| # Reading and Operating this file | |
| # ~ is basically Null/None/False e.t.c, you can either use ~ as the value, or comment out the line | |
| # A commented-out line's value is typically it's default but is disabled due to it being commented out | |
| # When a line starts with two #'s (comment line starting with #), then when un-commenting, it's intended to be disabled | |
| # for example, under network_G everything but ESRGAN's values are commented-out, so to show which should still be | |
| # enabled/disabled, a double-comment is used | |
| name: sollevante # put "debug" in the name to enable debug mode, only enable to test code works!!! | |
| use_tb_logger: true | |
| model: srragan # srragan | sr | srgan | ppon | asrragan | |
| scale: 4 | |
| gpu_ids: | |
| - 0 | |
| datasets: | |
| train: | |
| name: sollevante-train | |
| mode: LRHR | |
| # high resolution / ground truth images | |
| dataroot_HR: | |
| - /mnt/8tb-hdd-1/datasets/sollevante/hr/train | |
| # low resolution images. If there are missing LR images, they will be generated on the fly from HR | |
| dataroot_LR: | |
| - /mnt/8tb-hdd-1/datasets/sollevante/lr/train | |
| subset_file: ~ | |
| use_shuffle: true | |
| znorm: false # true | false // To normalize images in [-1, 1] range. Default = None (range [0,1]). Can use with activation function like tanh. | |
| n_workers: 8 # 0 to disable CPU multithreading, or an integrer representing CPU threads to use for dataloading | |
| batch_size: 32 | |
| HR_size: 128 #patch size. Default: 128. Needs to be coordinated with the patch size of the features network | |
| # Color space conversion: "color" for both LR and HR, "color_LR" for LR independently, "color_HR" for HR independently | |
| #color: y # remove for no conversion (RGB) | "y" for Y in YCbCr | "gray" to convert RGB to grayscale | "RGB" to convert gray to RGB | |
| #color_LR: y # remove for no conversion (RGB) | "y" for Y in YCbCr | "gray" to convert RGB to grayscale | "RGB" to convert gray to RGB | |
| #color_HR: y # remove for no conversion (RGB) | "y" for Y in YCbCr | "gray" to convert RGB to grayscale | "RGB" to convert gray to RGB | |
| # LR and HR modifiers. Random flip LR and HR or ignore provided LRs and generate new ones on the fly with defined probability: | |
| #rand_flip_LR_HR: false # flip LR and HR during training. | |
| #flip_chance: 0.05 # Example: 0.05 = 5% chance of LR and HR flipping during training. | |
| #aug_downscale: 0.2 # Example: 0.6 = 60% chance of generating LR on the fly, even if LR dataset exists | |
| # If manually configuring on the fly generation of LR: (else, it will automatically default to Matlab-like downscale algorithm (777) when/if required | |
| lr_downscale: true | |
| lr_downscale_types: | |
| # select from [0,1,2,3,4,5,777] where each number is: (scale algorithm) | |
| # - cv2.INTER_NEAREST | |
| # - cv2.INTER_LINEAR | |
| # - cv2.INTER_CUBIC | |
| # - cv2.INTER_AREA | |
| # - cv2.INTER_LANCZOS4 | |
| # - cv2.INTER_LINEAR_EXACT | |
| # - matlab.bicubic | |
| - 1 | |
| - 2 | |
| - 777 | |
| # Rotations augmentations: | |
| use_flip: true # flip lr with hr and hr with lr | |
| use_rot: true # rotate images in 90 degree angles, | |
| hr_rrot: false # rotate images in random degrees between -45 and 45 | |
| # Noise and blur augmentations: | |
| lr_blur: false | |
| lr_blur_types: | |
| # select from: average | box | gaussian | bilateral | clean, `median` and `motion` aren't working yet | |
| - gaussian | |
| - clean | |
| - clean | |
| - clean | |
| lr_noise: false | |
| lr_noise_types: | |
| # select from: gaussian | JPEG | quantize | poisson | dither | s&p | speckle | clean | |
| - gaussian | |
| - clean | |
| - clean | |
| - clean | |
| - clean | |
| lr_noise2: false | |
| lr_noise_types2: | |
| # select from: gaussian | JPEG | quantize | poisson | dither | s&p | speckle | clean | |
| - dither | |
| - dither | |
| - clean | |
| - clean | |
| hr_noise: false | |
| hr_noise_types: | |
| # select from: gaussian | JPEG | quantize | poisson | dither | s&p | speckle | clean | |
| - gaussian | |
