Delete utils_data

utils_data/crop_paired_data.py

@@ -1,73 +0,0 @@
-import os
-import sys
-sys.path.append(os.getcwd())
-import cv2
-
-import torch
-import torch.nn.functional as F
-from pytorch_lightning import seed_everything
-import argparse
-
-from basicsr.utils import FileClient, get_root_logger, imfrombytes, img2tensor
-from basicsr.utils import DiffJPEG, USMSharp
-parser = argparse.ArgumentParser()
-
-parser.add_argument("--save_dir", type=str, default='preset/datasets/train_datasets/training_for_seesr', help='the save path of the training dataset.')
-
-args = parser.parse_args()
-
-gt_path = os.path.join(args.save_dir, 'gt')
-lr_path = os.path.join(args.save_dir, 'lr')
-sr_bicubic_path = os.path.join(args.save_dir, 'sr_bicubic')
-print(gt_path)
-os.makedirs(gt_path, exist_ok=True)
-os.makedirs(lr_path, exist_ok=True)
-os.makedirs(sr_bicubic_path, exist_ok=True)
-hr_dir = '/media/ssd8T/wyw/Data/NTIRE2025/test/hr'
-lr_dir = '/media/ssd8T/wyw/Data/NTIRE2025/test/lr'
-hr_files = sorted(os.listdir(hr_dir))
-lr_files = sorted(os.listdir(lr_dir))
-usm_sharpener = USMSharp().cuda()
-step = 0
-for i, (hr_file, lr_file) in enumerate(zip(hr_files, lr_files)):
-    step += 1
-    print('process {} images...'.format(step))
-    
-    with open(os.path.join(hr_dir, hr_file), 'rb') as f:
-        img_bytes = f.read()
-    img_gt = imfrombytes(img_bytes, float32=True)
-    with open(os.path.join(lr_dir, lr_file), 'rb') as f:
-        img_bytes = f.read()
-    img_lr = imfrombytes(img_bytes, float32=True)
-    
-    h, w = img_gt.shape[0:2]
-    crop_pad_size = 512
-    if h < crop_pad_size or w < crop_pad_size:
-        pad_h = max(0, crop_pad_size - h)
-        pad_w = max(0, crop_pad_size - w)
-        img_gt = cv2.copyMakeBorder(img_gt, 0, pad_h, 0, pad_w, cv2.BORDER_REFLECT_101)
-
-    if img_gt.shape[0] > crop_pad_size or img_gt.shape[1] > crop_pad_size:
-        h, w = img_gt.shape[0:2]
-        top = 500
-        left = 250
-        img_gt = img_gt[top:top + crop_pad_size, left:left + crop_pad_size, ...]
-        img_lr = img_lr[top//4:top//4 + crop_pad_size//4, left//4:left//4 + crop_pad_size//4, ...]
-        
-    
-    img_gt = img2tensor([img_gt], bgr2rgb=True, float32=True)[0]
-    img_lr = img2tensor([img_lr], bgr2rgb=True, float32=True)[0]
-    img_gt = img_gt.unsqueeze(0).cuda() 
-    img_gt = usm_sharpener(img_gt).squeeze(0)
-    gt = torch.clamp(img_gt, 0, 1)
-    lr = torch.clamp(img_lr, 0, 1)
-    
-    sr_bicubic = F.interpolate(lr.unsqueeze(0), size=(gt.size(-2), gt.size(-1)), mode='bicubic',).squeeze(0)
-
-    lr_save_path =  os.path.join(lr_path,'{}.png'.format(str(step).zfill(7)))
-    gt_save_path =  os.path.join(gt_path, '{}.png'.format(str(step).zfill(7)))
-    sr_bicubic_save_path =  os.path.join(sr_bicubic_path, '{}.png'.format(str(step).zfill(7)))
-
-    cv2.imwrite(lr_save_path, 255*lr.detach().cpu().squeeze().permute(1,2,0).numpy()[..., ::-1])
-    cv2.imwrite(gt_save_path, 255*gt.detach().cpu().squeeze().permute(1,2,0).numpy()[..., ::-1])
-    cv2.imwrite(sr_bicubic_save_path, 255*sr_bicubic.detach().cpu().squeeze().permute(1,2,0).numpy()[..., ::-1])

utils_data/make_paired_data.py

@@ -1,330 +0,0 @@
-'''
- * SeeSR: Towards Semantics-Aware Real-World Image Super-Resolution 
- * Modified from diffusers by Rongyuan Wu
- * 24/12/2023
-'''
-import os
-import sys
-sys.path.append(os.getcwd())
-import cv2
-
-import torch
-import torch.nn.functional as F
-from pytorch_lightning import seed_everything
-    
-import argparse
-from basicsr.data.realesrgan_dataset import RealESRGANDataset
-from ram.models import ram
-from ram import inference_ram as inference
-
-parser = argparse.ArgumentParser()
-parser.add_argument("--gt_path", nargs='+', default=['PATH 1', 'PATH 2'], help='the path of high-resolution images')
-parser.add_argument("--save_dir", type=str, default='preset/datasets/train_datasets/training_for_seesr', help='the save path of the training dataset.')
