idmvton

BasicSR/options/train/HiFaceGAN/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: ~

BasicSR/options/train/LDL/train_LDL_Real_x4.yml

+# general settings
+name: train_LDL_realworld_RRDB
+model_type: RealESRGANModel
+scale: 4
+num_gpu: 4
+manual_seed: 0
+
+# ----------------- options for synthesizing training data in RealESRGANModel ----------------- #
+# USM the ground-truth
+l1_gt_usm: True
+percep_gt_usm: True
+gan_gt_usm: False
+
+# the first degradation process
+resize_prob: [0.2, 0.7, 0.1]  # up, down, keep
+resize_range: [0.15, 1.5]
+gaussian_noise_prob: 0.5
+noise_range: [1, 30]
+poisson_scale_range: [0.05, 3]
+gray_noise_prob: 0.4
+jpeg_range: [30, 95]
+
+# the second degradation process
+second_blur_prob: 0.8
+resize_prob2: [0.3, 0.4, 0.3]  # up, down, keep
+resize_range2: [0.3, 1.2]
+gaussian_noise_prob2: 0.5
+noise_range2: [1, 25]
+poisson_scale_range2: [0.05, 2.5]
+gray_noise_prob2: 0.4
+jpeg_range2: [30, 95]
+
+gt_size: 256
+queue_size: 180
+
+# dataset and data loader settings
+datasets:
+  train:
+    name: DF2K+OST
+    type: RealESRGANDataset
+    dataroot_gt: datasets/DF2K
+    meta_info: datasets/DF2K/meta_info/meta_info_DF2Kmultiscale+OST_sub.txt
+    io_backend:
+      type: disk
+
+    blur_kernel_size: 21
+    kernel_list: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    sinc_prob: 0.1
+    blur_sigma: [0.2, 3]
+    betag_range: [0.5, 4]
+    betap_range: [1, 2]
+
+    blur_kernel_size2: 21
+    kernel_list2: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob2: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    sinc_prob2: 0.1
+    blur_sigma2: [0.2, 1.5]
+    betag_range2: [0.5, 4]
+    betap_range2: [1, 2]
+
+    final_sinc_prob: 0.8
+
+    gt_size: 256
+    use_hflip: True
+    use_rot: False
+
+    # data loader
+    num_worker_per_gpu: 4
+    batch_size_per_gpu: 4
+    dataset_enlarge_ratio: 1
+    prefetch_mode: ~
+
+  val:
+    name: RealWorld38
+    type: SingleImageDataset
+    dataroot_lq: datasets/RealWorld38/LR
+    io_backend:
+      type: disk
+
+# network structures
+network_g:
+  type: RRDBNet
+  num_in_ch: 3
+  num_out_ch: 3
+  num_feat: 64
+  num_block: 23
+  num_grow_ch: 32
+
+
+network_d:
+  type: UNetDiscriminatorSN
+  num_in_ch: 3
+  num_feat: 64
+  skip_connection: True
+
+# path
+path:
+  # use the pre-trained Real-ESRNet model
+  pretrain_network_g: experiments/pretrained_models/RealESRGAN/RealESRNet_x4plus.pth
+  param_key_g: params_ema
+  strict_load_g: true
+  resume_state: ~
+
+# training settings
+train:
+  ema_decay: 0.999
+  optim_g:
+    type: Adam
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+  optim_d:
+    type: Adam
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+
+  scheduler:
+    type: MultiStepLR
+    milestones: [400000]
+    gamma: 0.5
+
+  total_iter: 400000
+  warmup_iter: -1  # no warm up
+
+  # losses
+  pixel_opt:
+    type: L1Loss
+    loss_weight: !!float 1e-2
+    reduction: mean
+  ldl_opt:
+    type: L1Loss
+    loss_weight: !!float 1.0
+    reduction: mean
+  # perceptual loss (content and style losses)
+  perceptual_opt:
+    type: PerceptualLoss
+    layer_weights:
+      # before relu
+      'conv1_2': 0.1
+      'conv2_2': 0.1
+      'conv3_4': 1
+      'conv4_4': 1
+      'conv5_4': 1
+    vgg_type: vgg19
+    use_input_norm: true
+    perceptual_weight: !!float 1.0
+    style_weight: 0
+    range_norm: false
+    criterion: l1
+  # gan loss
+  gan_opt:
+    type: GANLoss
+    gan_type: vanilla
+    real_label_val: 1.0
+    fake_label_val: 0.0
+    loss_weight: !!float 1e-1
+
+  net_d_iters: 1
+  net_d_init_iters: 0
+
+# Uncomment these for validation
+# validation settings
+val:
+ val_freq: !!float 5e3
+ save_img: True
+
+# logging settings
+logger:
+  print_freq: 100
+  save_checkpoint_freq: !!float 5e3
+  use_tb_logger: true
+  wandb:
+    project: ~
+    resume_id: ~
+
+# dist training settings
+dist_params:
+  backend: nccl
+  port: 29500

BasicSR/options/train/RCAN/train_RCAN_x2.yml

+# TODO
+# general settings
+name: 201_RCANx2_scratch_DIV2K_rand0
+model_type: SRModel
+scale: 2
+num_gpu: 1  # set num_gpu: 0 for cpu mode
+manual_seed: 10
+
+# dataset and data loader settings
+datasets:
+  train:
+    name: DIV2K
+    type: PairedImageDataset
+    dataroot_gt: datasets/DIV2K/DIV2K_train_HR_sub
+    dataroot_lq: datasets/DIV2K/DIV2K_train_LR_bicubic/X2_sub
+    # (for lmdb)
+    # dataroot_gt: datasets/DIV2K/DIV2K_train_HR_sub.lmdb
+    # dataroot_lq: datasets/DIV2K/DIV2K_train_LR_bicubic_X2_sub.lmdb
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+      # (for lmdb)
+      # type: lmdb
+
+    gt_size: 96
+    use_hflip: true
+    use_rot: true
+
+    # data loader
+    num_worker_per_gpu: 6
+    batch_size_per_gpu: 16
+    dataset_enlarge_ratio: 100
+    prefetch_mode: ~
+
+  val:
+    name: Set5
+    type: PairedImageDataset
+    dataroot_gt: datasets/Set5/GTmod12
+    dataroot_lq: datasets/Set5/LRbicx2
+    io_backend:
+      type: disk
+
+# network structures
+network_g:
+  type: RCAN
+  num_in_ch: 3
+  num_out_ch: 3
+  num_feat: 64
+  num_group: 10
+  num_block: 20
+  squeeze_factor: 16
+  upscale: 2
+  res_scale: 1
+  img_range: 255.
+  rgb_mean: [0.4488, 0.4371, 0.4040]
+
+# path
+path:
+  pretrain_network_g: ~
+  strict_load_g: true
+  resume_state: ~
+
+# training settings
+train:
+  ema_decay: 0.999
+  optim_g:
+    type: Adam
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+
+  scheduler:
+    type: MultiStepLR
+    milestones: [200000]
+    gamma: 0.5
+
+  total_iter: 300000
+  warmup_iter: -1  # no warm up
+
+  # losses
+  pixel_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+
+# validation settings
+val:
+  val_freq: !!float 5e3
+  save_img: false
+
+  metrics:
+    psnr: # metric name, can be arbitrary
+      type: calculate_psnr
+      crop_border: 2
+      test_y_channel: false
+
+# logging settings
+logger:
+  print_freq: 100
+  save_checkpoint_freq: !!float 5e3
+  use_tb_logger: true
+  wandb:
+    project: ~
+    resume_id: ~
+
+# dist training settings
+dist_params:
+  backend: nccl
+  port: 29500

