defaults:
  - override /network: null
  - override /criterion: null
  - override /datamodule: null
  - override /model: lit_vqgan
  - override /callbacks: wandb_vae
  - override /optimizers: null
  - override /lr_schedulers: null

criterion:
  _target_: text_recognizer.criterions.vqgan_loss.VQGANLoss
  reconstruction_loss:
    _target_: torch.nn.BCEWithLogitsLoss
    reduction: mean
  discriminator:
    _target_: text_recognizer.criterions.n_layer_discriminator.NLayerDiscriminator
    in_channels: 1
    num_channels: 64
    num_layers: 3
  commitment_weight: 0.25
  discriminator_weight: 0.8
  discriminator_factor: 1.0
  discriminator_iter_start: 1.5e4

datamodule:
  _target_: text_recognizer.data.iam_lines.IAMLines
  batch_size: 24
  num_workers: 12
  train_fraction: 0.8
  augment: true
  pin_memory: false

lr_schedulers:
  generator:
    _target_: torch.optim.lr_scheduler.CosineAnnealingLR
    T_max: 64
    eta_min: 4.5e-6
    last_epoch: -1
    interval: epoch
    monitor: val/loss
#   discriminator:
#     _target_: torch.optim.lr_scheduler.CosineAnnealingLR
#     T_max: 64
#     eta_min: 0.0
#     last_epoch: -1
# 
#     interval: epoch
#     monitor: val/loss

optimizers:
  generator:
    _target_: madgrad.MADGRAD
    lr: 1.0e-4
    momentum: 0.5
    weight_decay: 0
    eps: 1.0e-7
    parameters: network

  discriminator:
    _target_: madgrad.MADGRAD
    lr: 4.5e-6
    momentum: 0.5
    weight_decay: 0
    eps: 1.0e-6
    parameters: loss_fn.discriminator

network:
  _target_: text_recognizer.networks.vqvae.vqvae.VQVAE
  hidden_dim: 256
  embedding_dim: 32
  num_embeddings: 512
  decay: 0.99
  encoder:
    _target_: text_recognizer.networks.vqvae.encoder.Encoder
    in_channels: 1 
    hidden_dim: 32
    channels_multipliers: [1, 4, 8]
    dropout_rate: 0.0
    activation: mish
    use_norm: true
    num_residuals: 2
    residual_channels: 32
  decoder:
    _target_: text_recognizer.networks.vqvae.decoder.Decoder
    out_channels: 1 
    hidden_dim: 32
    channels_multipliers: [8, 4, 1]
    dropout_rate: 0.0
    activation: mish
    use_norm: true
    num_residuals: 2 
    residual_channels: 32

trainer:
  max_epochs: 64
  # limit_train_batches: 0.1
  # limit_val_batches: 0.1
  # gradient_clip_val: 100

# tune: false
# train: true
# test: false
summary: [2, 1, 56, 1024]