# @package _global_

defaults:
  - override /mapping: null
  - override /criterion: null
  - override /callbacks: htr
  - override /datamodule: iam_extended_paragraphs
  - override /network: null
  - override /model: null
  - override /lr_schedulers: null
  - override /optimizers: null


epochs: &epochs 720
ignore_index: &ignore_index 3
num_classes: &num_classes 58
max_output_len: &max_output_len 682
summary: [[1, 1, 576, 640], [1, 682]]

criterion:
  _target_: torch.nn.CrossEntropyLoss
  ignore_index: *ignore_index
    
mapping: &mapping
  mapping:
    _target_: text_recognizer.data.mappings.emnist.EmnistMapping
    extra_symbols: [ "\n" ]

callbacks:
  stochastic_weight_averaging:
    _target_: pytorch_lightning.callbacks.StochasticWeightAveraging
    swa_epoch_start: 0.75
    swa_lrs: 1.0e-5
    annealing_epochs: 10
    annealing_strategy: cos
    device: null

optimizers:
  madgrad:
    _target_: madgrad.MADGRAD
    lr: 1.0e-4
    momentum: 0.9
    weight_decay: 5.0e-6
    eps: 1.0e-6
    parameters: network

lr_schedulers:
  network:
    _target_: torch.optim.lr_scheduler.OneCycleLR
    max_lr: 1.0e-4
    total_steps: null
    epochs: *epochs
    steps_per_epoch: 211
    pct_start: 0.03
    anneal_strategy: cos
    cycle_momentum: true
    base_momentum: 0.85
    max_momentum: 0.95
    div_factor: 25
    final_div_factor: 1.0e2
    three_phase: false
    last_epoch: -1
    verbose: false
    interval: step
    monitor: val/loss

datamodule:
  _target_: text_recognizer.data.iam_extended_paragraphs.IAMExtendedParagraphs
  batch_size: 6
  num_workers: 12
  train_fraction: 0.8
  pin_memory: true
  << : *mapping

network:
  _target_: text_recognizer.networks.conv_transformer.ConvTransformer
  input_dims: [1, 576, 640]
  hidden_dim: &hidden_dim 192
  encoder_dim: 1280
  dropout_rate: 0.1
  num_classes: *num_classes
  pad_index: *ignore_index
  encoder:
    _target_: text_recognizer.networks.encoders.efficientnet.EfficientNet
    arch: b0
    out_channels: 1280
    stochastic_dropout_rate: 0.2
    bn_momentum: 0.99
    bn_eps: 1.0e-3
  decoder:
    _target_: text_recognizer.networks.transformer.layers.Decoder
    dim: *hidden_dim
    depth: 3
    num_heads: 4
    attn_fn: text_recognizer.networks.transformer.attention.Attention
    attn_kwargs:
      dim_head: 32
      dropout_rate: 0.05
    local_attn_fn: text_recognizer.networks.transformer.local_attention.LocalAttention
    local_attn_kwargs:
      dim_head: 32
      dropout_rate: 0.05
      window_size: 11
      look_back: 2
      depth: 2
    norm_fn: text_recognizer.networks.transformer.norm.ScaleNorm
    ff_fn: text_recognizer.networks.transformer.mlp.FeedForward
    ff_kwargs:
      dim_out: null
      expansion_factor: 4
      glu: true
      dropout_rate: 0.05
    cross_attend: true
    pre_norm: true
    rotary_emb:
      _target_: text_recognizer.networks.transformer.embeddings.rotary.RotaryEmbedding
      dim: 32
  pixel_pos_embedding:
    _target_: text_recognizer.networks.transformer.embeddings.axial.AxialPositionalEmbedding
    dim: *hidden_dim 
    shape: [18, 20]
  token_pos_embedding:
    _target_: text_recognizer.networks.transformer.embeddings.fourier.PositionalEncoding
    hidden_dim: *hidden_dim 
    dropout_rate: 0.05
    max_len: *max_output_len

model:
  _target_: text_recognizer.models.transformer.TransformerLitModel
  << : *mapping
  max_output_len: *max_output_len
  start_token: <s>
  end_token: <e>
  pad_token: <p>

trainer:
  _target_: pytorch_lightning.Trainer
  stochastic_weight_avg: true
  auto_scale_batch_size: binsearch
  auto_lr_find: false
  gradient_clip_val: 0.5
  fast_dev_run: false
  gpus: 1
  precision: 16
  max_epochs: *epochs
  terminate_on_nan: true
  weights_summary: null
  limit_train_batches: 1.0 
  limit_val_batches: 1.0
  limit_test_batches: 1.0
  resume_from_checkpoint: null
  accumulate_grad_batches: 16
  overfit_batches: 0