Init commit

app.py

@@ -0,0 +1,45 @@
+import os
+import cv2
+import gradio as gr
+from PIL import Image
+import torch
+
+model_path = 'experiments/pretrained_models/Swin2SR_RealworldSR_X4_64_BSRGAN_PSNR.pth'
+
+if os.path.exists(model_path):
+    print(f'loading model from {model_path}')        
+else:
+    os.makedirs(os.path.dirname(model_path), exist_ok=True)
+    url = 'https://github.com/mv-lab/swin2sr/releases/download/v0.0.1/{}'.format(os.path.basename(model_path))
+    r = requests.get(url, allow_redirects=True)
+    print(f'downloading model {model_path}')
+    open(model_path, 'wb').write(r.content)
+
+os.makedirs("test", exist_ok=True)
+    
+def inference(img):
+    
+    cv2.imwrite("test/1.png", cv2.cvtColor(img, cv2.COLOR_BGR2RGB))
+    # basewidth = 256
+    # wpercent = (basewidth/float(img.size[0]))
+    # hsize = int((float(img.size[1])*float(wpercent)))
+    # img = img.resize((basewidth,hsize), Image.ANTIALIAS)
+    #img.save("test/1.jpg", "JPEG")
+    os.system('python main_test_swin2sr.py --task real_sr --model_path experiments/pretrained_models/Swin2SR_RealworldSR_X4_64_BSRGAN_PSNR.pth --folder_lq test --scale 4')
+    return 'results/swin2sr_real_sr_x4/1_Swin2SR.png'
+        
+title = "Swin2SR"
+description = "Gradio demo for Swin2SR."
+article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2209.11345' target='_blank'>Swin2SR: SwinV2 Transformer for Compressed Image Super-Resolution and Restoration></a> | <a href='https://github.com/mv-lab/swin2sr' target='_blank'>Github Repo</a></p>"
+
+examples=[['butterflyx4.png']]
+gr.Interface(
+    inference, 
+    "image", 
+    "image",
+    title=title,
+    description=description,
+    article=article,
+    examples=examples,
+    ).launch(enable_queue=True,
+             share=True)

main_test_swin2sr.py

@@ -0,0 +1,302 @@
+import argparse
+import cv2
+import glob
+import numpy as np
+from collections import OrderedDict
+import os
+import torch
+import requests
+
+from models.network_swin2sr import Swin2SR as net
+from utils import util_calculate_psnr_ssim as util
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--task', type=str, default='color_dn', help='classical_sr, lightweight_sr, real_sr, '
+                                                                     'gray_dn, color_dn, jpeg_car, color_jpeg_car')
+    parser.add_argument('--scale', type=int, default=1, help='scale factor: 1, 2, 3, 4, 8') # 1 for dn and jpeg car
+    parser.add_argument('--noise', type=int, default=15, help='noise level: 15, 25, 50')
+    parser.add_argument('--jpeg', type=int, default=40, help='scale factor: 10, 20, 30, 40')
+    parser.add_argument('--training_patch_size', type=int, default=128, help='patch size used in training Swin2SR. '
+                                       'Just used to differentiate two different settings in Table 2 of the paper. '
+                                       'Images are NOT tested patch by patch.')
+    parser.add_argument('--large_model', action='store_true', help='use large model, only provided for real image sr')
+    parser.add_argument('--model_path', type=str,
+                        default='model_zoo/swin2sr/Swin2SR_ClassicalSR_X2_64.pth')
+    parser.add_argument('--folder_lq', type=str, default=None, help='input low-quality test image folder')
+    parser.add_argument('--folder_gt', type=str, default=None, help='input ground-truth test image folder')
+    parser.add_argument('--tile', type=int, default=None, help='Tile size, None for no tile during testing (testing as a whole)')
+    parser.add_argument('--tile_overlap', type=int, default=32, help='Overlapping of different tiles')
+    parser.add_argument('--save_img_only', default=False, action='store_true', help='save image and do not evaluate')
+    args = parser.parse_args()
+
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    # set up model
+    if os.path.exists(args.model_path):
+        print(f'loading model from {args.model_path}')        
+    else:
+        os.makedirs(os.path.dirname(args.model_path), exist_ok=True)
+        url = 'https://github.com/mv-lab/swin2sr/releases/download/v0.0.1/{}'.format(os.path.basename(args.model_path))
+        r = requests.get(url, allow_redirects=True)
+        print(f'downloading model {args.model_path}')
+        open(args.model_path, 'wb').write(r.content)
+
+    model = define_model(args)
+    model.eval()
+    model = model.to(device)
+
+    # setup folder and path
+    folder, save_dir, border, window_size = setup(args)
+    os.makedirs(save_dir, exist_ok=True)
+    test_results = OrderedDict()
+    test_results['psnr'] = []
+    test_results['ssim'] = []
+    test_results['psnr_y'] = []
+    test_results['ssim_y'] = []
+    test_results['psnrb'] = []
+    test_results['psnrb_y'] = []
+    psnr, ssim, psnr_y, ssim_y, psnrb, psnrb_y = 0, 0, 0, 0, 0, 0
+
+    for idx, path in enumerate(sorted(glob.glob(os.path.join(folder, '*')))):
+        # read image
+        imgname, img_lq, img_gt = get_image_pair(args, path)  # image to HWC-BGR, float32
+        img_lq = np.transpose(img_lq if img_lq.shape[2] == 1 else img_lq[:, :, [2, 1, 0]], (2, 0, 1))  # HCW-BGR to CHW-RGB
+        img_lq = torch.from_numpy(img_lq).float().unsqueeze(0).to(device)  # CHW-RGB to NCHW-RGB
+
+        # inference
+        with torch.no_grad():
+            # pad input image to be a multiple of window_size
+            _, _, h_old, w_old = img_lq.size()
+            h_pad = (h_old // window_size + 1) * window_size - h_old
+            w_pad = (w_old // window_size + 1) * window_size - w_old
+            img_lq = torch.cat([img_lq, torch.flip(img_lq, [2])], 2)[:, :, :h_old + h_pad, :]