-parser.add_argument("--start_gpu", type=int, default=1, help='if you have 5 GPUs, you can set it to 1/2/3/4/5 on five gpus for parallel processing., which will save your time. ')  
-parser.add_argument("--batch_size", type=int, default=10, help='smaller batch size means much time but more extensive degradation for making the training dataset.')  
-parser.add_argument("--epoch", type=int, default=1, help='decide how many epochs to create for the dataset.')
-args = parser.parse_args()
-
-print(f'====== START GPU: {args.start_gpu} =========')
-seed_everything(24+args.start_gpu*1000)
-
-from torchvision.transforms import Normalize, Compose
-args_training_dataset = {}
-
-# Please set your gt path here. If you have multi dirs, you can set it as ['PATH1', 'PATH2', 'PATH3', ...]
-args_training_dataset['gt_path'] = args.gt_path
-
-#################### REALESRGAN SETTING ###########################
-args_training_dataset['queue_size'] = 160
-args_training_dataset['crop_size'] =  512
-args_training_dataset['io_backend'] = {}
-args_training_dataset['io_backend']['type'] = 'disk'
-
-# args_training_dataset['blur_kernel_size'] = 21
-# args_training_dataset['kernel_list'] = ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
-# args_training_dataset['kernel_prob'] = [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
-# args_training_dataset['sinc_prob'] = 0.1
-# args_training_dataset['blur_sigma'] = [0.2, 3]
-# args_training_dataset['betag_range'] = [0.5, 4]
-# args_training_dataset['betap_range'] = [1, 2]
-
-args_training_dataset['blur_kernel_size'] = 2
-args_training_dataset['kernel_list'] = ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
-args_training_dataset['kernel_prob'] = [0.6, 0.2, 0.1, 0.05, 0.05, 0]
-args_training_dataset['sinc_prob'] = 0.05
-args_training_dataset['blur_sigma'] = [0.01, 0.02]
-args_training_dataset['betag_range'] = [0.01, 0.02]
-args_training_dataset['betap_range'] = [0.01, 0.02]
-
-# args_training_dataset['blur_kernel_size2'] = 11
-# args_training_dataset['kernel_list2'] = ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
-# args_training_dataset['kernel_prob2'] = [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
-# args_training_dataset['sinc_prob2'] = 0.1
-# args_training_dataset['blur_sigma2'] = [0.2, 1.5]
-# args_training_dataset['betag_range2'] = [0.5, 4.0]
-# args_training_dataset['betap_range2'] = [1, 2]
-
-args_training_dataset['blur_kernel_size2'] = 2
-args_training_dataset['kernel_list2'] = ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
-args_training_dataset['kernel_prob2'] = [0.6, 0.2, 0.1, 0.05, 0.05, 0]
-args_training_dataset['sinc_prob2'] = 0.05
-args_training_dataset['blur_sigma2'] = [0.01, 0.02]
-args_training_dataset['betag_range2'] = [0.01, 0.02]
-args_training_dataset['betap_range2'] = [0.01, 0.02]
-
-args_training_dataset['final_sinc_prob'] = 0.4
-
-args_training_dataset['use_hflip'] = True
-args_training_dataset['use_rot'] = False
-
-train_dataset = RealESRGANDataset(args_training_dataset)
-batch_size = args.batch_size
-train_dataloader = torch.utils.data.DataLoader(
-    train_dataset,
-    shuffle=False,
-    batch_size=batch_size,
-    num_workers=4,
-    drop_last=True,
-)
-
-#################### REALESRGAN SETTING ###########################
-args_degradation = {}
-# the first degradation process
-# args_degradation['resize_prob'] = [0.2, 0.7, 0.1]  # up, down, keep
-# args_degradation['resize_range'] = [0.15, 1.5]
-# args_degradation['gaussian_noise_prob'] = 0.5
-# args_degradation['noise_range'] = [1, 30]
-# args_degradation['poisson_scale_range'] = [0.05, 3.0]
-# args_degradation['gray_noise_prob'] = 0.4
-# args_degradation['jpeg_range'] = [30, 95]
-
-args_degradation['resize_prob'] = [0.4, 0.4, 0.2]  # up, down, keep
-args_degradation['resize_range'] = [0.95, 1.05]
-args_degradation['gaussian_noise_prob'] = 0.03
-args_degradation['noise_range'] = [0.01, 0.03]
-args_degradation['poisson_scale_range'] = [0.01, 0.1]
-args_degradation['gray_noise_prob'] = 0.1
-args_degradation['jpeg_range'] = [80, 100]
-
-# the second degradation process
-# args_degradation['second_blur_prob'] = 0.8
-# args_degradation['resize_prob2'] = [0.3, 0.4, 0.3]  # up, down, keep
-# args_degradation['resize_range2'] = [0.3, 1.2]
-# args_degradation['gaussian_noise_prob2'] = 0.5
-# args_degradation['noise_range2'] = [1, 25]
-# args_degradation['poisson_scale_range2'] = [0.05, 2.5]
-# args_degradation['gray_noise_prob2'] = 0.4
-# args_degradation['jpeg_range2'] = [30, 95]
-
-args_degradation['second_blur_prob'] = 0.1
-args_degradation['resize_prob2'] = [0.4, 0.4, 0.2]  # up, down, keep
-args_degradation['resize_range2'] = [0.95, 1.05]
-args_degradation['gaussian_noise_prob2'] = 0.03
-args_degradation['noise_range2'] = [0.01, 0.03]
-args_degradation['poisson_scale_range2'] = [0.01, 0.1]
-args_degradation['gray_noise_prob2'] = 0.1
