UltraSharp/Config/UltraSharp Config.yml

name: UltraSharp
use_tb_logger: true
model: sr
scale: 4
gpu_ids: [0]
use_amp: true
use_swa: false
use_cem: false

# Dataset options:
datasets:
  train:
    name: UltraSharp
    mode: aligned
    dataroot_HR: [
      '!!!! CHANGE THIS OR COMMENT OUT',
    dataroot_LR: [
      '!!!! CHANGE THIS OR COMMENT OUT'
      ] # low resolution images
    subset_file: null
    use_shuffle: true
    znorm: false
    n_workers: 4
    batch_size: 4
    virtual_batch_size: 8
    preprocess: crop
    crop_size: 128
    image_channels: 3

    # Color space conversion
    # color: 'y'
    # color_LR: 'y'
    # color_HR: 'y'
    
    # LR and HR modifiers.
    # aug_downscale: 0.2
    # shape_change: reshape_lr
    
    # Enable random downscaling of HR images (will fix LR pair to correct size)
    # hr_downscale: true
    # hr_downscale_amt: [2, 1.75, 1.5, 1]
    # #pre_crop: true
    
    # Presets and on the fly (OTF) augmentations
    #augs_strategy: combo
    add_blur_preset: custom_blur
    add_resize_preset: custom_resize
    add_noise_preset: custom_noise
    aug_downscale: 0.2
    resize_strat: pre
    
    # On the fly generation of LR:
    dataroot_kernels: 'KERNEL PATH !!!! CHANGE THIS OR COMMENT OUT'
    lr_downscale: true
    lr_downscale_types: ["linear", "bicubic", "nearest_aligned"]

    # Rotations augmentations:
    use_flip: true
    use_rot: true
    use_hrrot: false
    
    # Noise and blur augmentations:
    lr_blur: false
    lr_blur_types: {sinc: 0.2, iso: 0.2, ansio2: 0.4, sinc2: 0.2, clean: 3}
    noise_data: 'NOISE PATH !!!! CHANGE THIS OR COMMENT OUT'
    lr_noise: false
    lr_noise_types: {camera: 0.1, jpeg: 0.8, clean: 3}
    lr_noise2: false
    lr_noise_types2: {jpeg: 1, webp: 0, clean: 2, camera: 2}
    hr_noise: false
    hr_noise_types:  {gaussian: 1, clean: 4}
    
    # Color augmentations
    # lr_fringes: false
    # lr_fringes_chance: 0.4
    # auto_levels: HR
    # rand_auto_levels: 0.7
    #lr_unsharp_mask: true
    #lr_rand_unsharp: 0.7
    # hr_unsharp_mask: true
    # hr_rand_unsharp: 1
    
    # Augmentations for classification or (maybe) inpainting networks:
    # lr_cutout: false
    # lr_erasing: false

  #val: 
    #name: val_set14_part
    #mode: aligned
    #dataroot_B: '../datasets/val/hr'
    #dataroot_A: '../datasets/val/lr'
    
    #znorm: false
    
    # Color space conversion:
    # color: 'y'
    # color_LR: 'y'
    # color_HR: 'y'
    
    lr_downscale: true
    lr_downscale_types: ["linear", "bicubic", "nearest_aligned"]

path:
    root: '../'
    pretrain_model_G: 'UniScale-Balanced was used originally !!!! CHANGE THIS OR COMMENT OUT'
    #pretrain_model_G: '../experiments/pretrained_models//.pth'
    # pretrain_model_D: '../experiments/pretrained_models/patchgan.pth'
    #resume_state: '!!!! CHANGE THIS OR COMMENT OUT'

# Generator options:
network_G:  # configurations for the Generator network
    which_model_G: RRDB_net  # check:  

# Discriminator options:
network_D: unet

train:
    # Optimizer options:
    optim_G: AdamP
    optim_D: AdamP
    
    # Schedulers options:
    lr_scheme: MultiStepLR
    lr_steps_rel: [0.1, 0.2, 0.4, 0.6]
    lr_gamma: 0.4

    # For SWA scheduler
    swa_start_iter_rel: 0.05
    swa_lr: 1e-4
    swa_anneal_epochs: 10
    swa_anneal_strategy: "cos"
    
    # Losses:
    pixel_criterion: clipl1  # pixel (content) loss
    pixel_weight: 0.12
    feature_criterion: l1 # feature loss (VGG feature network)
    feature_weight: 0.3
    cx_type: contextual  # contextual loss
    cx_weight: 0.25
    cx_vgg_layers: {conv_3_2: 1, conv_4_2: 1}
    #hfen_criterion: l1  # hfen
    #hfen_weight: 1e-6 
    #grad_type: grad-4d-l1  # image gradient loss
    #grad_weight: 4e-1
    # tv_type: normal  # total variation
    # tv_weight: 1e-5
    # tv_norm: 1
    ssim_type: ssim  # structural similarity
    ssim_weight: 0.05
    lpips_weight: 0.25  # [.25] perceptual loss
    lpips_type: net-lin
    lpips_net: squeeze
    
    # Experimental losses
    # spl_type: spl  # spatial profile loss
    # spl_weight: 0.1
    #of_type: overflow  # overflow loss
    #of_weight: 0.1
    # range_weight: 1  # range loss
    fft_type: fft  # FFT loss
    fft_weight: 0.3 #[.2]
    # color_criterion: color-l1cosinesim  # color consistency loss
    # color_weight: 1
    # avg_criterion: avg-l1  # averaging downscale loss
    # avg_weight: 5
    # ms_criterion: multiscale-l1  # multi-scale pixel loss
    # ms_weight: 1e-2
    #fdpl_type: fdpl  # frequency domain-based perceptual loss
    #fdpl_weight: 1e-3
    
    # Adversarial loss:
    gan_type: vanilla
    gan_weight: 4e-3
    # freeze_loc: 4
    # For wgan-gp:
    # D_update_ratio: 1
    # D_init_iters: 0
    # gp_weigth: 10
    # Feature matching (if using the discriminator_vgg_128_fea or discriminator_vgg_fea):
    # gan_featmaps: true
    # dis_feature_criterion: cb  # discriminator feature loss
    # dis_feature_weight: 0.01
    
    # For PPON:
    # p1_losses: [pix]
    # p2_losses: [pix-multiscale, ms-ssim]
    # p3_losses: [fea]
    # ppon_stages: [1000, 2000]
    
    # Differentiable Augmentation for Data-Efficient GAN Training
    # diffaug: true
    # dapolicy: 'color,transl_zoom,flip,rotate,cutout'
    
    # Batch (Mixup) augmentations
    mixup: true
    mixopts: [blend, rgb, mixup, cutmix, cutmixup] # , "cutout", "cutblur"]
    mixprob: [1, 1, 1.0, 1.0, 1.0] #, 1.0, 1.0]
    #mixalpha: [0.6, 1.0, 1.2, 0.7, 0.7] #, 0.001, 0.7]
    aux_mixprob: 1.0
    #aux_mixalpha: 1.2
    mix_p: 1.2
    
    # Frequency Separator
    # fs: true
    # lpf_type: average
    # hpf_type: average
    
    # Other training options:
    manual_seed: 0
    niter: 5e5
    # warmup_iter: -1
    #val_freq: 5e3
    # overwrite_val_imgs: true
    # val_comparison: true
    metrics: 'psnr,ssim,lpips'
    grad_clip: norm
    grad_clip_value: 0.1 # "auto"

logger:
    print_freq: 100
    save_checkpoint_freq: 800
    overwrite_chkp: false