train_hifacegan.yml

# general settings
name: HiFaceGAN_SR4x_train
model_type: HiFaceGANModel
scale: 1    # HiFaceGAN does not resize lq input
num_gpu: 1  # set num_gpu: 0 for cpu mode
manual_seed: 0

datasets:
  train:
    name: FFHQ_sr4x
    type: PairedImageDataset
    dataroot_gt: datasets/FFHQ_512_gt
    dataroot_lq: datasets/FFHQ_512_lq_sr4x
    # (for lmdb)
    # dataroot_gt: datasets/FFHQ_512_gt.lmdb
    # dataroot_lq: datasets/FFHQ_512_lq_sr4x.lmdb
    io_backend:
      type: disk
      # (for lmdb)
      # type: lmdb

    gt_size: 512
    use_hflip: true
    use_rot: false

    # data loader
    num_worker_per_gpu: 1
    batch_size_per_gpu: 1
    dataset_enlarge_ratio: 1
    prefetch_mode: ~

  val:
    name: FFHQ_sr4x_val   # For now, use training dataset for validation
    type: PairedImageDataset
    dataroot_gt: datasets/FFHQ_512_gt
    dataroot_lq: datasets/FFHQ_512_lq_sr4x
    io_backend:
      type: disk

# network structures
network_g:
  type: HiFaceGAN
  num_in_ch: 3
  num_feat: 48
  use_vae: false
  z_dim: 256  # dummy var
  crop_size: 512
  #norm_g: 'spectralspadesyncbatch3x3'
  #norm_g: 'spectralspadeinstance3x3'
  norm_g: 'spectralspadebatch3x3'  # 20210519: Use batchnorm for now.
  is_train: false                  # HifaceGAN supports progressive training
                                   # so network architecture depends on it

network_d:
  type: HiFaceGANDiscriminator
  num_in_ch: 3
  num_out_ch: 3
  conditional_d: True
  num_feat: 64
  norm_d: 'spectralinstance'


# training settings
train:
  optim_g:
    type: Adam
    lr: !!float 2e-4
    weight_decay: 0
    betas: [0.9, 0.999]
  optim_d:
    type: Adam
    lr: !!float 2e-4
    weight_decay: 0
    betas: [0.9, 0.999]

  scheduler:
    # TODO: Integrate the exact scheduling system of HiFaceGAN
    # which involves a fixed lr followed by a linear decay
    # It is not supported in current BasicSR project.
    type: MultiStepLR
    milestones: [500, 1000, 2000, 3000]
    gamma: 0.5

  # By default HiFaceGAN trains for 50 epochs, need auto conversion with
  #    total_iter = #(epochs) * #(dataset_size) * enlarge_ratio / batch_size
  total_iter: 5000
  warmup_iter: -1  # no warm up

  # losses:
  #   Note: HiFaceGAN does not need pixel loss, use it at your own discretion
  # pixel_opt:
  #  type: L1Loss
  #  loss_weight: !!float 1e-2
  #  reduction: mean
  perceptual_opt:
    type: PerceptualLoss
    # vgg_layer_indices: 2,7,12,21,30
    # weights: 1/32, 1/16, 1/8, 1/4, 1
    layer_weights:
      'relu1_1': !!float 3.125e-2
      'relu2_1': !!float 6.25e-2
      'relu3_1': !!float 0.125
      'relu4_1': !!float 0.25
      'relu5_1': !!float 1.0
    vgg_type: vgg19
    use_input_norm: false # keep in [0,1] range
    range_norm: false
    perceptual_weight: !!float 10.0
    style_weight: 0
    criterion: l1
  gan_opt:
    type: MultiScaleGANLoss
    gan_type: lsgan
    real_label_val: 1.0
    fake_label_val: 0.0
    loss_weight: !!float 1.0
  feature_matching_opt:
    type: GANFeatLoss
    loss_weight: !!float 10.0
    criterion: l1

  net_d_iters: 1
  net_d_init_iters: 0

# path
path:
  pretrain_network_g: ~ # experiments/pretrained_models/4xsr/latest_net_G.pth
  strict_load_g: true

# validation settings
val:
  val_freq: !!float 5e3
  save_img: true

  metrics:
    psnr: # metric name, can be arbitrary
      type: calculate_psnr
      crop_border: 4
      test_y_channel: false

# logging settings
logger:
  print_freq: 100
  save_checkpoint_freq: !!float 5e3
  use_tb_logger: true
  wandb:
    project: ~
    resume_id: ~