# @package _global_

defaults:
  - override /criterion: cross_entropy
  - override /callbacks: htr
  - override /datamodule: iam_lines
  - override /network: conv_perceiver
  - override /model: lit_perceiver
  - override /lr_scheduler: null
  - override /optimizer: null

tags: [lines, perceiver]
epochs: &epochs 260
ignore_index: &ignore_index 3
num_classes: &num_classes 57
max_output_len: &max_output_len 89
summary: [[1, 1, 56, 1024]]

logger:
  wandb:
    tags: ${tags}

criterion:
  ignore_index: *ignore_index
  # label_smoothing: 0.1

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

optimizer:
  _target_: adan_pytorch.Adan
  lr: 1.0e-4
  betas: [0.02, 0.08, 0.01]
  weight_decay: 0.02

lr_scheduler:
  _target_: torch.optim.lr_scheduler.ReduceLROnPlateau
  mode: min
  factor: 0.8
  patience: 10
  threshold: 1.0e-4
  threshold_mode: rel
  cooldown: 0
  min_lr: 1.0e-5
  eps: 1.0e-8
  verbose: false
  interval: epoch
  monitor: val/cer

datamodule:
  batch_size: 8
  train_fraction: 0.95

network:
  input_dims: [1, 1, 56, 1024]
  num_classes: *num_classes
  pad_index: *ignore_index
  encoder:
    depth: 5
  decoder:
    depth: 6

model:
  max_output_len: *max_output_len

trainer:
  gradient_clip_val: 1.0
  stochastic_weight_avg: true
  max_epochs: *epochs
  accumulate_grad_batches: 1