DistilBERT
• ALBERT

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_nodel.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: 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): 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.

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.

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:

• GPT-2
• OpenAI-GPT
• BERT
• RoBERTa
• DistilBERT
• CamemBERT

Data format

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

Minimal Example

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")

LanguageModelingModel

class simpletransformers.language_modeling.LanguageModelingModel (model_type, model_name, args=None, use_cuda=True, cuda_device=-1)
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_nodel.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.
• 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(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 Question Answering tasks

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

    "line_by_line": False,
    "block_size": 512,
    "mlm": True,
    "mlm_probability": 0.15,
    "max_steps": -1,

line_by_line: bool

Whether distinct lines of text in the dataset are to be handled as distinct sequences.

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.

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_nodel.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: float

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_nodel.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

---

Regression

Regression tasks also use the ClassificationModel with 2 caveats.

1. num_labels should be 1.
2. 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,
}

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.

---

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 📖

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!