| - clean | |
| - clean | |
| - clean | |
| - clean | |
| # Color augmentations | |
| #lr_fringes: true | |
| #lr_fringes_chance: 0.4 | |
| #auto_levels: HR # "HR" | "LR" | "Both" //add auto levels to the images to expand dynamic range. Can use with SPL loss or (MS)-SSIM. | |
| #rand_auto_levels: 0.7 # Example: 0.4 = 40% chance of adding auto levels to images on the fly | |
| #unsharp_mask: true # add a un-sharpening mask to HR images. Can work well together with the HFEN loss function. | |
| #rand_unsharp: 1 # Example: 0.5 = 50% chance of adding un-sharpening mask to HR images on the fly | |
| # Augmentations for classification or (maybe) in-painting networks: | |
| #lr_cutout: false | |
| #lr_erasing: false | |
| val: | |
| name: sollevante-val | |
| mode: LRHR | |
| dataroot_HR: | |
| - /mnt/8tb-hdd-1/datasets/sollevante/hr/val | |
| dataroot_LR: | |
| - /mnt/8tb-hdd-1/datasets/sollevante/lr/val | |
| znorm: false # true | false // To normalize images in [-1, 1] range. Default = None (range [0,1]). Can use with activation function like tanh. | |
| # Color space conversion: "color" for both LR and HR, "color_LR" for LR independently, "color_HR" for HR independently | |
| #color: y # remove for no conversion (RGB) | "y" for Y in YCbCr | "gray" to convert RGB to grayscale | "RGB" to convert gray to RGB | |
| #color_LR: y # remove for no conversion (RGB) | "y" for Y in YCbCr | "gray" to convert RGB to grayscale | "RGB" to convert gray to RGB | |
| #color_HR: y # remove for no conversion (RGB) | "y" for Y in YCbCr | "gray" to convert RGB to grayscale | "RGB" to convert gray to RGB | |
| #hr_crop: false | |
| lr_downscale: false | |
| lr_downscale_types: | |
| # select from [0,1,2,3,4,5] where each number is: (scale algorithm) | |
| # - cv2.INTER_NEAREST | |
| # - cv2.INTER_LINEAR | |
| # - cv2.INTER_CUBIC | |
| # - cv2.INTER_AREA | |
| # - cv2.INTER_LANCZOS4 | |
| # - cv2.INTER_LINEAR_EXACT | |
| - 0 | |
| - 1 | |
| path: | |
| root: /home/owner/github/BasicSR | |
| pretrain_model_G: ../experiments/pretrained_models/RRDB_PSNR_x4.pth | |
| resume_state: ../experiments/sollevante/training_state/140000.state | |
| # Generator options: | |
| network_G: | |
| # ESRGAN: | |
| which_model_G: RRDB_net # RRDB_net | sr_resnet | |
| norm_type: ~ | |
| mode: CNA | |
| nf: 64 # number of discriminator filters in the first convolution layer | |
| nb: 23 | |
| in_nc: 3 # of input image channels: 3 for RGB and 1 for grayscale | |
| out_nc: 3 # of output image channels: 3 for RGB and 1 for grayscale | |
| gc: 32 | |
| group: 1 | |
| convtype: Conv2D # convolution, Conv2D | PartialConv2D | |
| net_act: leakyrelu # network activation, swish | leakyrelu | |
| #finalact: tanh # in testing, activation function to make outputs fit in [-1,1] range. Coordinate with znorm | |
| # ASRGAN: | |
| #which_model_G: asr_resnet # asr_resnet | asr_cnn | |
| #nf: 64 | |
| # PPON: | |
| #which_model_G: ppon | |
| ##norm_type: ~ | |
| #mode: CNA | |
| #nf: 64 | |
| #nb: 24 | |
| #in_nc: 3 | |
| #out_nc: 3 | |
| ##gc: 32 | |
| #group: 1 | |
| ##convtype: Conv2D # Conv2D | PartialConv2D | |
| # SRGAN: | |
| #which_model_G: sr_resnet # RRDB_net | sr_resnet | |
| #norm_type: ~ | |
| #mode: CNA | |
| #nf: 64 | |
| #nb: 16 | |
| #in_nc: 3 | |
| #out_nc: 3 | |
| # SR: | |
| #which_model_G: RRDB_net # RRDB_net | sr_resnet | |
| #norm_type: ~ | |
| #mode: CNA | |
| #nf: 64 | |
| #nb: 23 | |
| #in_nc: 3 | |
| #out_nc: 3 | |
| #gc: 32 | |
| #group: 1 | |
| # Discriminator options: | |
| network_D: | |
| # ESRGAN (default) | PPON | |
| which_model_D: discriminator_vgg_128 # discriminator_vgg_128 | discriminator_vgg | |
| norm_type: batch | |
| act_type: leakyrelu | |
| mode: CNA # CNA | NAC | |
| nf: 64 | |
| in_nc: 3 | |
| # ASRGAN (feature extraction) | |
| #which_model_D: discriminator_vgg_128_fea | |
| #norm_type: batch | |
| #act_type: leakyrelu | |
| #mode: CNA # CNA | NAC | |
| #nf: 64 | |
| #in_nc: 3 | |
| #spectral_norm: true | |
| #self_attention: true | |
| #max_pool: true | |
| #poolsize: 4 | |
| # Schedulers options: | |
| train: | |