BasicSR/options/train/RealESRGAN/train_realesrgan_x2plus.yml

+# general settings
+name: train_RealESRGANx2plus_400k_B12G4
+model_type: RealESRGANModel
+scale: 2
+num_gpu: auto  # auto: can infer from your visible devices automatically. official: 4 GPUs
+manual_seed: 0
+
+# ----------------- options for synthesizing training data in RealESRGANModel ----------------- #
+# USM the ground-truth
+l1_gt_usm: True
+percep_gt_usm: True
+gan_gt_usm: False
+
+# the first degradation process
+resize_prob: [0.2, 0.7, 0.1]  # up, down, keep
+resize_range: [0.15, 1.5]
+gaussian_noise_prob: 0.5
+noise_range: [1, 30]
+poisson_scale_range: [0.05, 3]
+gray_noise_prob: 0.4
+jpeg_range: [30, 95]
+
+# the second degradation process
+second_blur_prob: 0.8
+resize_prob2: [0.3, 0.4, 0.3]  # up, down, keep
+resize_range2: [0.3, 1.2]
+gaussian_noise_prob2: 0.5
+noise_range2: [1, 25]
+poisson_scale_range2: [0.05, 2.5]
+gray_noise_prob2: 0.4
+jpeg_range2: [30, 95]
+
+gt_size: 256
+queue_size: 180
+
+# dataset and data loader settings
+datasets:
+  train:
+    name: DF2K+OST
+    type: RealESRGANDataset
+    dataroot_gt: datasets/DF2K
+    meta_info: datasets/DF2K/meta_info/meta_info_DF2Kmultiscale+OST_sub.txt
+    io_backend:
+      type: disk
+
+    blur_kernel_size: 21
+    kernel_list: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    sinc_prob: 0.1
+    blur_sigma: [0.2, 3]
+    betag_range: [0.5, 4]
+    betap_range: [1, 2]
+
+    blur_kernel_size2: 21
+    kernel_list2: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob2: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    sinc_prob2: 0.1
+    blur_sigma2: [0.2, 1.5]
+    betag_range2: [0.5, 4]
+    betap_range2: [1, 2]
+
+    final_sinc_prob: 0.8
+
+    gt_size: 256
+    use_hflip: True
+    use_rot: False
+
+    # data loader
+    num_worker_per_gpu: 5
+    batch_size_per_gpu: 12
+    dataset_enlarge_ratio: 1
+    prefetch_mode: ~
+
+  # Uncomment these for validation
+  # val:
+  #   name: validation
+  #   type: PairedImageDataset
+  #   dataroot_gt: path_to_gt
+  #   dataroot_lq: path_to_lq
+  #   io_backend:
+  #     type: disk
+
+# network structures
+network_g:
+  type: RRDBNet
+  num_in_ch: 3
+  num_out_ch: 3
+  num_feat: 64
+  num_block: 23
+  num_grow_ch: 32
+  scale: 2
+
+network_d:
+  type: UNetDiscriminatorSN
+  num_in_ch: 3
+  num_feat: 64
+  skip_connection: True
+
+# path
+path:
+  # use the pre-trained Real-ESRNet model
+  pretrain_network_g: experiments/pretrained_models/RealESRNet_x2plus.pth
+  param_key_g: params_ema
+  strict_load_g: true
+  resume_state: ~
+
+# training settings
+train:
+  ema_decay: 0.999
+  optim_g:
+    type: Adam
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+  optim_d:
+    type: Adam
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+
+  scheduler:
+    type: MultiStepLR
+    milestones: [400000]
+    gamma: 0.5
+
+  total_iter: 400000
+  warmup_iter: -1  # no warm up
+
+  # losses
+  pixel_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+  # perceptual loss (content and style losses)
+  perceptual_opt:
+    type: PerceptualLoss
+    layer_weights:
+      # before relu
+      'conv1_2': 0.1
+      'conv2_2': 0.1
+      'conv3_4': 1
+      'conv4_4': 1
+      'conv5_4': 1
+    vgg_type: vgg19
+    use_input_norm: true
+    perceptual_weight: !!float 1.0
+    style_weight: 0
+    range_norm: false
+    criterion: l1
+  # gan loss
+  gan_opt:
+    type: GANLoss
+    gan_type: vanilla
+    real_label_val: 1.0
+    fake_label_val: 0.0
+    loss_weight: !!float 1e-1
+
+  net_d_iters: 1
+  net_d_init_iters: 0
+
+# Uncomment these for validation
+# validation settings
+# val:
+#   val_freq: !!float 5e3
+#   save_img: True
+
+#   metrics:
+#     psnr: # metric name
+#       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: ~
+
+# dist training settings
+dist_params:
+  backend: nccl
+  port: 29500

BasicSR/options/train/RealESRGAN/train_realesrgan_x4plus.yml

+# general settings
+name: train_RealESRGANx4plus_400k_B12G4
+model_type: RealESRGANModel
+scale: 4
+num_gpu: auto  # auto: can infer from your visible devices automatically. official: 4 GPUs
+manual_seed: 0
+
+# ----------------- options for synthesizing training data in RealESRGANModel ----------------- #
+# USM the ground-truth
+l1_gt_usm: True
+percep_gt_usm: True
+gan_gt_usm: False
+
+# the first degradation process
+resize_prob: [0.2, 0.7, 0.1]  # up, down, keep
+resize_range: [0.15, 1.5]
+gaussian_noise_prob: 0.5
+noise_range: [1, 30]
+poisson_scale_range: [0.05, 3]
+gray_noise_prob: 0.4
+jpeg_range: [30, 95]
+
+# the second degradation process
+second_blur_prob: 0.8
+resize_prob2: [0.3, 0.4, 0.3]  # up, down, keep
+resize_range2: [0.3, 1.2]
+gaussian_noise_prob2: 0.5
+noise_range2: [1, 25]
+poisson_scale_range2: [0.05, 2.5]
+gray_noise_prob2: 0.4
+jpeg_range2: [30, 95]
+
+gt_size: 256
+queue_size: 180
+
+# dataset and data loader settings
+datasets:
+  train:
+    name: DF2K+OST
+    type: RealESRGANDataset
+    dataroot_gt: datasets/DF2K
+    meta_info: datasets/DF2K/meta_info/meta_info_DF2Kmultiscale+OST_sub.txt
+    io_backend:
+      type: disk
+
+    blur_kernel_size: 21
+    kernel_list: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    sinc_prob: 0.1
+    blur_sigma: [0.2, 3]
+    betag_range: [0.5, 4]
+    betap_range: [1, 2]
+
+    blur_kernel_size2: 21
+    kernel_list2: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob2: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    sinc_prob2: 0.1
+    blur_sigma2: [0.2, 1.5]
+    betag_range2: [0.5, 4]
+    betap_range2: [1, 2]
+
+    final_sinc_prob: 0.8
+
+    gt_size: 256
+    use_hflip: True
+    use_rot: False
+
+    # data loader
+    num_worker_per_gpu: 5
+    batch_size_per_gpu: 12
+    dataset_enlarge_ratio: 1
+    prefetch_mode: ~
+
+  # Uncomment these for validation
+  # val:
+  #   name: validation
+  #   type: PairedImageDataset
+  #   dataroot_gt: path_to_gt
+  #   dataroot_lq: path_to_lq
+  #   io_backend:
+  #     type: disk
+
+# network structures
+network_g:
+  type: RRDBNet
+  num_in_ch: 3
+  num_out_ch: 3
+  num_feat: 64
+  num_block: 23
+  num_grow_ch: 32
+
+network_d:
+  type: UNetDiscriminatorSN
+  num_in_ch: 3
+  num_feat: 64
+  skip_connection: True
+
+# path
+path:
+  # use the pre-trained Real-ESRNet model
+  pretrain_network_g: experiments/pretrained_models/RealESRNet_x4plus.pth
+  param_key_g: params_ema
+  strict_load_g: true
+  resume_state: ~
+
+# training settings
+train:
+  ema_decay: 0.999
+  optim_g:
+    type: Adam
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+  optim_d:
+    type: Adam
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+
+  scheduler:
+    type: MultiStepLR
+    milestones: [400000]
+    gamma: 0.5
+
+  total_iter: 400000
+  warmup_iter: -1  # no warm up
+
+  # losses
+  pixel_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+  # perceptual loss (content and style losses)
+  perceptual_opt:
+    type: PerceptualLoss
+    layer_weights:
+      # before relu
+      'conv1_2': 0.1
+      'conv2_2': 0.1
+      'conv3_4': 1
+      'conv4_4': 1
+      'conv5_4': 1
+    vgg_type: vgg19
+    use_input_norm: true
+    perceptual_weight: !!float 1.0
+    style_weight: 0
+    range_norm: false
+    criterion: l1
+  # gan loss
+  gan_opt:
+    type: GANLoss
+    gan_type: vanilla
+    real_label_val: 1.0
+    fake_label_val: 0.0
+    loss_weight: !!float 1e-1
+
+  net_d_iters: 1
+  net_d_init_iters: 0
+
+# Uncomment these for validation
+# validation settings
+# val:
+#   val_freq: !!float 5e3
+#   save_img: True
+
+#   metrics:
+#     psnr: # metric name
+#       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: ~
+
+# dist training settings
+dist_params:
+  backend: nccl
+  port: 29500

