name: TAFG-vox2
engine: TAFG
result_dir: ./result
max_pairs: 1000000

handler:
  clear_cuda_cache: True
  set_epoch_for_dist_sampler: True
  checkpoint:
    epoch_interval: 1 # checkpoint once per `epoch_interval` epoch
    n_saved: 2
  tensorboard:
    scalar: 100 # log scalar `scalar` times per epoch
    image: 4 # log image `image` times per epoch
  test:
    random: True
    images: 10


misc:
  random_seed: 1004
  add_new_loss_epoch: -1

model:
  generator:
    _type: TAFG-Generator
    _bn_to_sync_bn: False
    style_in_channels: 3
    content_in_channels: 24
    use_spectral_norm: False
    style_encoder_type: StyleEncoder
    num_style_conv: 4
    style_dim: 8
    num_adain_blocks: 4
    num_res_blocks: 4
    
  discriminator:
    _type: MultiScaleDiscriminator
    num_scale: 2
    discriminator_cfg:
      _type: PatchDiscriminator
      in_channels: 3
      base_channels: 64
      use_spectral: True
      need_intermediate_feature: True

loss:
  gan:
    loss_type: hinge
    real_label_val: 1.0
    fake_label_val: 0.0
    weight: 1.0
  perceptual:
    layer_weights:
      "1": 0.03125
      "6": 0.0625
      "11": 0.125
      "20": 0.25
      "29": 1
    criterion: 'L1'
    style_loss: False
    perceptual_loss: True
    weight: 0
  style:
    layer_weights:
      "3": 1
    criterion: 'L1'
    style_loss: True
    perceptual_loss: False
    weight: 10
  recon:
    level: 1
    weight: 10
  style_recon:
    level: 1
    weight: 1
  content_recon:
    level: 1
    weight: 1
  edge:
    weight: 5
    hed_pretrained_model_path: ./network-bsds500.pytorch
  cycle:
    level: 1
    weight: 10

optimizers:
  generator:
    _type: Adam
    lr: 0.0001
    betas: [ 0, 0.9 ]
    weight_decay: 0.0001
  discriminator:
    _type: Adam
    lr: 4e-4
    betas: [ 0, 0.9 ]
    weight_decay: 0.0001

data:
  train:
    scheduler:
      start_proportion: 0.5
      target_lr: 0
    buffer_size: 50
    dataloader:
      batch_size: 8
      shuffle: True
      num_workers: 1
      pin_memory: True
      drop_last: True
    dataset:
      _type: GenerationUnpairedDatasetWithEdge
      root_a: "/data/i2i/VoxCeleb2Anime/trainA"
      root_b: "/data/i2i/VoxCeleb2Anime/trainB"
      edges_path: "/data/i2i/VoxCeleb2Anime/edges"
      landmarks_path: "/data/i2i/VoxCeleb2Anime/landmarks"
      edge_type: "landmark_hed"
      size: [ 128, 128 ]
      random_pair: True
      pipeline:
        - Load
        - Resize:
            size: [ 128, 128 ]
        - ToTensor
        - Normalize:
            mean: [ 0.5, 0.5, 0.5 ]
            std: [ 0.5, 0.5, 0.5 ]
  test:
    which: video_dataset
    dataloader:
      batch_size: 1
      shuffle: False
      num_workers: 1
      pin_memory: False
      drop_last: False
    dataset:
      _type: GenerationUnpairedDatasetWithEdge
      root_a: "/data/i2i/VoxCeleb2Anime/testA"
      root_b: "/data/i2i/VoxCeleb2Anime/testB"
      edges_path: "/data/i2i/VoxCeleb2Anime/edges"
      landmarks_path: "/data/i2i/VoxCeleb2Anime/landmarks"
      edge_type: "landmark_hed"
      random_pair: False
      size: [ 128, 128 ]
      pipeline:
        - Load
        - Resize:
            size: [ 128, 128 ]
        - ToTensor
        - Normalize:
            mean: [ 0.5, 0.5, 0.5 ]
            std: [ 0.5, 0.5, 0.5 ]
    video_dataset:
      _type: SingleFolderDataset
      root: "/data/i2i/VoxCeleb2Anime/test_video_frames/"
      with_path: True
      pipeline:
        - Load
        - Resize:
            size: [ 128, 128 ]
        - ToTensor
        - Normalize:
            mean: [ 0.5, 0.5, 0.5 ]
            std: [ 0.5, 0.5, 0.5 ]