Neural Machine Translation with RNN variants.

Introduction

Encoder-Decoder architecture for neural machine translation BN->EN.

• Both LSTM and GRU can be used as Encoder or Decoder.
• Scratch implementation of vocab language.py and trainer trainer.py class.
• Implementation of Bahdanau attention decoder.

Setup

Install the following if not installed.

• python 3.x
• torch cuda version
• scarceblue
• configparser

Training

• Keep your preprocess data in the data folder check out the sample.txt for data format.
• Change the appropriate variable inside experiment.ini i.e. lang1, rnn, hidden_size.
• reverse a bool variable will change the model training i.e. BN->EN to EN->BN
• Training: python train.py

Resume Training

For resume training change the obj_path from false to logs and checkpoint false to model path i.e. path/to/model.pt.

Train New Language

For train in new language you only need to change the tokenizer in train.py. You can use spacy tokenizer or custom made tokernizer.