defaults:
  - override /mapping: null
  - override /network: null
  - override /model: null

mapping:
  _target_: text_recognizer.data.emnist_mapping.EmnistMapping
  extra_symbols: [ "\n" ]

datamodule:
  word_pieces: false
  batch_size: 8
  augment: false

criterion:
  ignore_index: 3

network:
  _target_: text_recognizer.networks.vq_transformer.VqTransformer
  input_dims: [1, 576, 640]
  encoder_dim: 32
  hidden_dim: 256
  dropout_rate: 0.1
  num_classes: 58
  pad_index: 3
  no_grad: true
  decoder:
    _target_: text_recognizer.networks.transformer.Decoder
    dim: 256
    depth: 2
    num_heads: 8
    attn_fn: text_recognizer.networks.transformer.attention.Attention
    attn_kwargs:
      dim_head: 32
      dropout_rate: 0.2
    norm_fn: torch.nn.LayerNorm
    ff_fn: text_recognizer.networks.transformer.mlp.FeedForward
    ff_kwargs:
      dim_out: null
      expansion_factor: 4
      glu: true
      dropout_rate: 0.2
    cross_attend: true
    pre_norm: true
    rotary_emb: null
  pretrained_encoder_path: "training/logs/runs/2021-09-25/23-07-28"

model:
  _target_: text_recognizer.models.vq_transformer.VqTransformerLitModel
  start_token: <s>
  end_token: <e>
  pad_token: <p>
  max_output_len: 682  # 451
  alpha: 1.0

trainer:
  max_epochs: 64
  limit_train_batches: 0.1
  limit_val_batches: 0.1