txt2txt - An extremely easy to use seq2seq implementation with Attention for text to text use cases

Documentation can be found at http://bpraneeth.com/docs/txt2txt/

Examples

1. Adding two numbers
2. More Complex Math and fit_generator

Installation

pip install txt2txt

Training a model

from txt2txt import build_params, build_model, convert_training_data

input_data = ['123', '213', '312', '321', '132', '231']
output_data = ['123', '123', '123', '123', '123', '123']

build_params(input_data = input_data, output_data = output_data, params_path = 'test/params', max_lenghts=(10, 10))
    
model, params = build_model(params_path='test/params')

input_data, output_data = convert_training_data(input_data, output_data, params)
    
checkpoint = ModelCheckpoint('test/checkpoint', monitor='val_acc', verbose=1, save_best_only=True, mode='max')
callbacks_list = [checkpoint]

model.fit(input_data, output_data, validation_data=(input_data, output_data), batch_size=2, epochs=20, callbacks=callbacks_list)

Loading a trained model and running inference

from txt2txt import build_model, infer
model, params = build_model(params_path='test/params')
model.load_weights('path_to_checkpoint_file')
infer(input_text, model, params)

Note: Checkout https://github.com/bedapudi6788/deepcorrect for pre-trained models for english punctuation correction and grammar correction.

Requirements

This module needs Keras and Tensorflow. (tested with tf>=1.8.0, keras>=2.2.0).

Tensorflow is not included in setup.py and needs to be installed seperately.

What's the use of this module

Working with seq2seq tasks in NLP, I realised there aren't any easy to use, simple to understand and good performing libraries available for this. Though libraries like FairSeq or transformer are available they are in general either too complex for a newbie to understand or most probably overkill (and are very tough to train) for simple projects.

This module provides pre-built seq2seq model with Attention that performs excellently on most of the "simple" NLP taks. (Tested with Punctuation correction, transliteration and spell correction)

To Do

Make number of encoder and decoder layers configurable

Give option to add language model probability in beam search