-args_degradation['jpeg_range2'] = [80,100]
-
-args_degradation['gt_size']= 512
-args_degradation['no_degradation_prob']= 0.01
-
-
-from basicsr.utils import DiffJPEG, USMSharp
-from basicsr.utils.img_process_util import filter2D
-from basicsr.data.transforms import paired_random_crop, triplet_random_crop
-from basicsr.data.degradations import random_add_gaussian_noise_pt, random_add_poisson_noise_pt, random_add_speckle_noise_pt, random_add_saltpepper_noise_pt, bivariate_Gaussian
-import random
-import torch.nn.functional as F
-
-def realesrgan_degradation(batch,  args_degradation, use_usm=True, sf=4, resize_lq=True):
-    jpeger = DiffJPEG(differentiable=False).cuda()
-    usm_sharpener = USMSharp().cuda()  # do usm sharpening
-    im_gt = batch['gt'].cuda()
-    im_gt_ori = im_gt
-    if use_usm:
-        im_gt = usm_sharpener(im_gt)
-    im_gt_ori = im_gt_ori.to(memory_format=torch.contiguous_format).float()
-    im_gt = im_gt.to(memory_format=torch.contiguous_format).float()
-    kernel1 = batch['kernel1'].cuda()
-    kernel2 = batch['kernel2'].cuda()
-    sinc_kernel = batch['sinc_kernel'].cuda()
-
-    ori_h, ori_w = im_gt.size()[2:4]
-
-    # ----------------------- The first degradation process ----------------------- #
-    
-    if random.random()<0.5:
-        out = im_gt_ori
-        out = F.interpolate(
-                out,
-                size=(ori_h // sf,ori_w // sf),
-                mode='bicubic',
-                )
-    else:
-        print('degration')
-        # blur
-        out = filter2D(im_gt_ori, kernel1)
-        # random resize
-        updown_type = random.choices(
-                ['up', 'down', 'keep'],
-                args_degradation['resize_prob'],
-                )[0]
-        if updown_type == 'up':
-            scale = random.uniform(1, args_degradation['resize_range'][1])
-        elif updown_type == 'down':
-            scale = random.uniform(args_degradation['resize_range'][0], 1)
-        else:
-            scale = 1
-        mode = random.choice(['area', 'bilinear', 'bicubic'])
-        out = F.interpolate(out, scale_factor=scale, mode=mode)
-        # add noise
-        gray_noise_prob = args_degradation['gray_noise_prob']
-        if random.random() < args_degradation['gaussian_noise_prob']:
-            out = random_add_gaussian_noise_pt(
-                out,
-                sigma_range=args_degradation['noise_range'],
-                clip=True,
-                rounds=False,
-                gray_prob=gray_noise_prob,
-                )
-        else:
-            out = random_add_poisson_noise_pt(
-                out,
-                scale_range=args_degradation['poisson_scale_range'],
-                gray_prob=gray_noise_prob,
-                clip=True,
-                rounds=False)
-        # JPEG compression
-        jpeg_p = out.new_zeros(out.size(0)).uniform_(*args_degradation['jpeg_range'])
-        out = torch.clamp(out, 0, 1)  # clamp to [0, 1], otherwise JPEGer will result in unpleasant artifacts
-        out = jpeger(out, quality=jpeg_p)
-
-        # ----------------------- The second degradation process ----------------------- #
-        # blur
-        if random.random() < args_degradation['second_blur_prob']:
-            out = filter2D(out, kernel2)
-        # random resize
-        updown_type = random.choices(
-                ['up', 'down', 'keep'],
-                args_degradation['resize_prob2'],
-                )[0]
-        if updown_type == 'up':
-            scale = random.uniform(1, args_degradation['resize_range2'][1])
-        elif updown_type == 'down':
-            scale = random.uniform(args_degradation['resize_range2'][0], 1)
-        else:
-            scale = 1
-        mode = random.choice(['area', 'bilinear', 'bicubic'])
-        out = F.interpolate(
-                out,
-                size=(int(ori_h / sf * scale),
-                        int(ori_w / sf * scale)),
-                mode=mode,
-                )
-        # add noise
-        gray_noise_prob = args_degradation['gray_noise_prob2']
-        if random.random() < args_degradation['gaussian_noise_prob2']:
-            out = random_add_gaussian_noise_pt(
-                out,
-                sigma_range=args_degradation['noise_range2'],
-                clip=True,
-                rounds=False,
-                gray_prob=gray_noise_prob,
-                )
-        else:
-            out = random_add_poisson_noise_pt(
-                out,
-                scale_range=args_degradation['poisson_scale_range2'],
-                gray_prob=gray_noise_prob,
-                clip=True,
-                rounds=False,
-                )
-
-        # JPEG compression + the final sinc filter
-        # We also need to resize images to desired sizes. We group [resize back + sinc filter] together
-        # as one operation.