BasicSR/options/train/RealESRGAN/train_realesrnet_x2plus.yml

+# general settings
+name: train_RealESRNetx2plus_1000k_B12G4
+model_type: RealESRNetModel
+scale: 2
+num_gpu: auto  # auto: can infer from your visible devices automatically. official: 4 GPUs
+manual_seed: 0
+
+# ----------------- options for synthesizing training data in RealESRNetModel ----------------- #
+gt_usm: True  # USM the ground-truth
+
+# the first degradation process
+resize_prob: [0.2, 0.7, 0.1]  # up, down, keep
+resize_range: [0.15, 1.5]
+gaussian_noise_prob: 0.5
+noise_range: [1, 30]
+poisson_scale_range: [0.05, 3]
+gray_noise_prob: 0.4
+jpeg_range: [30, 95]
+
+# the second degradation process
+second_blur_prob: 0.8
+resize_prob2: [0.3, 0.4, 0.3]  # up, down, keep
+resize_range2: [0.3, 1.2]
+gaussian_noise_prob2: 0.5
+noise_range2: [1, 25]
+poisson_scale_range2: [0.05, 2.5]
+gray_noise_prob2: 0.4
+jpeg_range2: [30, 95]
+
+gt_size: 256
+queue_size: 180
+
+# dataset and data loader settings
+datasets:
+  train:
+    name: DF2K+OST
+    type: RealESRGANDataset
+    dataroot_gt: datasets/DF2K
+    meta_info: datasets/DF2K/meta_info/meta_info_DF2Kmultiscale+OST_sub.txt
+    io_backend:
+      type: disk
+
+    blur_kernel_size: 21
+    kernel_list: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    sinc_prob: 0.1
+    blur_sigma: [0.2, 3]
+    betag_range: [0.5, 4]
+    betap_range: [1, 2]
+
+    blur_kernel_size2: 21
+    kernel_list2: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob2: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    sinc_prob2: 0.1
+    blur_sigma2: [0.2, 1.5]
+    betag_range2: [0.5, 4]
+    betap_range2: [1, 2]
+
+    final_sinc_prob: 0.8
+
+    gt_size: 256
+    use_hflip: True
+    use_rot: False
+
+    # data loader
+    num_worker_per_gpu: 5
+    batch_size_per_gpu: 12
+    dataset_enlarge_ratio: 1
+    prefetch_mode: ~
+
+  # Uncomment these for validation
+  # val:
+  #   name: validation
+  #   type: PairedImageDataset
+  #   dataroot_gt: path_to_gt
+  #   dataroot_lq: path_to_lq
+  #   io_backend:
+  #     type: disk
+
+# network structures
+network_g:
+  type: RRDBNet
+  num_in_ch: 3
+  num_out_ch: 3
+  num_feat: 64
+  num_block: 23
+  num_grow_ch: 32
+  scale: 2
+
+# path
+path:
+  pretrain_network_g: experiments/pretrained_models/ESRGAN/RealESRGAN_x4plus.pth
+  param_key_g: params_ema
+  strict_load_g: False
+  resume_state: ~
+
+# training settings
+train:
+  ema_decay: 0.999
+  optim_g:
+    type: Adam
+    lr: !!float 2e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+
+  scheduler:
+    type: MultiStepLR
+    milestones: [1000000]
+    gamma: 0.5
+
+  total_iter: 1000000
+  warmup_iter: -1  # no warm up
+
+  # losses
+  pixel_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+
+# Uncomment these for validation
+# validation settings
+# val:
+#   val_freq: !!float 5e3
+#   save_img: True
+
+#   metrics:
+#     psnr: # metric name
+#       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: ~
+
+# dist training settings
+dist_params:
+  backend: nccl
+  port: 29500

BasicSR/options/train/RealESRGAN/train_realesrnet_x4plus.yml

+# general settings
+name: train_RealESRNetx4plus_1000k_B12G4
+model_type: RealESRNetModel
+scale: 4
+num_gpu: auto  # auto: can infer from your visible devices automatically. official: 4 GPUs
+manual_seed: 0
+
+# ----------------- options for synthesizing training data in RealESRNetModel ----------------- #
+gt_usm: True  # USM the ground-truth
+
+# the first degradation process
+resize_prob: [0.2, 0.7, 0.1]  # up, down, keep
+resize_range: [0.15, 1.5]
+gaussian_noise_prob: 0.5
+noise_range: [1, 30]
+poisson_scale_range: [0.05, 3]
+gray_noise_prob: 0.4
+jpeg_range: [30, 95]
+
+# the second degradation process
+second_blur_prob: 0.8
+resize_prob2: [0.3, 0.4, 0.3]  # up, down, keep
+resize_range2: [0.3, 1.2]
+gaussian_noise_prob2: 0.5
+noise_range2: [1, 25]
+poisson_scale_range2: [0.05, 2.5]
+gray_noise_prob2: 0.4
+jpeg_range2: [30, 95]
+
+gt_size: 256
+queue_size: 180
+
+# dataset and data loader settings
+datasets:
+  train:
+    name: DF2K+OST
+    type: RealESRGANDataset
+    dataroot_gt: datasets/DF2K
+    meta_info: datasets/DF2K/meta_info/meta_info_DF2Kmultiscale+OST_sub.txt
+    io_backend:
+      type: disk
+
+    blur_kernel_size: 21
+    kernel_list: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    sinc_prob: 0.1
+    blur_sigma: [0.2, 3]
+    betag_range: [0.5, 4]
+    betap_range: [1, 2]
+
+    blur_kernel_size2: 21
+    kernel_list2: ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
+    kernel_prob2: [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
+    sinc_prob2: 0.1
+    blur_sigma2: [0.2, 1.5]
+    betag_range2: [0.5, 4]
+    betap_range2: [1, 2]
+
+    final_sinc_prob: 0.8
+
+    gt_size: 256
+    use_hflip: True
+    use_rot: False
+
+    # data loader
+    num_worker_per_gpu: 5
+    batch_size_per_gpu: 12
+    dataset_enlarge_ratio: 1
+    prefetch_mode: ~
+
+  # Uncomment these for validation
+  # val:
+  #   name: validation
+  #   type: PairedImageDataset
+  #   dataroot_gt: path_to_gt
+  #   dataroot_lq: path_to_lq
+  #   io_backend:
+  #     type: disk
+
+# network structures
+network_g:
+  type: RRDBNet
+  num_in_ch: 3
+  num_out_ch: 3
+  num_feat: 64
+  num_block: 23
+  num_grow_ch: 32
+
+# path
+path:
+  pretrain_network_g: experiments/pretrained_models/ESRGAN/ESRGAN_SRx4_DF2KOST_official-ff704c30.pth
+  param_key_g: params_ema
+  strict_load_g: true
+  resume_state: ~
+
+# training settings
+train:
+  ema_decay: 0.999
+  optim_g:
+    type: Adam
+    lr: !!float 2e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+
+  scheduler:
+    type: MultiStepLR
+    milestones: [1000000]
+    gamma: 0.5
+
+  total_iter: 1000000
+  warmup_iter: -1  # no warm up
+
+  # losses
+  pixel_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+
+# Uncomment these for validation
+# validation settings
+# val:
+#   val_freq: !!float 5e3
+#   save_img: True
+
+#   metrics:
+#     psnr: # metric name
+#       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: ~
+
+# dist training settings
+dist_params:
+  backend: nccl
+  port: 29500

BasicSR/options/train/SRResNet_SRGAN/README.md

+# SRResNet and SRGAN
+
+Experiment results and pre-trained model descriptions are put here.