+            img_lq = torch.cat([img_lq, torch.flip(img_lq, [3])], 3)[:, :, :, :w_old + w_pad]
+            output = test(img_lq, model, args, window_size)
+            
+            if args.task == 'compressed_sr':
+                output = output[0][..., :h_old * args.scale, :w_old * args.scale]
+            else:
+                output = output[..., :h_old * args.scale, :w_old * args.scale]
+
+        # save image
+        output = output.data.squeeze().float().cpu().clamp_(0, 1).numpy()
+        if output.ndim == 3:
+            output = np.transpose(output[[2, 1, 0], :, :], (1, 2, 0))  # CHW-RGB to HCW-BGR
+        output = (output * 255.0).round().astype(np.uint8)  # float32 to uint8
+        cv2.imwrite(f'{save_dir}/{imgname}_Swin2SR.png', output)
+
+            
+        # evaluate psnr/ssim/psnr_b
+        if img_gt is not None:
+            img_gt = (img_gt * 255.0).round().astype(np.uint8)  # float32 to uint8
+            img_gt = img_gt[:h_old * args.scale, :w_old * args.scale, ...]  # crop gt
+            img_gt = np.squeeze(img_gt)
+
+            psnr = util.calculate_psnr(output, img_gt, crop_border=border)
+            ssim = util.calculate_ssim(output, img_gt, crop_border=border)
+            test_results['psnr'].append(psnr)
+            test_results['ssim'].append(ssim)
+            if img_gt.ndim == 3:  # RGB image
+                psnr_y = util.calculate_psnr(output, img_gt, crop_border=border, test_y_channel=True)
+                ssim_y = util.calculate_ssim(output, img_gt, crop_border=border, test_y_channel=True)
+                test_results['psnr_y'].append(psnr_y)
+                test_results['ssim_y'].append(ssim_y)
+            if args.task in ['jpeg_car', 'color_jpeg_car']:
+                psnrb = util.calculate_psnrb(output, img_gt, crop_border=border, test_y_channel=False)
+                test_results['psnrb'].append(psnrb)
+                if args.task in ['color_jpeg_car']:
+                    psnrb_y = util.calculate_psnrb(output, img_gt, crop_border=border, test_y_channel=True)
+                    test_results['psnrb_y'].append(psnrb_y)
+            print('Testing {:d} {:20s} - PSNR: {:.2f} dB; SSIM: {:.4f}; PSNRB: {:.2f} dB;'
+                  'PSNR_Y: {:.2f} dB; SSIM_Y: {:.4f}; PSNRB_Y: {:.2f} dB.'.
+                  format(idx, imgname, psnr, ssim, psnrb, psnr_y, ssim_y, psnrb_y))
+        else:
+            print('Testing {:d} {:20s}'.format(idx, imgname))
+
+    # summarize psnr/ssim
+    if img_gt is not None:
+        ave_psnr = sum(test_results['psnr']) / len(test_results['psnr'])
+        ave_ssim = sum(test_results['ssim']) / len(test_results['ssim'])
+        print('\n{} \n-- Average PSNR/SSIM(RGB): {:.2f} dB; {:.4f}'.format(save_dir, ave_psnr, ave_ssim))
+        if img_gt.ndim == 3:
+            ave_psnr_y = sum(test_results['psnr_y']) / len(test_results['psnr_y'])
+            ave_ssim_y = sum(test_results['ssim_y']) / len(test_results['ssim_y'])
+            print('-- Average PSNR_Y/SSIM_Y: {:.2f} dB; {:.4f}'.format(ave_psnr_y, ave_ssim_y))
+        if args.task in ['jpeg_car', 'color_jpeg_car']:
+            ave_psnrb = sum(test_results['psnrb']) / len(test_results['psnrb'])
+            print('-- Average PSNRB: {:.2f} dB'.format(ave_psnrb))
+            if args.task in ['color_jpeg_car']:
+                ave_psnrb_y = sum(test_results['psnrb_y']) / len(test_results['psnrb_y'])
+                print('-- Average PSNRB_Y: {:.2f} dB'.format(ave_psnrb_y))
+
+
+def define_model(args):
+    # 001 classical image sr
+    if args.task == 'classical_sr':
+        model = net(upscale=args.scale, in_chans=3, img_size=args.training_patch_size, 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')
+        param_key_g = 'params'
+
+    # 002 lightweight image sr
+    # use 'pixelshuffledirect' to save parameters
+    elif args.task in ['lightweight_sr']:
+        model = net(upscale=args.scale, in_chans=3, img_size=64, window_size=8,
+                    img_range=1., depths=[6, 6, 6, 6], embed_dim=60, num_heads=[6, 6, 6, 6],
+                    mlp_ratio=2, upsampler='pixelshuffledirect', resi_connection='1conv')
+        param_key_g = 'params'
+        
+    elif args.task == 'compressed_sr':
+        model = net(upscale=args.scale, in_chans=3, img_size=args.training_patch_size, 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_aux', resi_connection='1conv')
+        param_key_g = 'params'                
+
+    # 003 real-world image sr
+    elif args.task == 'real_sr':
+        if not args.large_model:
+            # use 'nearest+conv' to avoid block artifacts
+            model = net(upscale=args.scale, in_chans=3, img_size=64, 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='nearest+conv', resi_connection='1conv')
+        else:
+            # larger model size; use '3conv' to save parameters and memory; use ema for GAN training
+            model = net(upscale=args.scale, in_chans=3, img_size=64, window_size=8,
+                        img_range=1., depths=[6, 6, 6, 6, 6, 6, 6, 6, 6], embed_dim=240,
+                        num_heads=[8, 8, 8, 8, 8, 8, 8, 8, 8],
+                        mlp_ratio=2, upsampler='nearest+conv', resi_connection='3conv')
+        param_key_g = 'params_ema'
+
+    # 006 grayscale JPEG compression artifact reduction
+    # use window_size=7 because JPEG encoding uses 8x8; use img_range=255 because it's sligtly better than 1
+    elif args.task == 'jpeg_car':
+        model = net(upscale=1, in_chans=1, img_size=126, window_size=7,
+                    img_range=255., depths=[6, 6, 6, 6, 6, 6], embed_dim=180, num_heads=[6, 6, 6, 6, 6, 6],
