# network architecture
# encoder related
encoder: transformer
encoder_conf:
    output_size: 256    # dimension of attention
    attention_heads: 4
    linear_units: 2048  # the number of units of position-wise feed forward
    num_blocks: 12      # the number of encoder blocks
    dropout_rate: 0.1
    positional_dropout_rate: 0.1
    attention_dropout_rate: 0.0
    input_layer: conv2d # encoder architecture type
    normalize_before: true

# decoder related
decoder: transformer
decoder_conf:
    attention_heads: 4
    linear_units: 2048
    num_blocks: 6
    dropout_rate: 0.1
    positional_dropout_rate: 0.1
    self_attention_dropout_rate: 0.0
    src_attention_dropout_rate: 0.0

# hybrid CTC/attention
model_conf:
    ctc_weight: 0.3
    lsm_weight: 0.1     # label smoothing option
    length_normalized_loss: false

# minibatch related
batch_type: length
batch_bins: 32000
num_workers: 8

# optimization related
accum_grad: 1
grad_clip: 5
patience: 3
max_epoch: 20
val_scheduler_criterion:
    - valid
    - acc
best_model_criterion:
-   - valid
    - acc
    - max
keep_nbest_models: 10

# NoamLR is deprecated. Use WarmupLR.
# The following is equivalent setting for NoamLR:
#
#    optim: adam
#    optim_conf:
#        lr: 10.
#    scheduler: noamlr
#    scheduler_conf:
#        model_size: 256
#        warmup_steps: 25000
#
optim: adam
optim_conf:
    lr: 0.002
scheduler: warmuplr     # pytorch v1.1.0+ required
scheduler_conf:
    warmup_steps: 25000

log_interval: 50
normalize: None