BasicSR/options/train/SRResNet_SRGAN/train_MSRGAN_x4.yml

+# Modified SRResNet w/o BN from:
+# Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network
+# Use perceptual loss before ReLU as ESRGAN does
+
+# general settings
+name: 004_MSRGAN_x4_f64b16_DIV2K_400k_B16G1_wandb
+model_type: SRGANModel
+scale: 4
+num_gpu: 1  # set num_gpu: 0 for cpu mode
+manual_seed: 0
+
+# dataset and data loader settings
+datasets:
+  train:
+    name: DIV2K
+    type: PairedImageDataset
+    dataroot_gt: datasets/DIV2K/DIV2K_train_HR_sub
+    dataroot_lq: datasets/DIV2K/DIV2K_train_LR_bicubic/X4_sub
+    # (for lmdb)
+    # dataroot_gt: datasets/DIV2K/DIV2K_train_HR_sub.lmdb
+    # dataroot_lq: datasets/DIV2K/DIV2K_train_LR_bicubic_X4_sub.lmdb
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+      # (for lmdb)
+      # type: lmdb
+
+    gt_size: 128
+    use_hflip: true
+    use_rot: true
+
+    # data loader
+    num_worker_per_gpu: 6
+    batch_size_per_gpu: 16
+    dataset_enlarge_ratio: 100
+    prefetch_mode: ~
+
+  val:
+    name: Set14
+    type: PairedImageDataset
+    dataroot_gt: datasets/Set14/GTmod12
+    dataroot_lq: datasets/Set14/LRbicx4
+    io_backend:
+      type: disk
+
+# network structures
+network_g:
+  type: MSRResNet
+  num_in_ch: 3
+  num_out_ch: 3
+  num_feat: 64
+  num_block: 16
+  upscale: 4
+
+network_d:
+  type: VGGStyleDiscriminator
+  num_in_ch: 3
+  num_feat: 64
+
+# path
+path:
+  pretrain_network_g: experiments/001_MSRResNet_x4_f64b16_DIV2K_1000k_B16G1_wandb/models/net_g_1000000.pth
+  strict_load_g: true
+  resume_state: ~
+
+# training settings
+train:
+  ema_decay: 0.999
+  optim_g:
+    type: Adam
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+  optim_d:
+    type: Adam
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+
+  scheduler:
+    type: MultiStepLR
+    milestones: [50000, 100000, 200000, 300000]
+    gamma: 0.5
+
+  total_iter: 400000
+  warmup_iter: -1  # no warm up
+
+  # losses
+  pixel_opt:
+    type: L1Loss
+    loss_weight: !!float 1e-2
+    reduction: mean
+  perceptual_opt:
+    type: PerceptualLoss
+    layer_weights:
+      'conv5_4': 1  # before relu
+    vgg_type: vgg19
+    use_input_norm: true
+    range_norm: false
+    perceptual_weight: 1.0
+    style_weight: 0
+    criterion: l1
+  gan_opt:
+    type: GANLoss
+    gan_type: vanilla
+    real_label_val: 1.0
+    fake_label_val: 0.0
+    loss_weight: !!float 5e-3
+
+  net_d_iters: 1
+  net_d_init_iters: 0
+
+# 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: ~
+
+# dist training settings
+dist_params:
+  backend: nccl
+  port: 29500

BasicSR/options/train/SRResNet_SRGAN/train_MSRResNet_x2.yml

+# Modified SRResNet w/o BN from:
+# Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network
+
+# general settings
+name: 002_MSRResNet_x2_f64b16_DIV2K_1000k_B16G1_001pretrain_wandb
+model_type: SRModel
+scale: 2
+num_gpu: 1  # set num_gpu: 0 for cpu mode
+manual_seed: 0
+
+# dataset and data loader settings
+datasets:
+  train:
+    name: DIV2K
+    type: PairedImageDataset
+    dataroot_gt: datasets/DIV2K/DIV2K_train_HR_sub
+    dataroot_lq: datasets/DIV2K/DIV2K_train_LR_bicubic/X2_sub
+    # (for lmdb)
+    # dataroot_gt: datasets/DIV2K/DIV2K_train_HR_sub.lmdb
+    # dataroot_lq: datasets/DIV2K/DIV2K_train_LR_bicubic_X2_sub.lmdb
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+      # (for lmdb)
+      # type: lmdb
+
+    gt_size: 64
+    use_hflip: true
+    use_rot: true
+
+    # data loader
+    num_worker_per_gpu: 6
+    batch_size_per_gpu: 16
+    dataset_enlarge_ratio: 100
+    prefetch_mode: ~
+
+  val:
+    name: Set5
+    type: PairedImageDataset
+    dataroot_gt: datasets/Set5/GTmod12
+    dataroot_lq: datasets/Set5/LRbicx2
+    io_backend:
+      type: disk
+
+# network structures
+network_g:
+  type: MSRResNet
+  num_in_ch: 3
+  num_out_ch: 3
+  num_feat: 64
+  num_block: 16
+  upscale: 2
+
+# path
+path:
+  pretrain_network_g: experiments/001_MSRResNet_x4_f64b16_DIV2K_1000k_B16G1_wandb/models/net_g_1000000.pth
+  strict_load_g: false
+  resume_state: ~
+
+# training settings
+train:
+  ema_decay: 0.999
+  optim_g:
+    type: Adam
+    lr: !!float 2e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+
+  scheduler:
+    type: CosineAnnealingRestartLR
+    periods: [250000, 250000, 250000, 250000]
+    restart_weights: [1, 1, 1, 1]
+    eta_min: !!float 1e-7
+
+  total_iter: 1000000
+  warmup_iter: -1  # no warm up
+
+  # losses
+  pixel_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+
+# validation settings
+val:
+  val_freq: !!float 5e3
+  save_img: false
+
+  metrics:
+    psnr: # metric name, can be arbitrary
+      type: calculate_psnr
+      crop_border: 2
+      test_y_channel: false
+
+# logging settings
+logger:
+  print_freq: 100
+  save_checkpoint_freq: !!float 5e3
+  use_tb_logger: true
+  wandb:
+    project: ~
+    resume_id: ~
+
+# dist training settings
+dist_params:
+  backend: nccl
+  port: 29500

BasicSR/options/train/SRResNet_SRGAN/train_MSRResNet_x3.yml

+# Modified SRResNet w/o BN from:
+# Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network
+
+# general settings
+name: 003_MSRResNet_x3_f64b16_DIV2K_1000k_B16G1_001pretrain_wandb
+model_type: SRModel
+scale: 3
+num_gpu: 1  # set num_gpu: 0 for cpu mode
+manual_seed: 0
+
+# dataset and data loader settings
+datasets:
+  train:
+    name: DIV2K
+    type: PairedImageDataset
+    dataroot_gt: datasets/DIV2K/DIV2K_train_HR_sub
+    dataroot_lq: datasets/DIV2K/DIV2K_train_LR_bicubic/X3_sub
+    # (for lmdb)
+    # dataroot_gt: datasets/DIV2K/DIV2K_train_HR_sub.lmdb
+    # dataroot_lq: datasets/DIV2K/DIV2K_train_LR_bicubic_X3_sub.lmdb
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+      # (for lmdb)
+      # type: lmdb
+
+    gt_size: 96
+    use_hflip: true
+    use_rot: true
+
+    # data loader
+    num_worker_per_gpu: 6
+    batch_size_per_gpu: 16
+    dataset_enlarge_ratio: 100
+    prefetch_mode: ~
+
+  val:
+    name: Set5
+    type: PairedImageDataset
+    dataroot_gt: datasets/Set5/GTmod12
+    dataroot_lq: datasets/Set5/LRbicx3
+    io_backend:
+      type: disk
+
+# network structures
+network_g:
+  type: MSRResNet
+  num_in_ch: 3
+  num_out_ch: 3
+  num_feat: 64
+  num_block: 16
+  upscale: 3
+
+# path
+path:
+  pretrain_network_g: experiments/001_MSRResNet_x4_f64b16_DIV2K_1000k_B16G1_wandb/models/net_g_1000000.pth
+  strict_load_g: false
+  resume_state: ~
+
+# training settings
+train:
+  ema_decay: 0.999
+  optim_g:
+    type: Adam
+    lr: !!float 2e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+
+  scheduler:
+    type: CosineAnnealingRestartLR
+    periods: [250000, 250000, 250000, 250000]
+    restart_weights: [1, 1, 1, 1]
+    eta_min: !!float 1e-7
+
+  total_iter: 1000000
+  warmup_iter: -1  # no warm up
+
+  # losses
+  pixel_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+
+# validation settings
+val:
+  val_freq: !!float 5e3
+  save_img: false
+
+  metrics:
+    psnr: # metric name, can be arbitrary
+      type: calculate_psnr
+      crop_border: 3
+      test_y_channel: false
+
+# logging settings
+logger:
+  print_freq: 100
+  save_checkpoint_freq: !!float 5e3
+  use_tb_logger: true
+  wandb:
+    project: ~
+    resume_id: ~
+
+# dist training settings
+dist_params:
+  backend: nccl
+  port: 29500