+                    mlp_ratio=2, upsampler='', resi_connection='1conv')
+        param_key_g = 'params'
+
+    # 006 color JPEG compression artifact reduction
+    # use window_size=7 because JPEG encoding uses 8x8; use img_range=255 because it's sligtly better than 1
+    elif args.task == 'color_jpeg_car':
+        model = net(upscale=1, in_chans=3, img_size=126, window_size=7,
+                    img_range=255., depths=[6, 6, 6, 6, 6, 6], embed_dim=180, num_heads=[6, 6, 6, 6, 6, 6],
+                    mlp_ratio=2, upsampler='', resi_connection='1conv')
+        param_key_g = 'params'
+
+    pretrained_model = torch.load(args.model_path)
+    model.load_state_dict(pretrained_model[param_key_g] if param_key_g in pretrained_model.keys() else pretrained_model, strict=True)
+
+    return model
+
+
+def setup(args):
+    # 001 classical image sr/ 002 lightweight image sr
+    if args.task in ['classical_sr', 'lightweight_sr', 'compressed_sr']:
+        save_dir = f'results/swin2sr_{args.task}_x{args.scale}'
+        if args.save_img_only:
+            folder = args.folder_lq
+        else:
+            folder = args.folder_gt
+        border = args.scale
+        window_size = 8
+
+    # 003 real-world image sr
+    elif args.task in ['real_sr']:
+        save_dir = f'results/swin2sr_{args.task}_x{args.scale}'
+        if args.large_model:
+            save_dir += '_large'
+        folder = args.folder_lq
+        border = 0
+        window_size = 8
+
+    # 006 JPEG compression artifact reduction
+    elif args.task in ['jpeg_car', 'color_jpeg_car']:
+        save_dir = f'results/swin2sr_{args.task}_jpeg{args.jpeg}'
+        folder = args.folder_gt
+        border = 0
+        window_size = 7
+
+    return folder, save_dir, border, window_size
+
+
+def get_image_pair(args, path):
+    (imgname, imgext) = os.path.splitext(os.path.basename(path))
+
+    # 001 classical image sr/ 002 lightweight image sr (load lq-gt image pairs)
+    if args.task in ['classical_sr', 'lightweight_sr']:
+        if args.save_img_only:
+            img_gt = None
+            img_lq = cv2.imread(path, cv2.IMREAD_COLOR).astype(np.float32) / 255.            
+        else:
+            img_gt = cv2.imread(path, cv2.IMREAD_COLOR).astype(np.float32) / 255.
+            img_lq = cv2.imread(f'{args.folder_lq}/{imgname}x{args.scale}{imgext}', cv2.IMREAD_COLOR).astype(
+                np.float32) / 255.            
+        
+    elif args.task in ['compressed_sr']:
+        if args.save_img_only:
+            img_gt = None
+            img_lq = cv2.imread(path, cv2.IMREAD_COLOR).astype(np.float32) / 255.            
+        else:
+            img_gt = cv2.imread(path, cv2.IMREAD_COLOR).astype(np.float32) / 255.
+            img_lq = cv2.imread(f'{args.folder_lq}/{imgname}.jpg', cv2.IMREAD_COLOR).astype(
+                np.float32) / 255.        
+
+    # 003 real-world image sr (load lq image only)
+    elif args.task in ['real_sr', 'lightweight_sr_infer']:
+        img_gt = None
+        img_lq = cv2.imread(path, cv2.IMREAD_COLOR).astype(np.float32) / 255.
+
+    # 006 grayscale JPEG compression artifact reduction (load gt image and generate lq image on-the-fly)
+    elif args.task in ['jpeg_car']:
+        img_gt = cv2.imread(path, cv2.IMREAD_UNCHANGED)
+        if img_gt.ndim != 2:
+            img_gt = util.bgr2ycbcr(img_gt, y_only=True)
+        result, encimg = cv2.imencode('.jpg', img_gt, [int(cv2.IMWRITE_JPEG_QUALITY), args.jpeg])
+        img_lq = cv2.imdecode(encimg, 0)
+        img_gt = np.expand_dims(img_gt, axis=2).astype(np.float32) / 255.
+        img_lq = np.expand_dims(img_lq, axis=2).astype(np.float32) / 255.
+
+    # 006 JPEG compression artifact reduction (load gt image and generate lq image on-the-fly)
+    elif args.task in ['color_jpeg_car']:
+        img_gt = cv2.imread(path)
+        result, encimg = cv2.imencode('.jpg', img_gt, [int(cv2.IMWRITE_JPEG_QUALITY), args.jpeg])
+        img_lq = cv2.imdecode(encimg, 1)
+        img_gt = img_gt.astype(np.float32)/ 255.
+        img_lq = img_lq.astype(np.float32)/ 255.
+
+    return imgname, img_lq, img_gt
+
+
+def test(img_lq, model, args, window_size):
+    if args.tile is None:
+        # test the image as a whole
+        output = model(img_lq)
+    else:
+        # test the image tile by tile
+        b, c, h, w = img_lq.size()
+        tile = min(args.tile, h, w)
+        assert tile % window_size == 0, "tile size should be a multiple of window_size"
+        tile_overlap = args.tile_overlap
+        sf = args.scale
+
+        stride = tile - tile_overlap
+        h_idx_list = list(range(0, h-tile, stride)) + [h-tile]
+        w_idx_list = list(range(0, w-tile, stride)) + [w-tile]
+        E = torch.zeros(b, c, h*sf, w*sf).type_as(img_lq)
+        W = torch.zeros_like(E)
+
+        for h_idx in h_idx_list:
+            for w_idx in w_idx_list:
+                in_patch = img_lq[..., h_idx:h_idx+tile, w_idx:w_idx+tile]
+                out_patch = model(in_patch)
+                out_patch_mask = torch.ones_like(out_patch)
+
+                E[..., h_idx*sf:(h_idx+tile)*sf, w_idx*sf:(w_idx+tile)*sf].add_(out_patch)
+                W[..., h_idx*sf:(h_idx+tile)*sf, w_idx*sf:(w_idx+tile)*sf].add_(out_patch_mask)
+        output = E.div_(W)
+
+    return output
+
+if __name__ == '__main__':
+    main()

models/network_swin2sr.py

@@ -0,0 +1,1010 @@
+# -----------------------------------------------------------------------------------
+# Swin2SR: Swin2SR: SwinV2 Transformer for Compressed Image Super-Resolution and Restoration, https://arxiv.org/abs/
+# Written by Conde and Choi et al.