-        # We consider two orders:
-        #   1. [resize back + sinc filter] + JPEG compression
-        #   2. JPEG compression + [resize back + sinc filter]
-        # Empirically, we find other combinations (sinc + JPEG + Resize) will introduce twisted lines.
-        if random.random() < 0.5:
-            # resize back + the final sinc filter
-            mode = random.choice(['area', 'bilinear', 'bicubic'])
-            out = F.interpolate(
-                    out,
-                    size=(ori_h // sf,
-                            ori_w // sf),
-                    mode=mode,
-                    )
-            out = filter2D(out, sinc_kernel)
-            # JPEG compression
-            jpeg_p = out.new_zeros(out.size(0)).uniform_(*args_degradation['jpeg_range2'])
-            out = torch.clamp(out, 0, 1)
-            out = jpeger(out, quality=jpeg_p)
-        else:
-            # JPEG compression
-            jpeg_p = out.new_zeros(out.size(0)).uniform_(*args_degradation['jpeg_range2'])
-            out = torch.clamp(out, 0, 1)
-            out = jpeger(out, quality=jpeg_p)
-            # resize back + the final sinc filter
-            mode = random.choice(['area', 'bilinear', 'bicubic'])
-            out = F.interpolate(
-                    out,
-                    size=(ori_h // sf,
-                            ori_w // sf),
-                    mode=mode,
-                    )
-            out = filter2D(out, sinc_kernel)
-
-    # clamp and round
-    im_lq = torch.clamp(out, 0, 1.0)
-
-    # random crop
-    gt_size = args_degradation['gt_size']
-    im_gt, im_lq = paired_random_crop(im_gt, im_lq, gt_size, sf)
-    lq, gt = im_lq, im_gt
-
-
-    gt = torch.clamp(gt, 0, 1)
-    lq = torch.clamp(lq, 0, 1)
-
-    return lq, gt
-
-
-root_path = args.save_dir
-gt_path = os.path.join(root_path, 'gt')
-lr_path = os.path.join(root_path, 'lr')
-sr_bicubic_path = os.path.join(root_path, 'sr_bicubic')
-os.makedirs(gt_path, exist_ok=True)
-os.makedirs(lr_path, exist_ok=True)
-os.makedirs(sr_bicubic_path, exist_ok=True)
-
-
-epochs = args.epoch
-step = len(train_dataset) * epochs * args.start_gpu
-step = 0
-with torch.no_grad():
-    for epoch in range(epochs):
-        for num_batch, batch in enumerate(train_dataloader):
-            lr_batch, gt_batch = realesrgan_degradation(batch, args_degradation=args_degradation)
-            sr_bicubic_batch = F.interpolate(lr_batch, size=(gt_batch.size(-2), gt_batch.size(-1)), mode='bicubic',)
-
-            for i in range(batch_size):
-                step += 1
-                print('process {} images...'.format(step))
-                lr = lr_batch[i, ...]
-                gt = gt_batch[i, ...]
-                sr_bicubic = sr_bicubic_batch[i, ...]
