name: bt_how2sign
wandb: false
data:
    base_path: ${DATA_ROOT}/lmdb
    dataset: How2Sign
    version: how2sign
    pseudo_gloss_dir: ${DATA_ROOT}/pseudo_glosses
    feature_size: 240
    level: word
    tokenizer: whitespace
    txt_lowercase: true
    max_sent_length: 400
    max_sgn_length: 300
    random_train_subset: -1
    random_dev_subset: -1
testing:
    recognition_beam_sizes:
    - 1
    translation_beam_sizes:
    - 1
    - 2
    - 3
    - 4
    - 5
    translation_beam_alphas:
    - -1
    - 0
    - 1
    - 2
    - 3
training:
    model_dir: ./checkpoints/How2Sign
    random_seed: 42
    reset_best_ckpt: false
    reset_scheduler: false
    reset_optimizer: false
    recognition_loss_weight: 1.0
    translation_loss_weight: 1.0
    eval_metric: bleu
    optimizer: adam
    learning_rate: 0.0005
    batch_size: 128
    batch_type: sentence
    num_workers: 4
    epochs: 250
    early_stopping_metric: eval_metric
    translation_normalization: batch
    eval_recognition_beam_size: 1
    eval_translation_beam_size: 3
    eval_translation_beam_alpha: -1
    num_valid_log: 5
    overwrite: false
    shuffle: true
    use_cuda: true
    translation_max_output_length: 30
    keep_last_ckpts: 3
    batch_multiplier: 1
    logging_freq: 100
    validation_freq: 100
    betas:
    - 0.9
    - 0.998
    scheduling: plateau
    learning_rate_min: 1.0e-08
    weight_decay: 0.001
    patience: 3
    decrease_factor: 0.8
    label_smoothing: 0.1
model:
    initializer: xavier
    bias_initializer: zeros
    init_gain: 1.0
    embed_initializer: xavier
    embed_init_gain: 1.0
    tied_softmax: false
    encoder:
        type: transformer
        num_layers: 3
        num_heads: 8
        embeddings:
            embedding_dim: 512
            scale: false
            dropout: 0.1
            norm_type: batch
            activation_type: softsign
        hidden_size: 512
        ff_size: 1024
        dropout: 0.1
    decoder:
        type: transformer
        num_layers: 3
        num_heads: 8
        embeddings:
            embedding_dim: 512
            scale: false
            dropout: 0.1
            norm_type: batch
            activation_type: softsign
        hidden_size: 512
        ff_size: 1024
        dropout: 0.1