Skip to main content

An abstract library for implementing text classification tasks based on various transformers based language models

Project description


ClassiTransformers is an abstract library based on Tensorflow implementation of BERT and ELECTRA, and transformers library of HuggingFace Inc.

Currently implemented models

  • BERT (Tensorflow)
  • ELECTRA (Tensorflow)
  • RoBERTa (PyTorch - transformers library)
  • ALBERT (PyTorch - transformers library)
  • DistilBERT (PyTorch - transformers library)


  • Sequence Classification


  • Works for N-class classification problem where N is any number of classes.
  • Easy to use. Takes away all the complexity of writing tensorflow or pytorch codes for training and testing classification models.
  • It provides an methods to easily train, test and create deployable models in .pb and .bin format in just 5 steps.
  • Hyperparameters can be easily modified without having to change the source code.

Table of contents


Assuming that anaconda environment is already installed,

  • with requirements.txt
pip install -r requirements.txt
  • with yml file, create conda environment
conda env create -f environment.yml
source activate env


Example notebooks can be found in the sample_notebooks directory.


  • classitransformers.pytransformers - Includes all pytorch-based text classification models from transformers library.
  • classitransformers.tfelectra - Includes tensorflow-based Electra model for text classification
  • classitransformers.tfbert - Includes tensorflow-based BERT model for text classification
  • - Used for reporting performance metrics. (precision, recall, F1, confusion matric)
  • classitransformers.configs - Used for initializing the hyperparameters of the language models. Also checkas and creates the necessary directories.
  • classitransformers.downloader - Used for downloading any of the 5 language models.

Qucik Start

Supports text classification with any number of labels.

from import metrics
from classitransformers.configs import Configs
from classitransformers.tfelectra import ElectraClassification

config = Configs(pretrained_model_dir = './models/Electra_base/',
              train_batch_size = 16,
              eval_batch_size = 8, 
              do_train = True, 
              do_eval = True, 
              label_list = ["0", "1", "2", "3", "4"],
              max_seq_length = 256,
              data_dir = "./datasets/bbcsports/", 
              output_dir = "./electra_output_sports/")

model = ElectraClassification(config) 

prediction = model.test()

y_pred = [np.argmax(tup) for tup in prediction]

Data Preparation

The directory for input files needs to be specified while creating the config object. The files should be named train.csv, dev.csv, test.csv .test.csv may or may not have the labels. Labels would be useful forgenerating the report after testing.

Please check data_preparation_format.txt for details.

Setting Configuration

All the Hyperparameters are listed in the Configs class, and have standard default values. The values can be changed by modifying the parameters passed in the Configs constructor for config object.

class Configs(object):
  """Fine-tuning hyperparameters."""

    def __init__(self, pretrained_model_dir = './Albert',
              data_dir = "./data/", output_dir = "./albert_output/",
              export_dir = None, model_name="albert", 
              model_size ="base", learning_rate = 5e-5, 
              num_train_epochs=3.0, train_batch_size = 16,
              eval_batch_size = 8, predict_batch_size = 8, do_train = True,
              do_eval = True, do_predict = False, label_list = ["0", "1"],
              do_lower_case = True, max_seq_length = 256, use_tpu = False,
              iterations_per_loop = 1000, save_checkpoint_steps = 1000000,
              warmup_proportion = 0.1, export_path ='./exported_bert_model')

These are the parameters to be specified for creating the config object of Configs class.


  • pretrained_model_dir : The path for pretrained directory.
  • data_dir : The path of the directory for the train,dev and test files.
  • output_dir (optional): The directory where the fine-tuned model will be saved. If not given, model will be saved in the current directory.(checkpoint for TF, .bin for pytorch)
  • export_dir (optional): The directory where the model to be deployed will be saved.(Currently only for BERT)
  • model_name : The name of the model. Either of these: 'albert', 'bert', 'electra', 'roberta', 'distilbert'
  • learning_rate: The learning rate required while training the model. Default is 5e-5.
  • num_training_epochs: The number of iterations for finetuning the pretrained model for classification task.
  • label_list: The list of the labels for text classification task.
  • max_seq_length: Max Sequence Length (multiples of 2) should be ideally just greater than the length of the longest text sentence, to prevent loss of information.
  • export_path: The export path directory where chkpt format is converted to .pb format. Only set for bert.

Class Methods

The class methods do not take any parameters. All the parameters are predefined to improve the clarity of the code.

train() Fine-Tunes(trains) the model and saves the model and config file in the output_dir directory. Validation is done after each epoch.

test() Tests the model for test dataset. Returns the prediction labels.

export_model() Exports checkpoint model to .pb fotmat. Used for tensorflow-serviing while inferencing.(Currently only for BERT)

inference() Inference on any input csv in batches using tensorflow serving for .pb model. (Currently only for BERT)

text_inference() Inference on list of sentences as input.

report() Prints and returns the accuracy and other metrics. Also prints Confusion Matrix (decorated matrix using matplotlib)

Getting Language Models.

from classitransformers import downloader

# pass name of the model ('albert', 'bert', 'electra', 'roberta', 'distilbert')
downloader('roberta') # Downloads to default dir '../models'

Real Dataset Examples

Support and Contributions

Please submit bug reports and feature requests as Issues. Contributions are very welcome.

For additional questions and feedback, please contact us at


ClassiTransformers is developed by Emerging Tech Team at Fidelity Investments. The part of the package was developed as part of an internship program at Fidelity. We thank Hrushikesh and Mayank for their contribution to the package.


ClassiTransformers is licensed under the Apache License 2.0.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

classitransformers-0.0.1.tar.gz (44.8 kB view hashes)

Uploaded source

Built Distribution

classitransformers-0.0.1-py3-none-any.whl (51.8 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Huawei Huawei PSF Sponsor Microsoft Microsoft PSF Sponsor NVIDIA NVIDIA PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page