BasicSR/options/train/SRResNet_SRGAN/train_MSRResNet_x4.yml

+# Modified SRResNet w/o BN from:
+# Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network
+
+# ----------- Commands for running
+# ----------- Single GPU with auto_resume
+# PYTHONPATH="./:${PYTHONPATH}"  CUDA_VISIBLE_DEVICES=0 python basicsr/train.py -opt options/train/SRResNet_SRGAN/train_MSRResNet_x4.yml --auto_resume
+
+# general settings
+name: 001_MSRResNet_x4_f64b16_DIV2K_1000k_B16G1_wandb
+model_type: SRModel
+scale: 4
+num_gpu: 1  # set num_gpu: 0 for cpu mode
+manual_seed: 0
+
+# dataset and data loader settings
+datasets:
+  train:
+    name: DIV2K
+    type: PairedImageDataset
+    dataroot_gt: datasets/DF2K/DIV2K_train_HR_sub
+    dataroot_lq: datasets/DF2K/DIV2K_train_LR_bicubic_X4_sub
+    meta_info_file: basicsr/data/meta_info/meta_info_DIV2K800sub_GT.txt
+    # (for lmdb)
+    # dataroot_gt: datasets/DIV2K/DIV2K_train_HR_sub.lmdb
+    # dataroot_lq: datasets/DIV2K/DIV2K_train_LR_bicubic_X4_sub.lmdb
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+      # (for lmdb)
+      # type: lmdb
+
+    gt_size: 128
+    use_hflip: true
+    use_rot: true
+
+    # data loader
+    num_worker_per_gpu: 6
+    batch_size_per_gpu: 16
+    dataset_enlarge_ratio: 100
+    prefetch_mode: ~
+
+  val:
+    name: Set5
+    type: PairedImageDataset
+    dataroot_gt: datasets/Set5/GTmod12
+    dataroot_lq: datasets/Set5/LRbicx4
+    io_backend:
+      type: disk
+
+  val_2:
+    name: Set14
+    type: PairedImageDataset
+    dataroot_gt: datasets/Set14/GTmod12
+    dataroot_lq: datasets/Set14/LRbicx4
+    io_backend:
+      type: disk
+
+# network structures
+network_g:
+  type: MSRResNet
+  num_in_ch: 3
+  num_out_ch: 3
+  num_feat: 64
+  num_block: 16
+  upscale: 4
+
+# path
+path:
+  pretrain_network_g: ~
+  param_key_g: params
+  strict_load_g: true
+  resume_state: ~
+
+# training settings
+train:
+  ema_decay: 0.999
+  optim_g:
+    type: Adam
+    lr: !!float 2e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+
+  scheduler:
+    type: CosineAnnealingRestartLR
+    periods: [250000, 250000, 250000, 250000]
+    restart_weights: [1, 1, 1, 1]
+    eta_min: !!float 1e-7
+
+  total_iter: 1000000
+  warmup_iter: -1  # no warm up
+
+  # losses
+  pixel_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+
+# validation settings
+val:
+  val_freq: !!float 5e3
+  save_img: false
+
+  metrics:
+    psnr: # metric name, can be arbitrary
+      type: calculate_psnr
+      crop_border: 4
+      test_y_channel: false
+      better: higher  # the higher, the better. Default: higher
+    niqe:
+      type: calculate_niqe
+      crop_border: 4
+      better: lower  # the lower, the better
+
+# logging settings
+logger:
+  print_freq: 100
+  save_checkpoint_freq: !!float 5e3
+  use_tb_logger: true
+  wandb:
+    project: ~
+    resume_id: ~
+
+# dist training settings
+dist_params:
+  backend: nccl
+  port: 29500

BasicSR/options/train/StyleGAN/train_StyleGAN2_256_Cmul2_FFHQ.yml

+# general settings
+name: 501_StyleGAN2_256_Cmul2_FFHQ_800k_B24G8_scratch
+model_type: StyleGAN2Model
+num_gpu: 8  # set num_gpu: 0 for cpu mode
+manual_seed: 0
+
+# dataset and data loader settings
+datasets:
+  train:
+    name: FFHQ
+    type: FFHQDataset
+    dataroot_gt: datasets/ffhq/ffhq_256.lmdb
+    io_backend:
+      type: lmdb
+
+    use_hflip: true
+    mean: [0.5, 0.5, 0.5]
+    std: [0.5, 0.5, 0.5]
+
+    # data loader
+    num_worker_per_gpu: 6
+    batch_size_per_gpu: 3
+    dataset_enlarge_ratio: 100
+    prefetch_mode: ~
+
+# network structures
+network_g:
+  type: StyleGAN2Generator
+  out_size: 256
+  num_style_feat: 512
+  num_mlp: 8
+  channel_multiplier: 2
+  resample_kernel: [1, 3, 3, 1]
+  lr_mlp: 0.01
+
+network_d:
+  type: StyleGAN2Discriminator
+  out_size: 256
+  channel_multiplier: 2
+  resample_kernel: [1, 3, 3, 1]
+
+# path
+path:
+  pretrain_network_g: ~
+  strict_load_g: true
+  resume_state: ~
+
+# training settings
+train:
+  optim_g:
+    type: Adam
+    lr: !!float 2e-3
+
+  optim_d:
+    type: Adam
+    lr: !!float 2e-3
+
+  scheduler:
+    type: MultiStepLR
+    milestones: [600000]
+    gamma: 0.5
+
+  total_iter: 800000
+  warmup_iter: -1  # no warm up
+
+  # losses
+  gan_opt:
+    type: GANLoss
+    gan_type: wgan_softplus
+    loss_weight: !!float 1
+  # r1 regularization for discriminator
+  r1_reg_weight: 10
+  # path length regularization for generator
+  path_batch_shrink: 2
+  path_reg_weight: 2
+
+  net_g_reg_every: 4
+  net_d_reg_every: 16
+  mixing_prob: 0.9
+
+  net_d_iters: 1
+  net_d_init_iters: 0
+
+# validation settings
+val:
+  val_freq: !!float 5e3
+  save_img: true
+
+# logging settings
+logger:
+  print_freq: 100
+  save_checkpoint_freq: !!float 5e3
+  use_tb_logger: true
+  wandb:
+    project: ~
+    resume_id: ~
+
+# dist training settings
+dist_params:
+  backend: nccl
+  port: 29500

BasicSR/options/train/SwinIR/train_SwinIR_SRx2_scratch.yml

+# general settings
+name: train_SwinIR_SRx2_scratch_P48W8_DIV2K_500k_B4G8
+model_type: SwinIRModel
+scale: 2
+num_gpu: auto
+manual_seed: 0
+
+# dataset and data loader settings
+datasets:
+  train:
+    name: DIV2K
+    type: PairedImageDataset
+    dataroot_gt: datasets/DF2K/DIV2K_train_HR_sub
+    dataroot_lq: datasets/DF2K/DIV2K_train_LR_bicubic_X2_sub
+    meta_info_file: basicsr/data/meta_info/meta_info_DIV2K800sub_GT.txt
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+
+    gt_size: 96
+    use_hflip: true
+    use_rot: true
+
+    # data loader
+    num_worker_per_gpu: 6
+    batch_size_per_gpu: 4
+    dataset_enlarge_ratio: 1
+    prefetch_mode: ~
+
+  val:
+    name: Set5
+    type: PairedImageDataset
+    dataroot_gt: datasets/Set5/GTmod12
+    dataroot_lq: datasets/Set5/LRbicx2
+    io_backend:
+      type: disk
+
+# network structures
+network_g:
+  type: SwinIR
+  upscale: 2
+  in_chans: 3
+  img_size: 48
+  window_size: 8
+  img_range: 1.
+  depths: [6, 6, 6, 6, 6, 6]
+  embed_dim: 180
+  num_heads: [6, 6, 6, 6, 6, 6]
+  mlp_ratio: 2
+  upsampler: 'pixelshuffle'
+  resi_connection: '1conv'
+
+# path
+path:
+  pretrain_network_g: ~
+  strict_load_g: true
+  resume_state: ~
+
+# training settings
+train:
+  ema_decay: 0.999
+  optim_g:
+    type: Adam
+    lr: !!float 2e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+
+  scheduler:
+    type: MultiStepLR
+    milestones: [250000, 400000, 450000, 475000]
+    gamma: 0.5
+
+  total_iter: 500000
+  warmup_iter: -1  # no warm up
+
+  # losses
+  pixel_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+
+# validation settings
+val:
+  val_freq: !!float 5e3
+  save_img: false
+
+  metrics:
+    psnr: # metric name, can be arbitrary
+      type: calculate_psnr
+      crop_border: 2
+      test_y_channel: false
+
+# logging settings
+logger:
+  print_freq: 100
+  save_checkpoint_freq: !!float 5e3
+  use_tb_logger: true
+  wandb:
+    project: ~
+    resume_id: ~
+
+# dist training settings
+dist_params:
+  backend: nccl
+  port: 29500