-
-                lr_save_path =  os.path.join(lr_path,'{}.png'.format(str(step).zfill(7)))
-                gt_save_path =  os.path.join(gt_path, '{}.png'.format(str(step).zfill(7)))
-                sr_bicubic_save_path =  os.path.join(sr_bicubic_path, '{}.png'.format(str(step).zfill(7)))
-
-                cv2.imwrite(lr_save_path, 255*lr.detach().cpu().squeeze().permute(1,2,0).numpy()[..., ::-1])
-                cv2.imwrite(gt_save_path, 255*gt.detach().cpu().squeeze().permute(1,2,0).numpy()[..., ::-1])
-                cv2.imwrite(sr_bicubic_save_path, 255*sr_bicubic.detach().cpu().squeeze().permute(1,2,0).numpy()[..., ::-1])
-
-
-            del lr_batch, gt_batch, sr_bicubic_batch
-            torch.cuda.empty_cache()
-    

utils_data/make_paired_data_DAPE.py

@@ -1,284 +0,0 @@
-'''
- * SeeSR: Towards Semantics-Aware Real-World Image Super-Resolution 
- * Modified from diffusers by Rongyuan Wu
- * 24/12/2023
-'''
-import os
-import cv2
-import torch
-import torch.nn.functional as F
-from pytorch_lightning import seed_everything
-import argparse
-import sys
-sys.path.append(os.getcwd())
-
-from basicsr.data.realesrgan_dataset import RealESRGANDataset
-from dataloaders.simple_dataset import SimpleDataset
-
-from ram.models import ram
-from ram import inference_ram as inference
-
-parser = argparse.ArgumentParser()
-parser.add_argument("--gt_path", nargs='+', default=['PATH 1', 'PATH 2'], help='the path of high-resolution images')
-parser.add_argument("--save_dir", type=str, default='preset/datasets/train_datasets/training_for_dape', help='the save path of the training dataset.')
-parser.add_argument("--start_gpu", type=int, default=1, help='if you have 5 GPUs, you can set it to 1/2/3/4/5 on five gpus for parallel processing., which will save your time. ')  
-parser.add_argument("--batch_size", type=int, default=10, help='smaller batch size means much time but more extensive degradation for making the training dataset.')  
-parser.add_argument("--epoch", type=int, default=1, help='decide how many epochs to create for the dataset.')
-args = parser.parse_args()
-
-print(f'====== START GPU: {args.start_gpu} =========')
-seed_everything(24+args.start_gpu*1000)
-
-from torchvision.transforms import Normalize, Compose
-args_training_dataset = {}
-
-# Please set your gt path here. If you have multi dirs, you can set it as ['PATH1', 'PATH2', 'PATH3', ...]
-args_training_dataset['gt_path'] = args.gt_path
-    
-#################### REALESRGAN SETTING ###########################
-args_training_dataset['queue_size'] = 160
-args_training_dataset['crop_size'] =  512
-args_training_dataset['io_backend'] = {}
-args_training_dataset['io_backend']['type'] = 'disk'
-
-args_training_dataset['blur_kernel_size'] = 21
-args_training_dataset['kernel_list'] = ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
-args_training_dataset['kernel_prob'] = [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
-args_training_dataset['sinc_prob'] = 0.1
-args_training_dataset['blur_sigma'] = [0.2, 3]
-args_training_dataset['betag_range'] = [0.5, 4]
-args_training_dataset['betap_range'] = [1, 2]
-
-args_training_dataset['blur_kernel_size2'] = 11
-args_training_dataset['kernel_list2'] = ['iso', 'aniso', 'generalized_iso', 'generalized_aniso', 'plateau_iso', 'plateau_aniso']
-args_training_dataset['kernel_prob2'] = [0.45, 0.25, 0.12, 0.03, 0.12, 0.03]
-args_training_dataset['sinc_prob2'] = 0.1
-args_training_dataset['blur_sigma2'] = [0.2, 1.5]
-args_training_dataset['betag_range2'] = [0.5, 4.0]
-args_training_dataset['betap_range2'] = [1, 2]
-
-args_training_dataset['final_sinc_prob'] = 0.8
-
-args_training_dataset['use_hflip'] = True
-args_training_dataset['use_rot'] = False
-
-train_dataset = SimpleDataset(args_training_dataset, fix_size=512)
-batch_size = args.batch_size
-train_dataloader = torch.utils.data.DataLoader(
-    train_dataset,
-    shuffle=False,
-    batch_size=batch_size,
-    num_workers=11,
-    drop_last=True,
-)
-
-
-#################### REALESRGAN SETTING ###########################
-args_degradation = {}
-# the first degradation process
-args_degradation['resize_prob'] = [0.2, 0.7, 0.1]  # up, down, keep
-args_degradation['resize_range'] = [0.15, 1.5]
-args_degradation['gaussian_noise_prob'] = 0.5
-args_degradation['noise_range'] = [1, 30]
-args_degradation['poisson_scale_range'] = [0.05, 3.0]
-args_degradation['gray_noise_prob'] = 0.4
-args_degradation['jpeg_range'] = [30, 95]
-
-# the second degradation process
-args_degradation['second_blur_prob'] = 0.8
-args_degradation['resize_prob2'] = [0.3, 0.4, 0.3]  # up, down, keep
-args_degradation['resize_range2'] = [0.3, 1.2]