+# -----------------------------------------------------------------------------------
+
+import math
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+from timm.models.layers import DropPath, to_2tuple, trunc_normal_
+
+
+class Mlp(nn.Module):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    return windows
+
+
+def window_reverse(windows, window_size, H, W):
+    """
+    Args:
+        windows: (num_windows*B, window_size, window_size, C)
+        window_size (int): Window size
+        H (int): Height of image
+        W (int): Width of image
+    Returns:
+        x: (B, H, W, C)
+    """
+    B = int(windows.shape[0] / (H * W / window_size / window_size))
+    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+class WindowAttention(nn.Module):
+    r""" Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+        pretrained_window_size (tuple[int]): The height and width of the window in pre-training.
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, attn_drop=0., proj_drop=0.,
+                 pretrained_window_size=[0, 0]):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.pretrained_window_size = pretrained_window_size
+        self.num_heads = num_heads
+
+        self.logit_scale = nn.Parameter(torch.log(10 * torch.ones((num_heads, 1, 1))), requires_grad=True)
+
+        # mlp to generate continuous relative position bias
+        self.cpb_mlp = nn.Sequential(nn.Linear(2, 512, bias=True),
+                                     nn.ReLU(inplace=True),
+                                     nn.Linear(512, num_heads, bias=False))
+
+        # get relative_coords_table
+        relative_coords_h = torch.arange(-(self.window_size[0] - 1), self.window_size[0], dtype=torch.float32)
+        relative_coords_w = torch.arange(-(self.window_size[1] - 1), self.window_size[1], dtype=torch.float32)
+        relative_coords_table = torch.stack(
+            torch.meshgrid([relative_coords_h,
+                            relative_coords_w])).permute(1, 2, 0).contiguous().unsqueeze(0)  # 1, 2*Wh-1, 2*Ww-1, 2
+        if pretrained_window_size[0] > 0:
+            relative_coords_table[:, :, :, 0] /= (pretrained_window_size[0] - 1)
+            relative_coords_table[:, :, :, 1] /= (pretrained_window_size[1] - 1)
+        else:
+            relative_coords_table[:, :, :, 0] /= (self.window_size[0] - 1)
+            relative_coords_table[:, :, :, 1] /= (self.window_size[1] - 1)
+        relative_coords_table *= 8  # normalize to -8, 8
+        relative_coords_table = torch.sign(relative_coords_table) * torch.log2(
+            torch.abs(relative_coords_table) + 1.0) / np.log2(8)
+
+        self.register_buffer("relative_coords_table", relative_coords_table)
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=False)
+        if qkv_bias:
+            self.q_bias = nn.Parameter(torch.zeros(dim))
+            self.v_bias = nn.Parameter(torch.zeros(dim))
+        else:
+            self.q_bias = None
+            self.v_bias = None
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv_bias = None
+        if self.q_bias is not None:
+            qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias))
+        qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
+        qkv = qkv.reshape(B_, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        # cosine attention
+        attn = (F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(-2, -1))
+        logit_scale = torch.clamp(self.logit_scale, max=torch.log(torch.tensor(1. / 0.01)).to(self.logit_scale.device)).exp()
+        attn = attn * logit_scale
+
+        relative_position_bias_table = self.cpb_mlp(self.relative_coords_table).view(-1, self.num_heads)
+        relative_position_bias = relative_position_bias_table[self.relative_position_index.view(-1)].view(
+            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        relative_position_bias = 16 * torch.sigmoid(relative_position_bias)
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+    def extra_repr(self) -> str:
+        return f'dim={self.dim}, window_size={self.window_size}, ' \
+               f'pretrained_window_size={self.pretrained_window_size}, num_heads={self.num_heads}'
+
+    def flops(self, N):
+        # calculate flops for 1 window with token length of N
+        flops = 0
+        # qkv = self.qkv(x)
+        flops += N * self.dim * 3 * self.dim
+        # attn = (q @ k.transpose(-2, -1))
+        flops += self.num_heads * N * (self.dim // self.num_heads) * N
+        #  x = (attn @ v)
+        flops += self.num_heads * N * N * (self.dim // self.num_heads)
+        # x = self.proj(x)
+        flops += N * self.dim * self.dim
+        return flops
+
+class SwinTransformerBlock(nn.Module):
+    r""" Swin Transformer Block.
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resulotion.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+        pretrained_window_size (int): Window size in pre-training.
+    """
+
+    def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
+                 mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0., drop_path=0.,
+                 act_layer=nn.GELU, norm_layer=nn.LayerNorm, pretrained_window_size=0):
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        if min(self.input_resolution) <= self.window_size:
+            # if window size is larger than input resolution, we don't partition windows
+            self.shift_size = 0
+            self.window_size = min(self.input_resolution)
+        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop,
+            pretrained_window_size=to_2tuple(pretrained_window_size))
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        if self.shift_size > 0:
+            attn_mask = self.calculate_mask(self.input_resolution)
+        else:
+            attn_mask = None
+
+        self.register_buffer("attn_mask", attn_mask)
+        
+    def calculate_mask(self, x_size):
+        # calculate attention mask for SW-MSA
+        H, W = x_size
+        img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
+        h_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size),
+                    slice(-self.window_size, -self.shift_size),
+                    slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        return attn_mask        
+
+    def forward(self, x, x_size):
+        H, W = x_size
+        B, L, C = x.shape
+        #assert L == H * W, "input feature has wrong size"
+
+        shortcut = x
+        x = x.view(B, H, W, C)
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+        else:
+            shifted_x = x
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size
+        if self.input_resolution == x_size:
+            attn_windows = self.attn(x_windows, mask=self.attn_mask)  # nW*B, window_size*window_size, C
+        else:
+            attn_windows = self.attn(x_windows, mask=self.calculate_mask(x_size).to(x.device))
+            
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
+        shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+        x = x.view(B, H * W, C)
+        x = shortcut + self.drop_path(self.norm1(x))
+
+        # FFN
+        x = x + self.drop_path(self.norm2(self.mlp(x)))
+
+        return x
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
+               f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"
+
+    def flops(self):
+        flops = 0
+        H, W = self.input_resolution
+        # norm1
+        flops += self.dim * H * W
+        # W-MSA/SW-MSA
+        nW = H * W / self.window_size / self.window_size
+        flops += nW * self.attn.flops(self.window_size * self.window_size)
+        # mlp
+        flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
+        # norm2
+        flops += self.dim * H * W
+        return flops
+
+class PatchMerging(nn.Module):
+    r""" Patch Merging Layer.