BasicSR/options/train/SwinIR/train_SwinIR_SRx4_scratch.yml

+# general settings
+name: train_SwinIR_SRx4_scratch_P48W8_DIV2K_500k_B4G8
+model_type: SwinIRModel
+scale: 4
+num_gpu: auto
+manual_seed: 0
+
+# dataset and data loader settings
+datasets:
+  train:
+    name: DIV2K
+    type: PairedImageDataset
+    dataroot_gt: datasets/DF2K/DIV2K_train_HR_sub
+    dataroot_lq: datasets/DF2K/DIV2K_train_LR_bicubic_X4_sub
+    meta_info_file: basicsr/data/meta_info/meta_info_DIV2K800sub_GT.txt
+    filename_tmpl: '{}'
+    io_backend:
+      type: disk
+
+    gt_size: 192
+    use_hflip: true
+    use_rot: true
+
+    # data loader
+    num_worker_per_gpu: 6
+    batch_size_per_gpu: 4
+    dataset_enlarge_ratio: 1
+    prefetch_mode: ~
+
+  val:
+    name: Set5
+    type: PairedImageDataset
+    dataroot_gt: datasets/Set5/GTmod12
+    dataroot_lq: datasets/Set5/LRbicx4
+    io_backend:
+      type: disk
+
+# network structures
+network_g:
+  type: SwinIR
+  upscale: 4
+  in_chans: 3
+  img_size: 48
+  window_size: 8
+  img_range: 1.
+  depths: [6, 6, 6, 6, 6, 6]
+  embed_dim: 180
+  num_heads: [6, 6, 6, 6, 6, 6]
+  mlp_ratio: 2
+  upsampler: 'pixelshuffle'
+  resi_connection: '1conv'
+
+# path
+path:
+  pretrain_network_g: ~
+  strict_load_g: true
+  resume_state: ~
+
+# training settings
+train:
+  ema_decay: 0.999
+  optim_g:
+    type: Adam
+    lr: !!float 2e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+
+  scheduler:
+    type: MultiStepLR
+    milestones: [250000, 400000, 450000, 475000]
+    gamma: 0.5
+
+  total_iter: 500000
+  warmup_iter: -1  # no warm up
+
+  # losses
+  pixel_opt:
+    type: L1Loss
+    loss_weight: 1.0
+    reduction: mean
+
+# validation settings
+val:
+  val_freq: !!float 5e3
+  save_img: false
+
+  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: ~
+
+# dist training settings
+dist_params:
+  backend: nccl
+  port: 29500

BasicSR/options/train/VideoRecurrentGAN/train_VideoRecurrentGANModel_REDS.yml

+# general settings
+name: VideoRecurrentGANModel_REDS
+model_type: VideoRecurrentGANModel
+scale: 4
+num_gpu: 1  # set num_gpu: 0 for cpu mode
+manual_seed: 0
+
+# dataset and data loader settings
+datasets:
+  train:
+    name: REDS
+    type: REDSRecurrentDataset
+    dataroot_gt: datasets/REDS/train_sharp
+    dataroot_lq: datasets/REDS/train_sharp_bicubic
+    dataroot_flow: ~
+    meta_info_file: basicsr/data/meta_info/meta_info_REDS_GT.txt
+    val_partition: REDS4  # set to 'official' when use the official validation partition
+    test_mode: False
+    io_backend:
+      type: disk
+
+    num_frame: 15
+    gt_size: 256
+    interval_list: [1]
+    random_reverse: false
+    use_hflip: true
+    use_rot: true
+
+    # data loader
+    num_worker_per_gpu: 6
+    batch_size_per_gpu: 4
+    dataset_enlarge_ratio: 200
+    prefetch_mode: ~
+
+  val:
+    name: REDS4
+    type: VideoRecurrentTestDataset
+    dataroot_gt: datasets/REDS4/GT
+    dataroot_lq: datasets/REDS4/sharp_bicubic
+
+    cache_data: true
+    io_backend:
+      type: disk
+
+    num_frame: -1  # For recurrent, num_frame has no effect. Set to -1
+
+# network structures
+network_g:
+  type: BasicVSR
+  num_feat: 64
+  num_block: 30
+  spynet_path: experiments/pretrained_models/flownet/spynet_sintel_final-3d2a1287.pth
+
+network_d:
+  type: VGGStyleDiscriminator
+  num_in_ch: 3
+  num_feat: 32
+  input_size: 256
+
+# path
+path:
+  pretrain_network_g: experiments/pretrained_models/BasicVSR_REDS4.pth
+  strict_load_g: true
+  resume_state: ~
+
+# training settings
+train:
+  ema_decay: 0.999
+  optim_g:
+    type: Adam
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+  lr_flow: !!float 0
+  optim_d:
+    type: Adam
+    lr: !!float 1e-4
+    weight_decay: 0
+    betas: [0.9, 0.99]
+
+  scheduler:
+    type: MultiStepLR
+    milestones: [50000, 100000, 200000, 300000]
+    gamma: 0.5
+
+  total_iter: 400000
+  warmup_iter: -1  # no warm up
+  fix_flow: ~
+
+  # losses
+  pixel_opt:
+    type: L1Loss
+    loss_weight: !!float 1e-2
+    reduction: mean
+  perceptual_opt:
+    type: PerceptualLoss
+    layer_weights:
+      'conv5_4': 1  # before relu
+    vgg_type: vgg19
+    use_input_norm: true
+    range_norm: false
+    perceptual_weight: 1.0
+    style_weight: 0
+    criterion: l1
+  gan_opt:
+    type: GANLoss
+    gan_type: vanilla
+    real_label_val: 1.0
+    fake_label_val: 0.0
+    loss_weight: !!float 5e-3
+
+  net_d_iters: 1
+  net_d_init_iters: 0
+
+# validation settings
+val:
+  val_freq: !!float 1e4
+  save_img: false
+
+  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: ~
+
+# dist training settings
+dist_params:
+  backend: nccl
+  port: 29500
+
+find_unused_parameters: true

BasicSR/requirements.txt

+addict
+future
+lmdb
+numpy>=1.17
+opencv-python
+Pillow
+pyyaml
+requests
+scikit-image
+scipy
+tb-nightly
+torch>=1.7
+torchvision
+tqdm
+yapf

