config/config_big.yaml

experiment_name: 'runs/transformer_big'

dataset:
  src_lang: 'lo'
  src_tokenizer: 'BPE'
  src_max_seq_len: 400
  tgt_lang: 'vi'
  tgt_tokenizer: 'WordLevel'
  tgt_max_seq_len: 350
  train_dataset: 'train_clean.dat'
  validate_dataset: 'dev_clean.dat'
  tokenizer_file: "tokenizer_{0}.json"
  bleu_dataset: 'test2023'

model: # 42688527 parameters
  d_model: 512
  num_heads: 8
  d_ff: 2048
  dropout_p: 0.3
  num_encoder_layers: 4
  num_decoder_layers: 2
  model_folder: "weights"
  model_basename: "transformer_"
  preload: "big"

train:
  lr: 0.0001 # 1e-4
  batch_size: 16
  num_epochs: 40
  label_smoothing: 0.1
  on_colab: True # are you training on Colab?
  patience: 100 # (steps)
  warm_up_steps: 700