+    Args:
+        input_resolution (tuple[int]): Resolution of input feature.
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(2 * dim)
+
+    def forward(self, x):
+        """
+        x: B, H*W, C
+        """
+        H, W = self.input_resolution
+        B, L, C = x.shape
+        assert L == H * W, "input feature has wrong size"
+        assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."
+
+        x = x.view(B, H, W, C)
+
+        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
+        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
+        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
+        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.reduction(x)
+        x = self.norm(x)
+
+        return x
+
+    def extra_repr(self) -> str:
+        return f"input_resolution={self.input_resolution}, dim={self.dim}"
+
+    def flops(self):
+        H, W = self.input_resolution
+        flops = (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
+        flops += H * W * self.dim // 2
+        return flops    
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+        pretrained_window_size (int): Local window size in pre-training.
+    """
+
+    def __init__(self, dim, input_resolution, depth, num_heads, window_size,
+                 mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0.,
+                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False,
+                 pretrained_window_size=0):
+
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(dim=dim, input_resolution=input_resolution,
+                                 num_heads=num_heads, window_size=window_size,
+                                 shift_size=0 if (i % 2 == 0) else window_size // 2,
+                                 mlp_ratio=mlp_ratio,
+                                 qkv_bias=qkv_bias,
+                                 drop=drop, attn_drop=attn_drop,
+                                 drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                                 norm_layer=norm_layer,
+                                 pretrained_window_size=pretrained_window_size)
+            for i in range(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x, x_size):
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x, x_size)
+            else:
+                x = blk(x, x_size)
+        if self.downsample is not None:
+            x = self.downsample(x)
+        return x
+
+    def extra_repr(self) -> str:
+        return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"
+
+    def flops(self):
+        flops = 0
+        for blk in self.blocks:
+            flops += blk.flops()
+        if self.downsample is not None:
+            flops += self.downsample.flops()
+        return flops
+
+    def _init_respostnorm(self):
+        for blk in self.blocks:
+            nn.init.constant_(blk.norm1.bias, 0)
+            nn.init.constant_(blk.norm1.weight, 0)
+            nn.init.constant_(blk.norm2.bias, 0)
+            nn.init.constant_(blk.norm2.weight, 0)
+            
+class PatchEmbed(nn.Module):
+    r""" Image to Patch Embedding
+    Args:
+        img_size (int): Image size.  Default: 224.
+        patch_size (int): Patch token size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patches_resolution = patches_resolution
+        self.num_patches = patches_resolution[0] * patches_resolution[1]
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        B, C, H, W = x.shape
+        # FIXME look at relaxing size constraints
+        # assert H == self.img_size[0] and W == self.img_size[1],
+        #     f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
+        x = self.proj(x).flatten(2).transpose(1, 2)  # B Ph*Pw C
+        if self.norm is not None:
+            x = self.norm(x)
+        return x
+
+    def flops(self):
+        Ho, Wo = self.patches_resolution
+        flops = Ho * Wo * self.embed_dim * self.in_chans * (self.patch_size[0] * self.patch_size[1])
+        if self.norm is not None:
+            flops += Ho * Wo * self.embed_dim
+        return flops           
+
+class RSTB(nn.Module):
+    """Residual Swin Transformer Block (RSTB).
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+        img_size: Input image size.
+        patch_size: Patch size.
+        resi_connection: The convolutional block before residual connection.
+    """
+
+    def __init__(self, dim, input_resolution, depth, num_heads, window_size,
+                 mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0.,
+                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False,
+                 img_size=224, patch_size=4, resi_connection='1conv'):
+        super(RSTB, self).__init__()
+
+        self.dim = dim
+        self.input_resolution = input_resolution
+
+        self.residual_group = BasicLayer(dim=dim,
+                                         input_resolution=input_resolution,
+                                         depth=depth,
+                                         num_heads=num_heads,
+                                         window_size=window_size,
+                                         mlp_ratio=mlp_ratio,
+                                         qkv_bias=qkv_bias, 
+                                         drop=drop, attn_drop=attn_drop,
+                                         drop_path=drop_path,
+                                         norm_layer=norm_layer,
+                                         downsample=downsample,
+                                         use_checkpoint=use_checkpoint)
+
+        if resi_connection == '1conv':
+            self.conv = nn.Conv2d(dim, dim, 3, 1, 1)
+        elif resi_connection == '3conv':
+            # to save parameters and memory
+            self.conv = nn.Sequential(nn.Conv2d(dim, dim // 4, 3, 1, 1), nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                                      nn.Conv2d(dim // 4, dim // 4, 1, 1, 0),
+                                      nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                                      nn.Conv2d(dim // 4, dim, 3, 1, 1))
+
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=dim, embed_dim=dim,
+            norm_layer=None)
+
+        self.patch_unembed = PatchUnEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=dim, embed_dim=dim,
+            norm_layer=None)
+
+    def forward(self, x, x_size):
+        return self.patch_embed(self.conv(self.patch_unembed(self.residual_group(x, x_size), x_size))) + x
+
+    def flops(self):
+        flops = 0
+        flops += self.residual_group.flops()
+        H, W = self.input_resolution
+        flops += H * W * self.dim * self.dim * 9
+        flops += self.patch_embed.flops()
+        flops += self.patch_unembed.flops()
+
+        return flops
+
+class PatchUnEmbed(nn.Module):
+    r""" Image to Patch Unembedding
+    Args:
+        img_size (int): Image size.  Default: 224.
+        patch_size (int): Patch token size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patches_resolution = patches_resolution
+        self.num_patches = patches_resolution[0] * patches_resolution[1]
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+    def forward(self, x, x_size):
+        B, HW, C = x.shape
+        x = x.transpose(1, 2).view(B, self.embed_dim, x_size[0], x_size[1])  # B Ph*Pw C
+        return x
+
+    def flops(self):
+        flops = 0
+        return flops
+
+
+class Upsample(nn.Sequential):
+    """Upsample module.
+    Args:
+        scale (int): Scale factor. Supported scales: 2^n and 3.
+        num_feat (int): Channel number of intermediate features.
