config_iwslt17_sparsemax.yaml

name: "get_attention_wmt_ende_best"

data:
    src: "en"
    trg: "de"
    train: "data/corpus/iwslt17-ben/en-de/train"
    dev: "data/corpus/iwslt17-ben/en-de/valid"
    test: "data/corpus/iwslt17-ben/en-de/train"
    level: "word"
    lowercase: False
    max_sent_length: 50
    src_voc_min_freq: 1
    trg_voc_min_freq: 1

testing:
    beam_size: 0  # greedy = 0
    alpha: 1.0
    ckpt: "data/saved-translation-models/iwslt-ende-bahdanau-sparsemax-new/best.ckpt"
    output_path: "data/saved-translation-models/iwslt-ende-bahdanau-sparsemax-new/out"
    save_attention: True

training:
    random_seed: 42
    optimizer: "adam"
    learning_rate: 0.001
    learning_rate_min: 0.00005
    batch_size: 32
    scheduling: "plateau"
    patience: 8
    decrease_factor: 0.5
    early_stopping_metric: "ppl"
    epochs: 100
    validation_freq: 2359
    logging_freq: 500
    eval_metric: "bleu"
    model_dir: "data/saved-translation-models/iwslt-ende-bahdanau-sparsemax-new"
    overwrite: False
    shuffle: True
    use_cuda: True
    max_output_length: 100
    print_valid_sents: [0, 1, 2, 3, 4]

model:
    tied_embeddings: False  # True only for bpe
    encoder:
        rnn_type: "lstm"
        embeddings:
            embedding_dim: 512
            scale: False
        hidden_size: 512
        bidirectional: True
        dropout: 0.3
        num_layers: 2
    decoder:
        rnn_type: "lstm"
        embeddings:
            embedding_dim: 512
            scale: False
        emb_scale: False
        hidden_size: 512
        dropout: 0.3
        hidden_dropout: 0.3
        num_layers: 2
        input_feeding: True
        init_hidden: "last"
        attention: "bahdanau"
        attn_func: "sparsemax"