Skip to content
Extremely easy to use sequence to sequence library with attention, for text to text conversion tasks.
Branch: master
Clone or download
Latest commit 8900a7a Jun 18, 2019
Type Name Latest commit message Commit time
Failed to load latest commit information.
txt2txt add batch inputs to beam search(generate function) Jun 17, 2019
.gitignore Initial commit Nov 26, 2018
LICENSE Initial commit Nov 26, 2018 Update Apr 27, 2019 update version Jun 17, 2019

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



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


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], 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')
infer(input_text, model, params)

Note: Checkout for pre-trained models for english punctuation correction and grammar correction.


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

Tensorflow is not included in 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


Although txt2txt is licensed under GPL, if you want to use it commercially without open sourcing your code please email me or raise a issue in this repo so that I can provide you explicit written permission to use as you wish. The only reason for doing this is, it would be nice to know if some company is using my work.

You can’t perform that action at this time.