This is a PyTorch implementation of the Transformer model in "Attention is All You Need" (Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin, arxiv, 2017).

A novel sequence to sequence framework utilizes the self-attention mechanism, instead of Convolution operation or Recurrent structure, and achieve the state-of-the-art performance on WMT 2014 English-to-German translation task. (2017/06/12)

The official Tensorflow Implementation can be found in: tensorflow/tensor2tensor.

To learn more about self-attention mechanism, you could read "A Structured Self-attentive Sentence Embedding".

The project support training and translation with trained model now.

Note that this project is still a work in progress.

BPE related parts are not yet fully tested.

If there is any suggestion or error, feel free to fire an issue to let me know. :)

• python 3.4+
• pytorch 1.3.1
• torchtext 0.4.0
• spacy 2.2.2+
• tqdm
• dill
• numpy

An example of training for the WMT'16 Multimodal Translation task (http://www.statmt.org/wmt16/multimodal-task.html).

# conda install -c conda-forge spacy 
python -m spacy download en
python -m spacy download de

python preprocess.py -lang_src de -lang_trg en -share_vocab -save_data m30k_deen_shr.pkl

python train.py -data_pkl m30k_deen_shr.pkl -log m30k_deen_shr -embs_share_weight -proj_share_weight -label_smoothing -save_model trained -b 256 -warmup 128000 -epoch 400

python translate.py -data_pkl m30k_deen_shr.pkl -model trained.chkpt -output prediction.txt

Since the interfaces is not unified, you need to switch the main function call from main_wo_bpe to main.

python preprocess.py -raw_dir /tmp/raw_deen -data_dir ./bpe_deen -save_data bpe_vocab.pkl -codes codes.txt -prefix deen

python train.py -data_pkl ./bpe_deen/bpe_vocab.pkl -train_path ./bpe_deen/deen-train -val_path ./bpe_deen/deen-val -log deen_bpe -embs_share_weight -proj_share_weight -label_smoothing -save_model trained -b 256 -warmup 128000 -epoch 400

• TODO:
  • Load vocabulary.
  • Perform decoding after the translation.

• Parameter settings:

  • default parameter and optimizer settings
  • label smoothing
  • target embedding / pre-softmax linear layer weight sharing.

• Elapse per epoch (on NVIDIA Titan X):

  • Training set: 0.888 minutes
  • Validation set: 0.011 minutes

• coming soon.

• Evaluation on the generated text.
• Attention weight plot.

• The byte pair encoding parts are borrowed from subword-nmt.
• The project structure, some scripts and the dataset preprocessing steps are heavily borrowed from OpenNMT/OpenNMT-py.
• Thanks for the suggestions from @srush, @iamalbert, @Zessay, @JulesGM and @ZiJianZhao.