BasicSR/scripts/data_preparation/create_lmdb.py

+import argparse
+from os import path as osp
+
+from basicsr.utils import scandir
+from basicsr.utils.lmdb_util import make_lmdb_from_imgs
+
+
+def create_lmdb_for_div2k():
+    """Create lmdb files for DIV2K dataset.
+
+    Usage:
+        Before run this script, please run `extract_subimages.py`.
+        Typically, there are four folders to be processed for DIV2K dataset.
+
+            * DIV2K_train_HR_sub
+            * DIV2K_train_LR_bicubic/X2_sub
+            * DIV2K_train_LR_bicubic/X3_sub
+            * DIV2K_train_LR_bicubic/X4_sub
+
+        Remember to modify opt configurations according to your settings.
+    """
+    # HR images
+    folder_path = 'datasets/DIV2K/DIV2K_train_HR_sub'
+    lmdb_path = 'datasets/DIV2K/DIV2K_train_HR_sub.lmdb'
+    img_path_list, keys = prepare_keys_div2k(folder_path)
+    make_lmdb_from_imgs(folder_path, lmdb_path, img_path_list, keys)
+
+    # LRx2 images
+    folder_path = 'datasets/DIV2K/DIV2K_train_LR_bicubic/X2_sub'
+    lmdb_path = 'datasets/DIV2K/DIV2K_train_LR_bicubic_X2_sub.lmdb'
+    img_path_list, keys = prepare_keys_div2k(folder_path)
+    make_lmdb_from_imgs(folder_path, lmdb_path, img_path_list, keys)
+
+    # LRx3 images
+    folder_path = 'datasets/DIV2K/DIV2K_train_LR_bicubic/X3_sub'
+    lmdb_path = 'datasets/DIV2K/DIV2K_train_LR_bicubic_X3_sub.lmdb'
+    img_path_list, keys = prepare_keys_div2k(folder_path)
+    make_lmdb_from_imgs(folder_path, lmdb_path, img_path_list, keys)
+
+    # LRx4 images
+    folder_path = 'datasets/DIV2K/DIV2K_train_LR_bicubic/X4_sub'
+    lmdb_path = 'datasets/DIV2K/DIV2K_train_LR_bicubic_X4_sub.lmdb'
+    img_path_list, keys = prepare_keys_div2k(folder_path)
+    make_lmdb_from_imgs(folder_path, lmdb_path, img_path_list, keys)
+
+
+def prepare_keys_div2k(folder_path):
+    """Prepare image path list and keys for DIV2K dataset.
+
+    Args:
+        folder_path (str): Folder path.
+
+    Returns:
+        list[str]: Image path list.
+        list[str]: Key list.
+    """
+    print('Reading image path list ...')
+    img_path_list = sorted(list(scandir(folder_path, suffix='png', recursive=False)))
+    keys = [img_path.split('.png')[0] for img_path in sorted(img_path_list)]
+
+    return img_path_list, keys
+
+
+def create_lmdb_for_reds():
+    """Create lmdb files for REDS dataset.
+
+    Usage:
+        Before run this script, please run :file:`merge_reds_train_val.py`.
+        We take two folders for example:
+
+            * train_sharp
+            * train_sharp_bicubic
+
+        Remember to modify opt configurations according to your settings.
+    """
+    # train_sharp
+    folder_path = 'datasets/REDS/train_sharp'
+    lmdb_path = 'datasets/REDS/train_sharp_with_val.lmdb'
+    img_path_list, keys = prepare_keys_reds(folder_path)
+    make_lmdb_from_imgs(folder_path, lmdb_path, img_path_list, keys, multiprocessing_read=True)
+
+    # train_sharp_bicubic
+    folder_path = 'datasets/REDS/train_sharp_bicubic'
+    lmdb_path = 'datasets/REDS/train_sharp_bicubic_with_val.lmdb'
+    img_path_list, keys = prepare_keys_reds(folder_path)
+    make_lmdb_from_imgs(folder_path, lmdb_path, img_path_list, keys, multiprocessing_read=True)
+
+
+def prepare_keys_reds(folder_path):
+    """Prepare image path list and keys for REDS dataset.
+
+    Args:
+        folder_path (str): Folder path.
+
+    Returns:
+        list[str]: Image path list.
+        list[str]: Key list.
+    """
+    print('Reading image path list ...')
+    img_path_list = sorted(list(scandir(folder_path, suffix='png', recursive=True)))
+    keys = [v.split('.png')[0] for v in img_path_list]  # example: 000/00000000
+
+    return img_path_list, keys
+
+
+def create_lmdb_for_vimeo90k():
+    """Create lmdb files for Vimeo90K dataset.
+
+    Usage:
+        Remember to modify opt configurations according to your settings.
+    """
+    # GT
+    folder_path = 'datasets/vimeo90k/vimeo_septuplet/sequences'
+    lmdb_path = 'datasets/vimeo90k/vimeo90k_train_GT_only4th.lmdb'
+    train_list_path = 'datasets/vimeo90k/vimeo_septuplet/sep_trainlist.txt'
+    img_path_list, keys = prepare_keys_vimeo90k(folder_path, train_list_path, 'gt')
+    make_lmdb_from_imgs(folder_path, lmdb_path, img_path_list, keys, multiprocessing_read=True)
+
+    # LQ
+    folder_path = 'datasets/vimeo90k/vimeo_septuplet_matlabLRx4/sequences'
+    lmdb_path = 'datasets/vimeo90k/vimeo90k_train_LR7frames.lmdb'
+    train_list_path = 'datasets/vimeo90k/vimeo_septuplet/sep_trainlist.txt'
+    img_path_list, keys = prepare_keys_vimeo90k(folder_path, train_list_path, 'lq')
+    make_lmdb_from_imgs(folder_path, lmdb_path, img_path_list, keys, multiprocessing_read=True)
+
+
+def prepare_keys_vimeo90k(folder_path, train_list_path, mode):
+    """Prepare image path list and keys for Vimeo90K dataset.
+
+    Args:
+        folder_path (str): Folder path.
+        train_list_path (str): Path to the official train list.
+        mode (str): One of 'gt' or 'lq'.
+
+    Returns:
+        list[str]: Image path list.
+        list[str]: Key list.
+    """
+    print('Reading image path list ...')
+    with open(train_list_path, 'r') as fin:
+        train_list = [line.strip() for line in fin]
+
+    img_path_list = []
+    keys = []
+    for line in train_list:
+        folder, sub_folder = line.split('/')
+        img_path_list.extend([osp.join(folder, sub_folder, f'im{j + 1}.png') for j in range(7)])
+        keys.extend([f'{folder}/{sub_folder}/im{j + 1}' for j in range(7)])
+
+    if mode == 'gt':
+        print('Only keep the 4th frame for the gt mode.')
+        img_path_list = [v for v in img_path_list if v.endswith('im4.png')]
+        keys = [v for v in keys if v.endswith('/im4')]
+
+    return img_path_list, keys
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        '--dataset',
+        type=str,
+        help=("Options: 'DIV2K', 'REDS', 'Vimeo90K' You may need to modify the corresponding configurations in codes."))
+    args = parser.parse_args()
+    dataset = args.dataset.lower()
+    if dataset == 'div2k':
+        create_lmdb_for_div2k()
+    elif dataset == 'reds':
+        create_lmdb_for_reds()
+    elif dataset == 'vimeo90k':
+        create_lmdb_for_vimeo90k()
+    else:
+        raise ValueError('Wrong dataset.')

BasicSR/scripts/data_preparation/download_datasets.py

+import argparse
+import glob
+import os
+from os import path as osp
+
+from basicsr.utils.download_util import download_file_from_google_drive
+
+
+def download_dataset(dataset, file_ids):
+    save_path_root = './datasets/'
+    os.makedirs(save_path_root, exist_ok=True)
+
+    for file_name, file_id in file_ids.items():
+        save_path = osp.abspath(osp.join(save_path_root, file_name))
+        if osp.exists(save_path):
+            user_response = input(f'{file_name} already exist. Do you want to cover it? Y/N\n')
+            if user_response.lower() == 'y':
+                print(f'Covering {file_name} to {save_path}')
+                download_file_from_google_drive(file_id, save_path)
+            elif user_response.lower() == 'n':
+                print(f'Skipping {file_name}')
+            else:
+                raise ValueError('Wrong input. Only accepts Y/N.')
+        else:
+            print(f'Downloading {file_name} to {save_path}')
+            download_file_from_google_drive(file_id, save_path)
+
+        # unzip
+        if save_path.endswith('.zip'):
+            extracted_path = save_path.replace('.zip', '')
+            print(f'Extract {save_path} to {extracted_path}')
+            import zipfile
+            with zipfile.ZipFile(save_path, 'r') as zip_ref:
+                zip_ref.extractall(extracted_path)
+
+            file_name = file_name.replace('.zip', '')
+            subfolder = osp.join(extracted_path, file_name)
+            if osp.isdir(subfolder):
+                print(f'Move {subfolder} to {extracted_path}')
+                import shutil
+                for path in glob.glob(osp.join(subfolder, '*')):
+                    shutil.move(path, extracted_path)
+                shutil.rmtree(subfolder)
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument(
+        'dataset',
+        type=str,
+        help=("Options: 'Set5', 'Set14'. "
+              "Set to 'all' if you want to download all the dataset."))
+    args = parser.parse_args()
+
+    file_ids = {
+        'Set5': {
+            'Set5.zip':  # file name
+            '1RtyIeUFTyW8u7oa4z7a0lSzT3T1FwZE9',  # file id
+        },
+        'Set14': {
+            'Set14.zip': '1vsw07sV8wGrRQ8UARe2fO5jjgy9QJy_E',
+        }
+    }
+
+    if args.dataset == 'all':
+        for dataset in file_ids.keys():
+            download_dataset(dataset, file_ids[dataset])
+    else:
+        download_dataset(args.dataset, file_ids[args.dataset])