-args_degradation['gaussian_noise_prob2'] = 0.5
-args_degradation['noise_range2'] = [1, 25]
-args_degradation['poisson_scale_range2'] = [0.05, 2.5]
-args_degradation['gray_noise_prob2'] = 0.4
-args_degradation['jpeg_range2'] = [30, 95]
-
-args_degradation['gt_size']= 512
-args_degradation['no_degradation_prob']= 0.01
-
-from basicsr.utils import DiffJPEG, USMSharp
-from basicsr.utils.img_process_util import filter2D
-from basicsr.data.transforms import paired_random_crop, triplet_random_crop
-from basicsr.data.degradations import random_add_gaussian_noise_pt, random_add_poisson_noise_pt, random_add_speckle_noise_pt, random_add_saltpepper_noise_pt, bivariate_Gaussian
-import random
-import torch.nn.functional as F
-
-def realesrgan_degradation(batch,  args_degradation, use_usm=True, sf=4, resize_lq=True):
-    jpeger = DiffJPEG(differentiable=False).cuda()
-    usm_sharpener = USMSharp().cuda()  # do usm sharpening
-    im_gt = batch['gt'].cuda()
-    if use_usm:
-        im_gt = usm_sharpener(im_gt)
-    im_gt = im_gt.to(memory_format=torch.contiguous_format).float()
-    kernel1 = batch['kernel1'].cuda()
-    kernel2 = batch['kernel2'].cuda()
-    sinc_kernel = batch['sinc_kernel'].cuda()
-
-    ori_h, ori_w = im_gt.size()[2:4]
-
-    # ----------------------- The first degradation process ----------------------- #
-    # blur
-    out = filter2D(im_gt, kernel1)
-    # random resize
-    updown_type = random.choices(
-            ['up', 'down', 'keep'],
-            args_degradation['resize_prob'],
-            )[0]
-    if updown_type == 'up':
-        scale = random.uniform(1, args_degradation['resize_range'][1])
-    elif updown_type == 'down':
-        scale = random.uniform(args_degradation['resize_range'][0], 1)
-    else:
-        scale = 1
-    mode = random.choice(['area', 'bilinear', 'bicubic'])
-    out = F.interpolate(out, scale_factor=scale, mode=mode)
-    # add noise
-    gray_noise_prob = args_degradation['gray_noise_prob']
-    if random.random() < args_degradation['gaussian_noise_prob']:
-        out = random_add_gaussian_noise_pt(
-            out,
-            sigma_range=args_degradation['noise_range'],
-            clip=True,
-            rounds=False,
-            gray_prob=gray_noise_prob,
-            )
-    else:
-        out = random_add_poisson_noise_pt(
-            out,
-            scale_range=args_degradation['poisson_scale_range'],
-            gray_prob=gray_noise_prob,
-            clip=True,
-            rounds=False)
-    # JPEG compression
-    jpeg_p = out.new_zeros(out.size(0)).uniform_(*args_degradation['jpeg_range'])
-    out = torch.clamp(out, 0, 1)  # clamp to [0, 1], otherwise JPEGer will result in unpleasant artifacts
-    out = jpeger(out, quality=jpeg_p)
-
-    # ----------------------- The second degradation process ----------------------- #
-    # blur
-    if random.random() < args_degradation['second_blur_prob']:
-        out = filter2D(out, kernel2)
-    # random resize
-    updown_type = random.choices(
-            ['up', 'down', 'keep'],
-            args_degradation['resize_prob2'],
-            )[0]
-    if updown_type == 'up':
-        scale = random.uniform(1, args_degradation['resize_range2'][1])
-    elif updown_type == 'down':
-        scale = random.uniform(args_degradation['resize_range2'][0], 1)
-    else:
-        scale = 1
-    mode = random.choice(['area', 'bilinear', 'bicubic'])
-    out = F.interpolate(
-            out,
-            size=(int(ori_h / sf * scale),
-                    int(ori_w / sf * scale)),
-            mode=mode,
-            )
-    # add noise
-    gray_noise_prob = args_degradation['gray_noise_prob2']
-    if random.random() < args_degradation['gaussian_noise_prob2']:
-        out = random_add_gaussian_noise_pt(
-            out,
-            sigma_range=args_degradation['noise_range2'],
-            clip=True,
-            rounds=False,
-            gray_prob=gray_noise_prob,
-            )
-    else:
-        out = random_add_poisson_noise_pt(
-            out,
-            scale_range=args_degradation['poisson_scale_range2'],
-            gray_prob=gray_noise_prob,
-            clip=True,
-            rounds=False,
-            )
-
-    # JPEG compression + the final sinc filter
-    # We also need to resize images to desired sizes. We group [resize back + sinc filter] together
-    # as one operation.
-    # We consider two orders:
-    #   1. [resize back + sinc filter] + JPEG compression
-    #   2. JPEG compression + [resize back + sinc filter]
-    # Empirically, we find other combinations (sinc + JPEG + Resize) will introduce twisted lines.
