/conv_seq2seq

A tensorflow implementation of Fairseq Convolutional Sequence to Sequence Learning(Gehring et al. 2017)

Primary LanguagePythonApache License 2.0Apache-2.0

Convolutional Seq2Seq

This is a tensorflow implementation of the convolutional seq2seq model released by Facebook. This model is orignially written via Torch/Lua in Fairseq. Considering Lua is not that popular as python in the industry and research community, I re-implemente this model with Tensorflow/Python after carefully reading the paper details and Torch/Lua codebase.

This implementation is based on the framework of Google seq2seq project, which has a detailed documentation on how to use this framework. In this conv seq2seq project, I implement the conv encoder, conv decoder, and attention mechanism, as well as other modules needed by the conv seq2seq model, which is not available in the original seq2seq project.

Requirement

  • Python 2.7.0+
  • Tensorflow 1.0+ (this version is strictly required)
  • and their dependencies

Please follow seq2seq project on how to install the Convolutional Sequence to Sequence Learning project.

How to use

For dataset, please follow seq2seq nmt guides to prepare your dataset

The following is an example of how to run iwslt de-en translation task.

Train

export PYTHONIOENCODING=UTF-8
export DATA_PATH="your iwslt de-en data path"

export VOCAB_SOURCE=${DATA_PATH}/vocab.de
export VOCAB_TARGET=${DATA_PATH}/vocab.en
export TRAIN_SOURCES=${DATA_PATH}/train.de
export TRAIN_TARGETS=${DATA_PATH}/train.en
export DEV_SOURCES=${DATA_PATH}/valid.de
export DEV_TARGETS=${DATA_PATH}/valid.en
export TEST_SOURCES=${DATA_PATH}/test.de
export TEST_TARGETS=${DATA_PATH}/test.en

export TRAIN_STEPS=1000000

export MODEL_DIR=${TMPDIR:-/tmp}/nmt_conv_seq2seq
mkdir -p $MODEL_DIR

python -m bin.train \
  --config_paths="
      ./example_configs/conv_seq2seq.yml,
      ./example_configs/train_seq2seq.yml,
      ./example_configs/text_metrics_bpe.yml" \
  --model_params "
      vocab_source: $VOCAB_SOURCE
      vocab_target: $VOCAB_TARGET" \
  --input_pipeline_train "
    class: ParallelTextInputPipelineFairseq
    params:
      source_files:
        - $TRAIN_SOURCES
      target_files:
        - $TRAIN_TARGETS" \
  --input_pipeline_dev "
    class: ParallelTextInputPipelineFairseq
    params:
       source_files:
        - $DEV_SOURCES
       target_files:
        - $DEV_TARGETS" \
  --batch_size 32 \
  --eval_every_n_steps 5000 \
  --train_steps $TRAIN_STEPS \
  --output_dir $MODEL_DIR

Test

export PRED_DIR=${MODEL_DIR}/pred
mkdir -p ${PRED_DIR}

decode with greedy search

python -m bin.infer \
  --tasks "
    - class: DecodeText" \
  --model_dir $MODEL_DIR \
  --model_params "
    inference.beam_search.beam_width: 1 
    decoder.class: seq2seq.decoders.ConvDecoderFairseq" \
  --input_pipeline "
    class: ParallelTextInputPipelineFairseq
    params:
      source_files:
        - $TEST_SOURCES" \
  > ${PRED_DIR}/predictions.txt

decode with beam search

python -m bin.infer \
  --tasks "
    - class: DecodeText
    - class: DumpBeams
      params:
        file: ${PRED_DIR}/beams.npz" \
  --model_dir $MODEL_DIR \
  --model_params "
    inference.beam_search.beam_width: 5 
    decoder.class: seq2seq.decoders.ConvDecoderFairseqBS" \
  --input_pipeline "
    class: ParallelTextInputPipelineFairseq
    params:
      source_files:
        - $TEST_SOURCES" \
  > ${PRED_DIR}/predictions.txt

calculate BLEU score

./bin/tools/multi-bleu.perl ${TEST_TARGETS} < ${PRED_DIR}/predictions.txt

For more detailed instructions, please refer to seq2seq project.

Issues and contributions are warmly welcome.