configs/bicx4_swinunet_lpips.yaml

trainer:
  target: trainer.TrainerDifIRLPIPS

autoencoder:
  target: ldm.models.autoencoder.VQModelTorch
  ckpt_path: weights/autoencoder/autoencoder_vq_f4.pth
  use_fp16: True
  params:
    embed_dim: 3
    n_embed: 8192
    ddconfig:
      double_z: False
      z_channels: 3
      resolution: 256
      in_channels: 3
      out_ch: 3
      ch: 128
      ch_mult:
      - 1
      - 2
      - 4
      num_res_blocks: 2
      attn_resolutions: []
      dropout: 0.0
      padding_mode: zeros

model:
  target: models.unet.UNetModelSwin
  ckpt_path: ~
  params:
    image_size: 64
    in_channels: 3
    model_channels: 160
    out_channels: 3
    attention_resolutions: [64,32,16,8]
    dropout: 0
    channel_mult: [1, 2, 2, 4]
    num_res_blocks: [2, 2, 2, 2]
    conv_resample: True
    dims: 2
    use_fp16: False
    num_head_channels: 32
    use_scale_shift_norm: True
    resblock_updown: False
    swin_depth: 2
    swin_embed_dim: 192
    window_size: 8
    mlp_ratio: 4
    cond_lq: True
    lq_size: 64

diffusion:
  target: models.script_util.create_gaussian_diffusion
  params:
    sf: 4
    schedule_name: exponential
    schedule_kwargs:
      power: 0.3
    etas_end: 0.99
    steps: 4
    min_noise_level: 0.2
    kappa: 2.0
    weighted_mse: False
    predict_type: xstart
    timestep_respacing: ~
    scale_factor: 1.0
    normalize_input: True
    latent_flag: True

data:
  train:
    type: bicubic
    params:
      source_path: ~ 
      source_txt_path: 
        - /mnt/sfs-common/zsyue/database/ImageNet/files_txt/path_train_all.txt
        - /mnt/sfs-common/zsyue/database/FFHQ/files_txt/files256.txt
      degrade_kwargs:
        scale: 0.25
        activate_matlab: True
        resize_back: False
        pch_size: 256
        pass_smallmaxresize: False
        pass_aug: False
        pass_crop: False
      transform_type: default
      transform_kwargs:
        mean: 0.5
        std: 0.5
      length: ~
      need_path: False
      im_exts: JPEG
      recursive: False
  val:
    type: bicubic
    params:
      source_path: /mnt/sfs-common/zsyue/projects/ResShift/SR/testingdata/imagenet256/gt
      degrade_kwargs:
        scale: 0.25
        activate_matlab: True
        resize_back: False
        pch_size: 256
        pass_smallmaxresize: True
        pass_aug: True
        pass_crop: True
      transform_type: default
      transform_kwargs:
        mean: 0.5
        std: 0.5
      length: 64
      need_path: False
      im_exts: png
      recursive: False

train:
  # learning rate
  lr: 5e-5                      # learning rate 
  lr_min: 2e-5                      # learning rate 
  lr_schedule: cosin
  warmup_iterations: 5000
  # dataloader
  batch: [96, 8]                
  microbatch: 12
  num_workers: 6
  prefetch_factor: 2            
  # optimization settings
  weight_decay: 0               
  ema_rate: 0.999
  iterations: 400000            # total iterations
  # save logging
  save_freq: 10000              
  log_freq: [1000, 2000, 1]         # [training loss, training images, val images]
  loss_coef: [1.0, 1.0]         # [mse, lpips]
  local_logging: True           # manually save images
  tf_logging: False             # tensorboard logging
  # validation settings
  use_ema_val: True            
  val_freq: ${train.save_freq}
  val_y_channel: True
  val_resolution: ${model.params.lq_size}
  val_padding_mode: reflect
  # training setting
  use_amp: True                # amp training
  seed: 123456                 # random seed
  global_seeding: False
  # model compile
  compile:
    flag: True
    mode: reduce-overhead      # default, reduce-overhead