config.yml

# MODEL DIR
model_dir: model 


# DATA
data:
    train_features_file: train.src
    train_labels_file: train.tgt
    eval_features_file: dev.src
    eval_labels_file: dev.tgt
    source_vocabulary: vocab
    target_vocabulary: vocab


# TRAINING PARAMS
train:
  batch_size: 4096
  batch_type: tokens 
  effective_batch_size: 25000   # Tune gradient accumulation to train with at least this effective batch size
  save_checkpoints_steps: 5000 
  keep_checkpoint_max: 2
  save_summary_steps: 100
  max_step: 1000000
  single_pass: false            # a single pass over the training data
  average_last_checkpoints: 0   # model/avg


# EVALUATION PARAMS
eval:
  batch_size: 32
  steps: 5000
  save_eval_predictions: false  
  external_evaluators: bleu     # sacreBLEU, BLEU, BLEU-detok, ROUGE
  early_stopping:
    metric: bleu                # bleu, loss
    min_improvement: 0.01
    steps: 4

# INFERENCE PARAMS
infer:
  batch_size: 64
  batch_type: examples 
  length_bucket_width: 5
  with_scores: false
  with_alignments: null


# MODEL/OPTIM PARAMS
params:
  beam_width: 5