+    """
+
+    def __init__(self, scale, num_feat):
+        m = []
+        if (scale & (scale - 1)) == 0:  # scale = 2^n
+            for _ in range(int(math.log(scale, 2))):
+                m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
+                m.append(nn.PixelShuffle(2))
+        elif scale == 3:
+            m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
+            m.append(nn.PixelShuffle(3))
+        else:
+            raise ValueError(f'scale {scale} is not supported. ' 'Supported scales: 2^n and 3.')
+        super(Upsample, self).__init__(*m)
+        
+class Upsample_hf(nn.Sequential):
+    """Upsample module.
+    Args:
+        scale (int): Scale factor. Supported scales: 2^n and 3.
+        num_feat (int): Channel number of intermediate features.
+    """
+
+    def __init__(self, scale, num_feat):
+        m = []
+        if (scale & (scale - 1)) == 0:  # scale = 2^n
+            for _ in range(int(math.log(scale, 2))):
+                m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
+                m.append(nn.PixelShuffle(2))
+        elif scale == 3:
+            m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
+            m.append(nn.PixelShuffle(3))
+        else:
+            raise ValueError(f'scale {scale} is not supported. ' 'Supported scales: 2^n and 3.')
+        super(Upsample_hf, self).__init__(*m)        
+
+
+class UpsampleOneStep(nn.Sequential):
+    """UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle)
+       Used in lightweight SR to save parameters.
+    Args:
+        scale (int): Scale factor. Supported scales: 2^n and 3.
+        num_feat (int): Channel number of intermediate features.
+    """
+
+    def __init__(self, scale, num_feat, num_out_ch, input_resolution=None):
+        self.num_feat = num_feat
+        self.input_resolution = input_resolution
+        m = []
+        m.append(nn.Conv2d(num_feat, (scale ** 2) * num_out_ch, 3, 1, 1))
+        m.append(nn.PixelShuffle(scale))
+        super(UpsampleOneStep, self).__init__(*m)
+
+    def flops(self):
+        H, W = self.input_resolution
+        flops = H * W * self.num_feat * 3 * 9
+        return flops
+    
+    
+
+class Swin2SR(nn.Module):
+    r""" Swin2SR
+        A PyTorch impl of : `Swin2SR: SwinV2 Transformer for Compressed Image Super-Resolution and Restoration`.
+    Args:
+        img_size (int | tuple(int)): Input image size. Default 64
+        patch_size (int | tuple(int)): Patch size. Default: 1
+        in_chans (int): Number of input image channels. Default: 3
+        embed_dim (int): Patch embedding dimension. Default: 96
+        depths (tuple(int)): Depth of each Swin Transformer layer.
+        num_heads (tuple(int)): Number of attention heads in different layers.
+        window_size (int): Window size. Default: 7
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        drop_rate (float): Dropout rate. Default: 0
+        attn_drop_rate (float): Attention dropout rate. Default: 0
+        drop_path_rate (float): Stochastic depth rate. Default: 0.1
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
+        upscale: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction
+        img_range: Image range. 1. or 255.
+        upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None
+        resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
+    """
+
+    def __init__(self, img_size=64, patch_size=1, in_chans=3,
+                 embed_dim=96, depths=[6, 6, 6, 6], num_heads=[6, 6, 6, 6],
+                 window_size=7, mlp_ratio=4., qkv_bias=True, 
+                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
+                 norm_layer=nn.LayerNorm, ape=False, patch_norm=True,
+                 use_checkpoint=False, upscale=2, img_range=1., upsampler='', resi_connection='1conv',
+                 **kwargs):
+        super(Swin2SR, self).__init__()
+        num_in_ch = in_chans
+        num_out_ch = in_chans
+        num_feat = 64
+        self.img_range = img_range
+        if in_chans == 3:
+            rgb_mean = (0.4488, 0.4371, 0.4040)
+            self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1)
+        else:
+            self.mean = torch.zeros(1, 1, 1, 1)
+        self.upscale = upscale
+        self.upsampler = upsampler
+        self.window_size = window_size
+
+        #####################################################################################################
+        ################################### 1, shallow feature extraction ###################################
+        self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1)
+
+        #####################################################################################################
+        ################################### 2, deep feature extraction ######################################
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.num_features = embed_dim
+        self.mlp_ratio = mlp_ratio
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+        num_patches = self.patch_embed.num_patches
+        patches_resolution = self.patch_embed.patches_resolution
+        self.patches_resolution = patches_resolution
+
+        # merge non-overlapping patches into image
+        self.patch_unembed = PatchUnEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=embed_dim, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+
+        # absolute position embedding
+        if self.ape:
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        # stochastic depth
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+
+        # build Residual Swin Transformer blocks (RSTB)
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            layer = RSTB(dim=embed_dim,
+                         input_resolution=(patches_resolution[0],
+                                           patches_resolution[1]),
+                         depth=depths[i_layer],
+                         num_heads=num_heads[i_layer],
+                         window_size=window_size,
+                         mlp_ratio=self.mlp_ratio,
+                         qkv_bias=qkv_bias, 
+                         drop=drop_rate, attn_drop=attn_drop_rate,
+                         drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],  # no impact on SR results
+                         norm_layer=norm_layer,
+                         downsample=None,
+                         use_checkpoint=use_checkpoint,
+                         img_size=img_size,
+                         patch_size=patch_size,
+                         resi_connection=resi_connection
+
+                         )
+            self.layers.append(layer)
+            
+        if self.upsampler == 'pixelshuffle_hf':
+            self.layers_hf = nn.ModuleList()
+            for i_layer in range(self.num_layers):
+                layer = RSTB(dim=embed_dim,
+                             input_resolution=(patches_resolution[0],
+                                               patches_resolution[1]),
+                             depth=depths[i_layer],
+                             num_heads=num_heads[i_layer],
+                             window_size=window_size,