-    if random.random() < 0.5:
-        # resize back + the final sinc filter
-        mode = random.choice(['area', 'bilinear', 'bicubic'])
-        out = F.interpolate(
-                out,
-                size=(ori_h // sf,
-                        ori_w // sf),
-                mode=mode,
-                )
-        out = filter2D(out, sinc_kernel)
-        # JPEG compression
-        jpeg_p = out.new_zeros(out.size(0)).uniform_(*args_degradation['jpeg_range2'])
-        out = torch.clamp(out, 0, 1)
-        out = jpeger(out, quality=jpeg_p)
-    else:
-        # JPEG compression
-        jpeg_p = out.new_zeros(out.size(0)).uniform_(*args_degradation['jpeg_range2'])
-        out = torch.clamp(out, 0, 1)
-        out = jpeger(out, quality=jpeg_p)
-        # resize back + the final sinc filter
-        mode = random.choice(['area', 'bilinear', 'bicubic'])
-        out = F.interpolate(
-                out,
-                size=(ori_h // sf,
-                        ori_w // sf),
-                mode=mode,
-                )
-        out = filter2D(out, sinc_kernel)
-
-    # clamp and round
-    im_lq = torch.clamp(out, 0, 1.0)
-
-    # random crop
-    gt_size = args_degradation['gt_size']
-    im_gt, im_lq = paired_random_crop(im_gt, im_lq, gt_size, sf)
-    lq, gt = im_lq, im_gt
-
-
-    gt = torch.clamp(gt, 0, 1)
-    lq = torch.clamp(lq, 0, 1)
-
-    return lq, gt
-
-
-root_path = args.save_dir
-gt_path = os.path.join(root_path, 'gt')
-lr_path = os.path.join(root_path, 'lr')
-sr_bicubic_path = os.path.join(root_path, 'sr_bicubic')
-os.makedirs(gt_path, exist_ok=True)
-os.makedirs(lr_path, exist_ok=True)
-os.makedirs(sr_bicubic_path, exist_ok=True)
-
-epochs = args.epoch
-step = len(train_dataset) * epochs * args.start_gpu
-with torch.no_grad():
-    for epoch in range(epochs):
-        for num_batch, batch in enumerate(train_dataloader):
-            lr_batch, gt_batch = realesrgan_degradation(batch, args_degradation=args_degradation)
-            sr_bicubic_batch = F.interpolate(lr_batch, size=(gt_batch.size(-2), gt_batch.size(-1)), mode='bicubic',)
-
-            for i in range(batch_size):
-                step += 1
-                print('process {} images...'.format(step))
-                lr = lr_batch[i, ...]
-                gt = gt_batch[i, ...]
-                sr_bicubic = sr_bicubic_batch[i, ...]
-
-                lr_save_path =  os.path.join(lr_path,'{}.png'.format(str(step).zfill(7)))
-                gt_save_path =  os.path.join(gt_path, '{}.png'.format(str(step).zfill(7)))
-                sr_bicubic_save_path =  os.path.join(sr_bicubic_path, '{}.png'.format(str(step).zfill(7)))
-
-                cv2.imwrite(lr_save_path, 255*lr.detach().cpu().squeeze().permute(1,2,0).numpy()[..., ::-1])
-                cv2.imwrite(gt_save_path, 255*gt.detach().cpu().squeeze().permute(1,2,0).numpy()[..., ::-1])
-                cv2.imwrite(sr_bicubic_save_path, 255*sr_bicubic.detach().cpu().squeeze().permute(1,2,0).numpy()[..., ::-1])
-               
-
-            del lr_batch, gt_batch, sr_bicubic_batch
-            torch.cuda.empty_cache()
-    

utils_data/make_tags.py

@@ -1,90 +0,0 @@
-'''
- * SeeSR: Towards Semantics-Aware Real-World Image Super-Resolution 
- * Modified from diffusers by Rongyuan Wu
- * 24/12/2023
-'''
-import torch
-import torch.nn as nn
-import torch.nn.functional as F
-import torch.utils.data as data
-from torch.utils.data import DataLoader
-from torchvision import transforms
-from typing import Mapping, Any
-
-import random
-import os
-import cv2
-import glob
-import json
-import math
-from tqdm import tqdm
-
-
-import numpy as np
-from PIL import Image
-
-import sys
-sys.path.append(os.getcwd())
-
-from ram.models.ram import ram
-from ram import inference_ram as inference
-from ram import get_transform
-from ram.utils import build_openset_label_embedding
-
-from basicsr.data.ram_tag_dataset import RAMTagDataset
-
-ram_transforms = transforms.Compose([
-            transforms.ToTensor(),
-            transforms.Resize((384, 384)),
-            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
-        ])
-
-import argparse
-parser = argparse.ArgumentParser()
-parser.add_argument("--root_path", type=str, default='preset/datasets/train_datasets/training_for_seesr', help='the dataset you want to tag.') # 
-parser.add_argument("--start_gpu", type=int, default=0, help='if you have 5 GPUs, you can set it to 0/1/2/3/4 when using different GPU for parallel processing. It will save your time.') 
