A simple Tensorflow implementation of the Transformer

This project is a simple implementation of Tensor2tensor (https://github.com/tensorflow/tensor2tensor) for machine translation.

Usage

• Preprosessing. Prepare the parallel data (token, bpe, vocab, and so on), run ./datagen.sh to generate data.

• Training. Modify the model params (transformer_params_big or transformer_params_base, basic params are set in models/common_hparms.py), and run ./train.sh.

• inference. Run the command to translate source sentence: ./test.sh.

C++ Version Translator

Once you have trained the model, you can also use a C++/CUDA Transformer-translator to translate source language.

Contact

If you have questions, suggestions and bug reports, please email wszlong@gmail.com or long.zhou@nlpr.ia.ac.cn.