| lr_G: !!float 1e-4 # starting lr_g | |
| weight_decay_G: 0 | |
| beta1_G: 0.9 | |
| lr_D: !!float 1e-4 # starting lr_d | |
| weight_decay_D: 0 | |
| beta1_D: 0.9 | |
| # For MultiStepLR (ESRGAN, default):, | |
| lr_scheme: MultiStepLR | |
| lr_steps: | |
| - 50000 | |
| - 100000 | |
| - 200000 | |
| - 300000 | |
| lr_gamma: 0.5 # lr change at every step (multiplied by) | |
| # For StepLR_Restart (PPON): | |
| # lr_gamma: 0.9 # lr change at every step (multiplied by) | |
| # lr_scheme: StepLR_Restart # MultiStepLR | MultiStepLR_Restart | StepLR | StepLR_Restart | CosineAnnealingLR_Restart | |
| # lr_step_sizes: # Steps for each restart for "StepLR_Restart" | |
| # - 200 | |
| # - 100 | |
| # - 250 | |
| # restarts: # Restart iterations for MultiStepLR_Restart | StepLR_Restart | CosineAnnealingLR_Restart | |
| # - 138000 | |
| # - 172500 | |
| # restart_weights: # lr_() * each weight in "restart_weights" for each restart in "restarts" | |
| # - 1 | |
| # - 0.5 | |
| # - 0.5 | |
| ## clear_state: true | |
| # For MultiStepLR_Restart: | |
| # lr_gamma: 0.9 | |
| # lr_scheme: MultiStepLR_Restart # MultiStepLR | MultiStepLR_Restart | StepLR | StepLR_Restart | CosineAnnealingLR_Restart | |
| # lr_steps: # For "MultiStepLR" and "MultiStepLR_Restart" | |
| # - 34500 | |
| # - 69000 | |
| # - 103500 | |
| # - 155250 | |
| # - 189750 | |
| # - 241500 | |
| # restarts: # Restart iterations for MultiStepLR_Restart | StepLR_Restart | CosineAnnealingLR_Restart | |
| # - 138000 | |
| # - 172500 | |
| # restart_weights: # lr_() * each weight in "restart_weights" for each restart in "restarts" | |
| # - 0.5 | |
| # - 0.5 | |
| ## clear_state: true | |
| # Losses: | |
| pixel_criterion: l1 # pixel loss, l1 | l2 | cb | elastic | relativel1 | l1cosinesim | |
| pixel_weight: !!float 1e-2 # 1e-2 | 1 | |
| feature_criterion: l1 # feature loss (VGG feature network), l1 | l2 | cb | elastic | |
| feature_weight: 1 | |
| #dis_feature_criterion: l1 # discriminator feature loss, asrragan only, l1 | l2 | cb | elastic | |
| #dis_feature_weight: 1 | |
| #hfen_criterion: l1 # helps in deblurring and finding edges/lines, l1 | l2 | rel_l1 | rel_l2 | |
| #hfen_weight: !!float 1e-1 | |
| #tv_type: normal # helps in denoising, reducing upscale artefacts | |
| #tv_weight: !!float 1e-6 | |
| #tv_norm: 1 # 1 == l1 or 2 == l2. Change "tv_weight" so the l_g_tv is around 1e-02 | |
| #ssim_type: ms-ssim # ssim | ms-ssim # helps to maintain luminance, contrast and covariance between SR and HR | |
| #ssim_weight: 1 | |
| gan_type: vanilla # vanilla | basic | |
| gan_weight: !!float 5e-3 | |
| #lpips_weight: 1 # perceptual loss | |
| #lpips_type: net-lin # net-lin | net | |
| #lpips_net: squeeze # vgg | alex | squeeze | |
| #spl_weight: !!float 1e-3 # SPL loss function. note: needs to add a cap in the generator (finalcap) or it becomes unstable | |
| #spl_type: spl # spl | gpl | cpl | |
| # for wgan-gp: | |
| #D_update_ratio: 1 | |
| #D_init_iters: 0 | |
| #gp_weigth: 10 | |
| # For PPON: | |
| #train_phase: 1 # Training starting phase, can skip the first phases | |
| #phase1_s: 100 # -1 to skip. Need to coordinate with the LR steps. //COBranch: lr = 2e−4, decreased by the factor of 2 for every 1000 epochs (1.38e+5 iterations) 138k | |
| #phase2_s: 200 # -1 to skip. Need to coordinate with the LR steps. //SOBranch: λ = 1e+3 (?), lr = 1e−4 and halved at every 250 epochs (3.45e+4iterations) 34.5k | |
| #phase3_s: 5000000 # -1 to skip. Need to coordinate with the LR steps. //POBranch: η = 5e−3, lr = 1e−4 and halved at every 250 epochs (3.45e+4iterations) 34.5k | |
| #phase4_s: 100100 | |
| # Other training options: | |
| #finalcap: tanh # Test. Cap Generator outputs to fit in: [-1,1] range ("tanh"), rescale tanh to [0,1] range ("scaltanh"), cap ("sigmoid") or clamp ("clamp") to [0,1] range. Default = None. Coordinate with znorm. | |
| #manual_seed: 0, # only set if you want reproducibility as it will incur a performance cut, 0 to 2^32-1 | |
| niter: !!float 5e5 | |
| val_freq: !!float 5e3 | |
| logger: | |
| print_freq: 200 | |
| save_checkpoint_freq: !!float 5e3 |