BasicSR/scripts/data_preparation/extract_images_from_tfrecords.py

+import argparse
+import cv2
+import glob
+import numpy as np
+import os
+
+from basicsr.utils.lmdb_util import LmdbMaker
+
+
+def convert_celeba_tfrecords(tf_file, log_resolution, save_root, save_type='img', compress_level=1):
+    """Convert CelebA tfrecords to images or lmdb files.
+
+    Args:
+        tf_file (str): Input tfrecords file in glob pattern.
+            Example: 'datasets/celeba/celeba_tfrecords/validation/validation-r08-s-*-of-*.tfrecords'  # noqa:E501
+        log_resolution (int): Log scale of resolution.
+        save_root (str): Path root to save.
+        save_type (str): Save type. Options: img | lmdb. Default: img.
+        compress_level (int):  Compress level when encoding images. Default: 1.
+    """
+    if 'validation' in tf_file:
+        phase = 'validation'
+    else:
+        phase = 'train'
+    if save_type == 'lmdb':
+        save_path = os.path.join(save_root, f'celeba_{2**log_resolution}_{phase}.lmdb')
+        lmdb_maker = LmdbMaker(save_path)
+    elif save_type == 'img':
+        save_path = os.path.join(save_root, f'celeba_{2**log_resolution}_{phase}')
+    else:
+        raise ValueError('Wrong save type.')
+
+    os.makedirs(save_path, exist_ok=True)
+
+    idx = 0
+    for record in sorted(glob.glob(tf_file)):
+        print('Processing record: ', record)
+        record_iterator = tf.python_io.tf_record_iterator(record)
+        for string_record in record_iterator:
+            example = tf.train.Example()
+            example.ParseFromString(string_record)
+
+            # label = example.features.feature['label'].int64_list.value[0]
+            # attr = example.features.feature['attr'].int64_list.value
+            # male = attr[20]
+            # young = attr[39]
+
+            shape = example.features.feature['shape'].int64_list.value
+            h, w, c = shape
+            img_str = example.features.feature['data'].bytes_list.value[0]
+            img = np.fromstring(img_str, dtype=np.uint8).reshape((h, w, c))
+
+            img = img[:, :, [2, 1, 0]]
+
+            if save_type == 'img':
+                cv2.imwrite(os.path.join(save_path, f'{idx:08d}.png'), img)
+            elif save_type == 'lmdb':
+                _, img_byte = cv2.imencode('.png', img, [cv2.IMWRITE_PNG_COMPRESSION, compress_level])
+                key = f'{idx:08d}/r{log_resolution:02d}'
+                lmdb_maker.put(img_byte, key, (h, w, c))
+
+            idx += 1
+            print(idx)
+
+    if save_type == 'lmdb':
+        lmdb_maker.close()
+
+
+def convert_ffhq_tfrecords(tf_file, log_resolution, save_root, save_type='img', compress_level=1):
+    """Convert FFHQ tfrecords to images or lmdb files.
+
+    Args:
+        tf_file (str): Input tfrecords file.
+        log_resolution (int): Log scale of resolution.
+        save_root (str): Path root to save.
+        save_type (str): Save type. Options: img | lmdb. Default: img.
+        compress_level (int):  Compress level when encoding images. Default: 1.
+    """
+
+    if save_type == 'lmdb':
+        save_path = os.path.join(save_root, f'ffhq_{2**log_resolution}.lmdb')
+        lmdb_maker = LmdbMaker(save_path)
+    elif save_type == 'img':
+        save_path = os.path.join(save_root, f'ffhq_{2**log_resolution}')
+    else:
+        raise ValueError('Wrong save type.')
+
+    os.makedirs(save_path, exist_ok=True)
+
+    idx = 0
+    for record in sorted(glob.glob(tf_file)):
+        print('Processing record: ', record)
+        record_iterator = tf.python_io.tf_record_iterator(record)
+        for string_record in record_iterator:
+            example = tf.train.Example()
+            example.ParseFromString(string_record)
+
+            shape = example.features.feature['shape'].int64_list.value
+            c, h, w = shape
+            img_str = example.features.feature['data'].bytes_list.value[0]
+            img = np.fromstring(img_str, dtype=np.uint8).reshape((c, h, w))
+
+            img = img.transpose(1, 2, 0)
+            img = img[:, :, [2, 1, 0]]
+            if save_type == 'img':
+                cv2.imwrite(os.path.join(save_path, f'{idx:08d}.png'), img)
+            elif save_type == 'lmdb':
+                _, img_byte = cv2.imencode('.png', img, [cv2.IMWRITE_PNG_COMPRESSION, compress_level])
+                key = f'{idx:08d}/r{log_resolution:02d}'
+                lmdb_maker.put(img_byte, key, (h, w, c))
+
+            idx += 1
+            print(idx)
+
+    if save_type == 'lmdb':
+        lmdb_maker.close()
+
+
+def make_ffhq_lmdb_from_imgs(folder_path, log_resolution, save_root, save_type='lmdb', compress_level=1):
+    """Make FFHQ lmdb from images.
+
+    Args:
+        folder_path (str): Folder path.
+        log_resolution (int): Log scale of resolution.
+        save_root (str): Path root to save.
+        save_type (str): Save type. Options: img | lmdb. Default: img.
+        compress_level (int):  Compress level when encoding images. Default: 1.
+    """
+
+    if save_type == 'lmdb':
+        save_path = os.path.join(save_root, f'ffhq_{2**log_resolution}_crop1.2.lmdb')
+        lmdb_maker = LmdbMaker(save_path)
+    else:
+        raise ValueError('Wrong save type.')
+
+    os.makedirs(save_path, exist_ok=True)
+
+    img_list = sorted(glob.glob(os.path.join(folder_path, '*')))
+    for idx, img_path in enumerate(img_list):
+        print(f'Processing {idx}: ', img_path)
+        img = cv2.imread(img_path)
+        h, w, c = img.shape
+
+        if save_type == 'lmdb':
+            _, img_byte = cv2.imencode('.png', img, [cv2.IMWRITE_PNG_COMPRESSION, compress_level])
+            key = f'{idx:08d}/r{log_resolution:02d}'
+            lmdb_maker.put(img_byte, key, (h, w, c))
+
+    if save_type == 'lmdb':
+        lmdb_maker.close()
+
+
+if __name__ == '__main__':
+    """Read tfrecords w/o define a graph.
+
+    We have tested it on TensorFlow 1.15
+
+    References: http://warmspringwinds.github.io/tensorflow/tf-slim/2016/12/21/tfrecords-guide/
+    """
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        '--dataset', type=str, default='ffhq', help="Dataset name. Options: 'ffhq' | 'celeba'. Default: 'ffhq'.")
+    parser.add_argument(
+        '--tf_file',
+        type=str,
+        default='datasets/ffhq/ffhq-r10.tfrecords',
+        help=(
+            'Input tfrecords file. For celeba, it should be glob pattern. '
+            'Put quotes around the wildcard argument to prevent the shell '
+            'from expanding it.'
+            "Example: 'datasets/celeba/celeba_tfrecords/validation/validation-r08-s-*-of-*.tfrecords'"  # noqa:E501
+        ))
+    parser.add_argument('--log_resolution', type=int, default=10, help='Log scale of resolution.')
+    parser.add_argument('--save_root', type=str, default='datasets/ffhq/', help='Save root path.')
+    parser.add_argument(
+        '--save_type', type=str, default='img', help="Save type. Options: 'img' | 'lmdb'. Default: 'img'.")
+    parser.add_argument(
+        '--compress_level', type=int, default=1, help='Compress level when encoding images. Default: 1.')
+    args = parser.parse_args()
+
+    try:
+        import tensorflow as tf
+    except Exception:
+        raise ImportError('You need to install tensorflow to read tfrecords.')
+
+    if args.dataset == 'ffhq':
+        convert_ffhq_tfrecords(
+            args.tf_file,
+            args.log_resolution,
+            args.save_root,
+            save_type=args.save_type,
+            compress_level=args.compress_level)
+    else:
+        convert_celeba_tfrecords(
+            args.tf_file,
+            args.log_resolution,
+            args.save_root,
+            save_type=args.save_type,
+            compress_level=args.compress_level)