-parser.add_argument("--all_gpu", type=int, default=1, help='if you set --start_gpu max to 5, please set it to 5') 
-args = parser.parse_args()
-
-gt_path = os.path.join(args.root_path, 'gt')
-tag_path = os.path.join(args.root_path, 'tag')
-os.makedirs(tag_path, exist_ok=True)
-
-lq_lists = glob.glob(os.path.join(gt_path, '*.png'))
-print(f'There are {len(lq_lists)} imgs' )
-
-model = ram(pretrained='/media/ssd8T/ly/SeeSR/preset/models/ram_swin_large_14m.pth',   
-                            image_size=384,
-                            vit='swin_l')
-model = model.eval()
-model = model.to('cuda')
-
-start_num = args.start_gpu * len(lq_lists)//args.all_gpu
-end_num = (args.start_gpu+1) * len(lq_lists)//args.all_gpu
-
-print(f'===== process [{start_num}   {end_num}] =====')
-with torch.no_grad():
-    for lq_idx, lq_path in enumerate(lq_lists[start_num:end_num]):
-        print(f' ====== process {lq_idx,lq_path} imgs... =====')
-        basename = os.path.basename(lq_path).split('.')[0]
-        lq = ram_transforms(Image.open(lq_path)).unsqueeze(0).to('cuda')
-        gt_captions = inference(lq, model)
-        gt_prompt = f"{gt_captions[0]},"
-        tag_save_path = tag_path + f'/{basename}.txt'
-        f = open(f"{tag_save_path}", "w")
-        f.write(gt_prompt)
-        f.close()
-        print(f'The GT tag of {basename}.txt: {gt_prompt}')
-
-
-
-
-
-
-
-
-
-
-
-
-        

utils_data/metrics.py

@@ -1,64 +0,0 @@
-import os
-import pyiqa
-import argparse
-from tqdm import tqdm
-
-def test_image_quality(image_dir, metrics, weight_paths):
-    """
-    测试指定文件夹中所有 PNG 图像的质量指标。
-
-    Args:
-        image_dir (str): 包含 PNG 图像的文件夹路径。
-        metrics (list): 需要测试的指标列表，例如 ['musiq', 'maniqa', 'clipiqa'].
-        weight_paths (dict): 每个指标的本地权重文件路径。
-    """
-    # 初始化指标模型
-    metric_models = {}
-    for metric in metrics:
-        if metric in weight_paths:
-            # 如果提供了本地权重路径，则加载本地权重
-            model = pyiqa.create_metric(metric, pretrained_model_path=weight_paths[metric])
-        else:
-            # 否则使用默认权重（需要网络下载）
-            model = pyiqa.create_metric(metric)
-        metric_models[metric] = model
-
-    # 获取所有 PNG 图像路径
-    image_paths = [os.path.join(image_dir, f) for f in os.listdir(image_dir) if f.endswith('.png')]
-    if not image_paths:
-        print(f"未找到 PNG 图像：{image_dir}")
-        return
-
-
-    # 遍历图像并计算指标
-    results = {metric: [] for metric in metrics}
-    for image_path in tqdm(image_paths, desc="Processing images"):
-        for metric, model in metric_models.items():
-            score = model(image_path)  # 计算指标分数
-            results[metric].append(score.item())  # 将分数添加到结果中
-
-    # 打印结果
-    print("\n测试结果：")
-    for metric, scores in results.items():
-        avg_score = sum(scores) / len(scores)
-        # print(f"{metric.upper()} - 平均分数: {avg_score:.4f}")
-        print(avg_score)
-        # print(f"{metric.upper()} - 单张图像分数: {scores}")
-
-if __name__ == "__main__":
-    # 解析命令行参数
-    parser = argparse.ArgumentParser(description="测试图像质量指标")
-    parser.add_argument("--image_dir", type=str, required=True, help="包含 PNG 图像的文件夹路径")
-    args = parser.parse_args()
-
-    # 需要测试的指标
-    metrics_to_test = ['musiq', 'maniqa', 'clipiqa']
-
-    # 每个指标的本地权重文件路径
-    weight_paths = {
-        'musiq': '/media/ssd8T/wyw/Pretrained/musiq/musiq_koniq_ckpt-e95806b9.pth',
-        'maniqa': '/media/ssd8T/wyw/Pretrained/clipiqa/ckpt_koniq10k.pt',
-    }
-
-    # 运行测试
-    test_image_quality(args.image_dir, metrics_to_test, weight_paths)