# @package _global_

defaults:
  - override /mapping: characters
  - override /criterion: cross_entropy
  - override /callbacks: htr
  - override /datamodule: iam_lines
  - override /network: conv_transformer
  - override /model: lit_transformer
  - override /lr_schedulers: null
  - override /optimizers: null

epochs: &epochs 512
ignore_index: &ignore_index 3
num_classes: &num_classes 57
max_output_len: &max_output_len 89
summary: [[1, 1, 56, 1024], [1, 89]]

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

optimizers:
  radam:
    _target_: torch.optim.RAdam
    lr: 3.0e-4
    betas: [0.9, 0.999]
    weight_decay: 0
    eps: 1.0e-8
    parameters: network

lr_schedulers:
  network:
    _target_: torch.optim.lr_scheduler.ReduceLROnPlateau
    mode: min
    factor: 0.5
    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/loss

datamodule:
  batch_size: 16

network:
  input_dims: [1, 1, 56, 1024]
  num_classes: *num_classes
  pad_index: *ignore_index
  decoder:
    depth: 10
  pixel_embedding:
    shape: [7, 128]

model:
  max_output_len: *max_output_len

trainer:
  gradient_clip_val: 0.5
  max_epochs: *epochs