name: "en_kin_transformer"

data:
    src: 
        lang: "en"
        level: "bpe"
        lowercase: False
        tokenizer_type: "subword-nmt"
        num_merges: 4000
        tokenizer_cfg:
            num_merges: 4000
            codes: "models/en_kin_transformer/bpe.codes.4000"
            pretokenizer: "none"

    trg: 
        lang: "kin"
        level: "bpe"
        lowercase: False
        tokenizer_type: "subword-nmt"
        num_merges: 4000
        tokenizer_cfg:
            num_merges: 4000
            codes:  "models/en_kin_transformer/bpe.codes.4000"
            pretokenizer: "none"

    train: "data/train/en_kin_train"
    dev:   "data/val/en_kin_val"
    test:  "data/test/en_kin_test"
    level: "bpe"
    # lowercase: False
    max_sent_length: 100
    # src_vocab: "vocab.txt"
    # trg_vocab: "vocab.txt"
    dataset_type: "tsv"

testing:
    beam_size: 15
    beam_alpha: 1.0
    eval_metrics: ["bleu"]
    batch_type: "sentence"
    sacrebleu_cfg:                      # sacrebleu options
        remove_whitespace: True     # `remove_whitespace` option in sacrebleu.corpus_chrf() function (defalut: True)
        tokenize: "none"            # `tokenize` option in sacrebleu.corpus_bleu() function (options include: "none" (use for already tokenized test data), "13a" (default minimal tokenizer), "intl" which mostly does punctuation and unicode, etc) 

training:
    #load_model: "{ models/{name}_transformer/1.ckpt" # if uncommented, load a pre-trained model from this checkpoint
    random_seed: 42
    optimizer: "adam"
    normalization: "tokens"
    adam_betas: [0.9, 0.999] 
    scheduling: "plateau"           # TODO: try switching from plateau to Noam scheduling
    patience: 5                     # For plateau: decrease learning rate by decrease_factor if validation score has not improved for this many validation rounds.
    learning_rate_factor: 0.5       # factor for Noam scheduler (used with Transformer)
    learning_rate_warmup: 1000      # warmup steps for Noam scheduler (used with Transformer)
    decrease_factor: 0.7
    loss: "crossentropy"
    learning_rate: 0.0003
    learning_rate_min: 0.00000001
    weight_decay: 0.0
    label_smoothing: 0.1
    batch_size: 256
    batch_type: "token"
    eval_batch_size: 3600
    eval_batch_type: "sentence"
    batch_multiplier: 1
    early_stopping_metric: "ppl"
    epochs: 30                     # TODO: Decrease for when playing around and checking of working. Around 30 is sufficient to check if its working at all
    validation_freq: 1000          # TODO: Set to at least once per epoch.
    logging_freq: 100
    eval_metric: "bleu"
    model_dir: "models/en_kin_transformer"
    overwrite: False               # TODO: Set to True if you want to overwrite possibly existing models. 
    shuffle: True
    use_cuda: True
    max_output_length: 100
    print_valid_sents: [0, 1, 2, 3]
    keep_last_ckpts: 3

model:
    initializer: "xavier_normal"
    bias_initializer: "zeros"
    init_gain: 1.0
    embed_initializer: "xavier_normal"
    embed_init_gain: 1.0
    tied_embeddings: False
    tied_softmax: True
    encoder:
        type: "transformer"
        num_layers: 6
        num_heads: 8
        embeddings:
            embedding_dim: 256
            scale: True
            dropout: 0.
        # typically ff_size = 4 x hidden_size
        hidden_size: 256
        ff_size: 1024
        dropout: 0.1
        layer_norm: "post"
    decoder:
        type: "transformer"
        num_layers: 6
        num_heads: 8
        embeddings:
            embedding_dim: 256
            scale: True
            dropout: 0.
        # typically ff_size = 4 x hidden_size
        hidden_size: 256
        ff_size: 1024
        dropout: 0.1
        layer_norm: "post"