+                             mlp_ratio=self.mlp_ratio,
+                             qkv_bias=qkv_bias, 
+                             drop=drop_rate, attn_drop=attn_drop_rate,
+                             drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],  # no impact on SR results
+                             norm_layer=norm_layer,
+                             downsample=None,
+                             use_checkpoint=use_checkpoint,
+                             img_size=img_size,
+                             patch_size=patch_size,
+                             resi_connection=resi_connection
+
+                             )
+                self.layers_hf.append(layer)
+                        
+        self.norm = norm_layer(self.num_features)
+
+        # build the last conv layer in deep feature extraction
+        if resi_connection == '1conv':
+            self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1)
+        elif resi_connection == '3conv':
+            # to save parameters and memory
+            self.conv_after_body = nn.Sequential(nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1),
+                                                 nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                                                 nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0),
+                                                 nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                                                 nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1))
+
+        #####################################################################################################
+        ################################ 3, high quality image reconstruction ################################
+        if self.upsampler == 'pixelshuffle':
+            # for classical SR
+            self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
+                                                      nn.LeakyReLU(inplace=True))
+            self.upsample = Upsample(upscale, num_feat)
+            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+        elif self.upsampler == 'pixelshuffle_aux':
+            self.conv_bicubic = nn.Conv2d(num_in_ch, num_feat, 3, 1, 1)
+            self.conv_before_upsample = nn.Sequential(
+                nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
+                nn.LeakyReLU(inplace=True))
+            self.conv_aux = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+            self.conv_after_aux = nn.Sequential(
+                nn.Conv2d(3, num_feat, 3, 1, 1),
+                nn.LeakyReLU(inplace=True))            
+            self.upsample = Upsample(upscale, num_feat)
+            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+            
+        elif self.upsampler == 'pixelshuffle_hf':
+            self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
+                                                      nn.LeakyReLU(inplace=True))
+            self.upsample = Upsample(upscale, num_feat)
+            self.upsample_hf = Upsample_hf(upscale, num_feat)
+            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+            self.conv_first_hf = nn.Sequential(nn.Conv2d(num_feat, embed_dim, 3, 1, 1),
+                                                      nn.LeakyReLU(inplace=True))
+            self.conv_after_body_hf = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1)
+            self.conv_before_upsample_hf = nn.Sequential(
+                nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
+                nn.LeakyReLU(inplace=True))
+            self.conv_last_hf = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+            
+        elif self.upsampler == 'pixelshuffledirect':
+            # for lightweight SR (to save parameters)
+            self.upsample = UpsampleOneStep(upscale, embed_dim, num_out_ch,
+                                            (patches_resolution[0], patches_resolution[1]))
+        elif self.upsampler == 'nearest+conv':
+            # for real-world SR (less artifacts)
+            assert self.upscale == 4, 'only support x4 now.'
+            self.conv_before_upsample = nn.Sequential(nn.Conv2d(embed_dim, num_feat, 3, 1, 1),
+                                                      nn.LeakyReLU(inplace=True))
+            self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+            self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+            self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+            self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
+        else:
+            # for image denoising and JPEG compression artifact reduction
+            self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'absolute_pos_embed'}
+
+    @torch.jit.ignore
+    def no_weight_decay_keywords(self):
+        return {'relative_position_bias_table'}
+
+    def check_image_size(self, x):
+        _, _, h, w = x.size()
+        mod_pad_h = (self.window_size - h % self.window_size) % self.window_size
+        mod_pad_w = (self.window_size - w % self.window_size) % self.window_size
+        x = F.pad(x, (0, mod_pad_w, 0, mod_pad_h), 'reflect')
+        return x
+
+    def forward_features(self, x):
+        x_size = (x.shape[2], x.shape[3])
+        x = self.patch_embed(x)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+
+        for layer in self.layers:
+            x = layer(x, x_size)
+
+        x = self.norm(x)  # B L C
+        x = self.patch_unembed(x, x_size)
+
+        return x
+    
+    def forward_features_hf(self, x):
+        x_size = (x.shape[2], x.shape[3])
+        x = self.patch_embed(x)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+
+        for layer in self.layers_hf:
+            x = layer(x, x_size)
+
+        x = self.norm(x)  # B L C
+        x = self.patch_unembed(x, x_size)
+
+        return x    
+
+    def forward(self, x):
+        H, W = x.shape[2:]
+        x = self.check_image_size(x)
+
+        self.mean = self.mean.type_as(x)
+        x = (x - self.mean) * self.img_range
+
+        if self.upsampler == 'pixelshuffle':
+            # for classical SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.conv_before_upsample(x)
+            x = self.conv_last(self.upsample(x))
+        elif self.upsampler == 'pixelshuffle_aux':
+            bicubic = F.interpolate(x, size=(H * self.upscale, W * self.upscale), mode='bicubic', align_corners=False)
+            bicubic = self.conv_bicubic(bicubic)
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.conv_before_upsample(x)
+            aux = self.conv_aux(x) # b, 3, LR_H, LR_W
+            x = self.conv_after_aux(aux)
+            x = self.upsample(x)[:, :, :H * self.upscale, :W * self.upscale] + bicubic[:, :, :H * self.upscale, :W * self.upscale]
+            x = self.conv_last(x)
+            aux = aux / self.img_range + self.mean
+        elif self.upsampler == 'pixelshuffle_hf':
+            # for classical SR with HF
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x_before = self.conv_before_upsample(x)
+            x_out = self.conv_last(self.upsample(x_before))
+            
