TensorFlow-Transformer

Attention Is All You Need

Translate EN into DE (Train with WMT17)

• newstest2014 BLEU: 28.51
• newstest2015 BLEU: 30.22
• newstest2016 BLEU: 33.88

Paper

• Attention Is All You Need: https://arxiv.org/abs/1706.03762
• Layer Normalization: https://arxiv.org/abs/1607.06450
• Label Smoothing: https://arxiv.org/abs/1512.00567
• Byte-Pair Encoding (BPE): https://arxiv.org/abs/1508.07909
• Beam-Search length penalty: https://arxiv.org/abs/1609.08144

Env

• GTX1080TI
• ubuntu 16.04
• CUDA 8.0
• cuDNN 5.1
• tensorflow 1.4
• numpy
• nltk (bleu)
• tqdm (iteration check bar)
• python 3

Dataset

• Preprocessed WMT17 en-de: http://data.statmt.org/wmt17/translation-task/preprocessed/

  • train_set: corpus.tc.[en, de]/corpus.tc.[en, de]
  • dev_set: dev.tar/newstest[2014, 2015, 2016].tc.[en, de]

• learn and apply Sentences were encoded using byte-pair encoding

  • -num_merges: 35000
  • -final_voca_threshold: 50
  • -train_voca_threshold: 1
  • make_file: bpe applied documents and voca

Code

• transformer.py

  • Transformer graph

• inference_helper.py

  • greedy
  • beam (length penalty applied)
  • bleu (nltk)

• bucket_data_helper.py

  • bucket으로 구성된 데이터를 쉽게 가져오도록 하는 class

• make_dataset.py

  • generate bucketed bpe2idx dataset for train, valid, test from bpe applied dataset
  • need MakeFile of Sentences were encoded using byte-pair encoding
  • command:
    • make bucket train_set wmt17

     python make_dataset.py 
       -mode train 
       -source_input_path path/bpe_wmt17.en (source bpe applied document data)
       -source_out_path path/source_idx_wmt17_en.csv (source bpe idx data)
       -target_input_path path/bpe_wmt17.de (target bpe applied document data)
       -target_out_path path/target_idx_wmt17_de.csv (target bpe idx data)
       -bucket_out_path ./bpe_dataset/train_set_wmt17 (bucket trainset from source bpe idx data, target bpe idx data)
       -voca_path voca_path/voca_file_name (bpe voca from bpe_learn.py)
    

    • make bucket valid_set newstest2014

     python make_dataset.py 
       -mode infer 
       -source_input_path path/bpe_newstest2014.en (source bpe applied document data)
       -source_out_path path/source_idx_newstest2014_en.csv (source bpe idx data)
       -target_input_path path/dev.tar/newstest2014.tc.de (target original raw data)
       -bucket_out_path ./bpe_dataset/valid_set_newstest2014 (bucket validset from source bpe idx data, target original raw data)
       -voca_path voca_path/voca_file_name (bpe voca from bpe_learn.py)
    

    • make bucket test_set newstest2015

     python make_dataset.py 
       -mode infer 
       -source_input_path path/bpe_newstest2015.en (source bpe applied document data)
       -source_out_path path/source_idx_newstest2015_en.csv (source bpe idx data)
       -target_input_path path/dev.tar/newstest2015.tc.de (target original raw data)
       -bucket_out_path ./bpe_dataset/test_set_newstest2015 (bucket testset from source bpe idx data, target original raw data)
       -voca_path voca_path/voca_file_name (bpe voca from bpe_learn.py)
    

    • make bucket test_set newstest2016

     python make_dataset.py 
       -mode infer 
       -source_input_path path/bpe_newstest2016.en (source bpe applied document data)
       -source_out_path path/source_idx_newstest2016_en.csv (source bpe idx data)
       -target_input_path path/dev.tar/newstest2016.tc.de (target original raw data)
       -bucket_out_path ./bpe_dataset/test_set_newstest2016 (bucket testset from source bpe idx data, target original raw data)
       -voca_path voca_path/voca_file_name (bpe voca from bpe_learn.py)
    

• translation_train.py

  • en -> de translation train, validation, test
  • command

     python translation_train.py 
       -train_path_2017 ./bpe_dataset/train_set_wmt17 
       -valid_path_2014 ./bpe_dataset/valid_set_newstest2014 
       -test_path_2015 ./bpe_dataset/test_set_newstest2015 
       -test_path_2016 ./bpe_dataset/test_set_newstest2016 
       -voca_path voca_path/voca_file_name
    

Training

1. WMT17 Dataset Download
2. apply byte-pair_encoding
3. run make_dataset.py
4. run translation_train.py

Reference

• https://jalammar.github.io/illustrated-transformer/
• https://github.com/Kyubyong/transformer