mt_training_args.yml

entry.class: trainer
entry.params:
  train_steps: 200000
  summary_steps: 200
  save_checkpoint_steps: 1000
  criterion.class: label_smoothed_cross_entropy
  criterion.params:
    label_smoothing: 0.1
  optimizer.class: adam
  optimizer.params:
    epsilon: 1.e-9
    beta_1: 0.9
    beta_2: 0.98
  lr_schedule.class: noam
  lr_schedule.params:
    initial_factor: 1.0
    dmodel: 512
    warmup_steps: 4000

dataset.class: ParallelTextDataset
dataset.params:
  src_file: DATA_PATH/mt/train/train.en.bpe.txt
  trg_file: DATA_PATH/mt/train/train.fr.tok.bpe.txt
  data_is_processed: True

task.class: seq2seq
task.params:
  batch_by_tokens: True
  batch_size: 25000
  max_src_len: 120
  max_trg_len: 150
  src_data_pipeline.class: TranscriptDataPipeline
  src_data_pipeline.params:
    remove_punctuation: True
    lowercase: True
    language: en
    tokenizer: moses
    subtokenizer: bpe
    subtokenizer_codes: DATA_PATH/mt/codes.bpe
    vocab_path: DATA_PATH/mt/vocab.en
  trg_data_pipeline.class: TextDataPipeline
  trg_data_pipeline.params:
    language: fr
    tokenizer: moses
    subtokenizer: bpe
    subtokenizer_codes: DATA_PATH/mt/codes.bpe
    vocab_path: DATA_PATH/mt/vocab.fr