Restormer/Chained/Chained_Restormer.yml

# general settings
name: Chained_Restormer
model_type: ChainedImageCleanModel
scale: 1
num_gpu: 2  # set num_gpu: 0 for cpu mode
manual_seed: 100

# dataset and data loader settings
datasets:
  train:
    name: TrainSet
    type: Dataset_MixedPairedImage
    dataroot_gt: /hdd/Restoration/data/MiO_train/HQ
    dataroot_lq:
      - /hdd/Restoration/data/MiO_train/LQ/d2/rain+noise
      - /hdd/Restoration/data/MiO_train/LQ/d2/defocusblur+noise
      - /hdd/Restoration/data/MiO_train/LQ/d2/noise+defocusblur
      - /hdd/Restoration/data/MiO_train/LQ/d2/defocusblur+rain
      - /hdd/Restoration/data/MiO_train/LQ/d2/rain+defocusblur
      - /hdd/Restoration/data/MiO_train/LQ/d1/rain
      - /hdd/Restoration/data/MiO_train/LQ/d1/noise
      - /hdd/Restoration/data/MiO_train/LQ/d1/defocusblur
    datalq_mask:
      - '210'
      - '201'
      - '102'
      - '021'
      - '012'
      - '010'
      - '100'
      - '001'
    geometric_augs: true

    filename_tmpl: '{}'
    io_backend:
      type: disk

    # data loader
    use_shuffle: true
    num_worker_per_gpu: 8
    batch_size_per_gpu: 8

    ### -------------Progressive training--------------------------
    # mini_batch_sizes: [8,5,4,2,1,1]             # Batch size per gpu   
    # iters: [92000,64000,48000,36000,36000,24000]
    # gt_size: 384   # Max patch size for progressive training
    # gt_sizes: [128,160,192,256,320,384]  # Patch sizes for progressive training.
    ### ------------------------------------------------------------

    ### ------- Training on single fixed-patch size 128x128---------
    mini_batch_sizes: [2]   
    iters: [10000]
    gt_size: 128   
    gt_sizes: [128]
    ### ------------------------------------------------------------

    dataset_enlarge_ratio: 1
    prefetch_mode: ~

  val:
    name: ValSet
    type: Dataset_MixedPairedImage
    dataroot_gt: /hdd/Restoration/data/MiO_test/original/HQ
    dataroot_lq:
      - /hdd/Restoration/data/MiO_test/original/LQ/d2/rain+noise
    datalq_mask:
      - '110'

    io_backend:
      type: disk

# network structures
network_g:
  type: ChainedRestormer
  opts:
    # Denoise
    - type: Restormer
      inp_channels: 3
      out_channels: 3
      dim: 48
      num_blocks: [4,6,6,8]
      num_refinement_blocks: 4
      heads: [1,2,4,8]
      ffn_expansion_factor: 2.66
      bias: False
      LayerNorm_type: BiasFree
      dual_pixel_task: False
    # Derain
    - type: Restormer
      inp_channels: 3
      out_channels: 3
      dim: 48
      num_blocks: [4,6,6,8]
      num_refinement_blocks: 4
      heads: [1,2,4,8]
      ffn_expansion_factor: 2.66
      bias: False
      LayerNorm_type: WithBias
      dual_pixel_task: False
    # Defocus Blur
    - type: Restormer
      inp_channels: 3
      out_channels: 3
      dim: 48
      num_blocks: [4,6,6,8]
      num_refinement_blocks: 4
      heads: [1,2,4,8]
      ffn_expansion_factor: 2.66
      bias: False
      LayerNorm_type: WithBias
      dual_pixel_task: False


# path
path:
  pretrain_network_g: 
    - /hdd/Restoration/Restormer/Denoising/pretrained_models/gaussian_color_denoising_sigma25.pth
    - /hdd/Restoration/Restormer/Deraining/pretrained_models/deraining.pth
    - /hdd/Restoration/Restormer/Defocus_Deblurring/pretrained_models/single_image_defocus_deblurring.pth
  strict_load_g: true
  resume_state: ~
  
# training settings
train:
  total_iter: 10000
  warmup_iter: 500 # no warm up
  use_grad_clip: true

  # Split 300k iterations into two cycles. 
  # 1st cycle: fixed 3e-4 LR for 92k iters. 
  # 2nd cycle: cosine annealing (3e-4 to 1e-6) for 208k iters.
  scheduler:
    type: CosineAnnealingRestartCyclicLR
    periods: [2000, 8000]       
    restart_weights: [1,1]
    eta_mins: [0.0001,0.000001]   
  
  mixing_augs:
    mixup: false
    mixup_beta: 1.2
    use_identity: true

  optim_g:
    type: AdamW
    lr: !!float 1e-4
    weight_decay: !!float 1e-4
    betas: [0.9, 0.999]
  
  # losses
  pixel_opt:
    type: L1Loss
    loss_weight: 1
    reduction: mean

# validation settings
val:
  window_size: 8
  val_freq: 100000
  save_img: false
  rgb2bgr: true
  use_image: true
  max_minibatch: 8

  metrics:
    psnr: # metric name, can be arbitrary
      type: calculate_psnr
      crop_border: 0
      test_y_channel: true

# logging settings
logger:
  log_file: /hdd/Restoration/Restormer/trainings/
  print_freq: 200
  save_checkpoint_freq: !!float 4e3
  use_tb_logger: true
  wandb:
    project: ~
    resume_id: ~

# dist training settings
dist_params:
  backend: nccl
  port: 29500