NTED/fashion_512.yaml

distributed: True
image_to_tensorboard: True
snapshot_save_iter: 50000
snapshot_save_epoch: 20
snapshot_save_start_iter: 20000
snapshot_save_start_epoch: 100
image_save_iter: 1000
max_epoch: 400
logging_iter: 100
amp: False

gen_optimizer:
    type: adam
    lr: 0.002
    adam_beta1: 0.
    adam_beta2: 0.99
    lr_policy:
        iteration_mode: False
        type: step
        step_size: 1000000
        gamma: 1

dis_optimizer:
    type: adam
    lr: 0.001882
    adam_beta1: 0.
    adam_beta2: 0.9905
    lr_policy:
        iteration_mode: False
        type: step
        step_size: 1000000
        gamma: 1


trainer:
    type: NTED.extraction_distribution_trainer::Trainer
    gan_mode: style_gan2
    gan_start_iteration: 1000  # 0
    face_crop_method: util.face_crop::crop_face_from_output
    hand_crop_method: util.face_crop::crop_hands_from_output
    d_reg_every: 16
    r1: 10
    loss_weight:
      weight_perceptual: 1
      weight_gan: 1.5
      weight_attn_rec: 15
      weight_face: 1
      weight_hand: 1
      weight_l1: 1
      weight_l1_hand: 0.8
      weight_edge: 100
    attn_weights: 
      8: 1 
      16: 1
      32: 1
      64: 1
      128: 1
      256: 1
    vgg_param:
      network: vgg19
      layers: ['relu_1_1', 'relu_2_1', 'relu_3_1', 'relu_4_1', 'relu_5_1']
      num_scales: 3
      use_style_loss: True
      style_to_perceptual: 1000
    vgg_hand_param:
      network: vgg19
      layers: ['relu_1_1', 'relu_2_1', 'relu_3_1','relu_3_3', 'relu_4_1', 'relu_4_3', 'relu_5_1']

gen:
    type: NTED.extraction_distribution_model::Generator
    param:
      size: 512
      semantic_dim: 20
      channels:
        16: 512
        32: 512
        64: 512
        128: 256
        256: 128
        512: 64
        1024: 32
      num_labels:
        16: 16
        32: 32
        64: 32
        128: 64
        256: 64
        512: False
      match_kernels:
        16: 1
        32: 3
        64: 3
        128: 3
        256: 3
        512: False
                
dis:
    type: generators.discriminator::Discriminator
    param:
      size: 512
      channels:      
        4: 512
        8: 512
        16: 512
        32: 512
        64: 512
        128: 256
        256: 128  
        512: 64     
      is_square_image: False


data:
    type: data.fashion_data::Dataset
    preprocess_mode: resize_and_crop  # resize_and_crop
    path: /apdcephfs/share_1474453/zejunzhang/dataset/pose_trans_dataset_2d
    num_workers: 16
    sub_path: 512-352
    resolution: 512
    scale_param: 0.1
    train:
      batch_size: 4   # real_batch_size: 2 * 2 (source-->target & target --> source) * 4 (GPUs) = 16
      distributed: True
    val:
      batch_size: 4
      distributed: True
    hand_keypoint: False