+            x_hf = self.conv_first_hf(x_before)
+            x_hf = self.conv_after_body_hf(self.forward_features_hf(x_hf)) + x_hf
+            x_hf = self.conv_before_upsample_hf(x_hf)
+            x_hf = self.conv_last_hf(self.upsample_hf(x_hf))
+            x = x_out + x_hf
+            x_hf = x_hf / self.img_range + self.mean
+
+        elif self.upsampler == 'pixelshuffledirect':
+            # for lightweight SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.upsample(x)
+        elif self.upsampler == 'nearest+conv':
+            # for real-world SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.conv_before_upsample(x)
+            x = self.lrelu(self.conv_up1(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
+            x = self.lrelu(self.conv_up2(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
+            x = self.conv_last(self.lrelu(self.conv_hr(x)))
+        else:
+            # for image denoising and JPEG compression artifact reduction
+            x_first = self.conv_first(x)
+            res = self.conv_after_body(self.forward_features(x_first)) + x_first
+            x = x + self.conv_last(res)
+        
+        x = x / self.img_range + self.mean
+        if self.upsampler == "pixelshuffle_aux":
+            return x[:, :, :H*self.upscale, :W*self.upscale], aux
+        
+        elif self.upsampler == "pixelshuffle_hf":
+            x_out = x_out / self.img_range + self.mean
+            return x_out[:, :, :H*self.upscale, :W*self.upscale], x[:, :, :H*self.upscale, :W*self.upscale], x_hf[:, :, :H*self.upscale, :W*self.upscale]
+        
+        else:
+            return x[:, :, :H*self.upscale, :W*self.upscale]
+
+    def flops(self):
+        flops = 0
+        H, W = self.patches_resolution
+        flops += H * W * 3 * self.embed_dim * 9
+        flops += self.patch_embed.flops()
+        for i, layer in enumerate(self.layers):
+            flops += layer.flops()
+        flops += H * W * 3 * self.embed_dim * self.embed_dim
+        flops += self.upsample.flops()
+        return flops
+
+
+if __name__ == '__main__':
+    upscale = 4
+    window_size = 8
+    height = (1024 // upscale // window_size + 1) * window_size
+    width = (720 // upscale // window_size + 1) * window_size
+    model = Swin2SR(upscale=2, img_size=(height, width),
+                   window_size=window_size, img_range=1., depths=[6, 6, 6, 6],
+                   embed_dim=60, num_heads=[6, 6, 6, 6], mlp_ratio=2, upsampler='pixelshuffledirect')
+    print(model)
+    print(height, width, model.flops() / 1e9)
+
+    x = torch.randn((1, 3, height, width))
+    x = model(x)
+    print(x.shape)

requirements.txt

@@ -0,0 +1,5 @@
+timm
+numpy
+torch
+opencv-python-headless
+Pillow

utils/util_calculate_psnr_ssim.py

@@ -0,0 +1,320 @@
+# -----------------------------------------------------------------------------------
+# https://github.com/JingyunLiang/SwinIR/blob/main/utils/util_calculate_psnr_ssim.py
+# -----------------------------------------------------------------------------------
+
+import cv2
+import torch
+import numpy as np
+
+def calculate_psnr(img1, img2, crop_border, input_order='HWC', test_y_channel=False):
+    """Calculate PSNR (Peak Signal-to-Noise Ratio).
+    Ref: https://en.wikipedia.org/wiki/Peak_signal-to-noise_ratio
+    Args:
+        img1 (ndarray): Images with range [0, 255].
+        img2 (ndarray): Images with range [0, 255].
+        crop_border (int): Cropped pixels in each edge of an image. These
+            pixels are not involved in the PSNR calculation.
+        input_order (str): Whether the input order is 'HWC' or 'CHW'.
+            Default: 'HWC'.
+        test_y_channel (bool): Test on Y channel of YCbCr. Default: False.
+    Returns:
+        float: psnr result.
+    """
+
+    assert img1.shape == img2.shape, (f'Image shapes are differnet: {img1.shape}, {img2.shape}.')
+    if input_order not in ['HWC', 'CHW']:
+        raise ValueError(f'Wrong input_order {input_order}. Supported input_orders are ' '"HWC" and "CHW"')
+    img1 = reorder_image(img1, input_order=input_order)
+    img2 = reorder_image(img2, input_order=input_order)
+    img1 = img1.astype(np.float64)
+    img2 = img2.astype(np.float64)
+
+    if crop_border != 0:
+        img1 = img1[crop_border:-crop_border, crop_border:-crop_border, ...]
+        img2 = img2[crop_border:-crop_border, crop_border:-crop_border, ...]
+
+    if test_y_channel:
+        img1 = to_y_channel(img1)
+        img2 = to_y_channel(img2)
+
+    mse = np.mean((img1 - img2) ** 2)
+    if mse == 0:
+        return float('inf')
+    return 20. * np.log10(255. / np.sqrt(mse))
+
+
+def _ssim(img1, img2):
+    """Calculate SSIM (structural similarity) for one channel images.
+    It is called by func:`calculate_ssim`.
+    Args:
+        img1 (ndarray): Images with range [0, 255] with order 'HWC'.
+        img2 (ndarray): Images with range [0, 255] with order 'HWC'.
+    Returns:
+        float: ssim result.
+    """
+
+    C1 = (0.01 * 255) ** 2
+    C2 = (0.03 * 255) ** 2
+
+    img1 = img1.astype(np.float64)
+    img2 = img2.astype(np.float64)
+    kernel = cv2.getGaussianKernel(11, 1.5)
+    window = np.outer(kernel, kernel.transpose())
+
+    mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5]
+    mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5]
+    mu1_sq = mu1 ** 2
+    mu2_sq = mu2 ** 2
+    mu1_mu2 = mu1 * mu2
+    sigma1_sq = cv2.filter2D(img1 ** 2, -1, window)[5:-5, 5:-5] - mu1_sq
+    sigma2_sq = cv2.filter2D(img2 ** 2, -1, window)[5:-5, 5:-5] - mu2_sq
+    sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2
+
+    ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) * (sigma1_sq + sigma2_sq + C2))
+    return ssim_map.mean()
+
+
+def calculate_ssim(img1, img2, crop_border, input_order='HWC', test_y_channel=False):
+    """Calculate SSIM (structural similarity).
+    Ref:
+    Image quality assessment: From error visibility to structural similarity
+    The results are the same as that of the official released MATLAB code in
+    https://ece.uwaterloo.ca/~z70wang/research/ssim/.
+    For three-channel images, SSIM is calculated for each channel and then
+    averaged.
+    Args:
+        img1 (ndarray): Images with range [0, 255].
+        img2 (ndarray): Images with range [0, 255].
+        crop_border (int): Cropped pixels in each edge of an image. These
+            pixels are not involved in the SSIM calculation.
+        input_order (str): Whether the input order is 'HWC' or 'CHW'.
+            Default: 'HWC'.
+        test_y_channel (bool): Test on Y channel of YCbCr. Default: False.
+    Returns:
+        float: ssim result.
+    """
+
+    assert img1.shape == img2.shape, (f'Image shapes are differnet: {img1.shape}, {img2.shape}.')
+    if input_order not in ['HWC', 'CHW']:
+        raise ValueError(f'Wrong input_order {input_order}. Supported input_orders are ' '"HWC" and "CHW"')
+    img1 = reorder_image(img1, input_order=input_order)
+    img2 = reorder_image(img2, input_order=input_order)
+    img1 = img1.astype(np.float64)
+    img2 = img2.astype(np.float64)
+
+    if crop_border != 0:
+        img1 = img1[crop_border:-crop_border, crop_border:-crop_border, ...]
+        img2 = img2[crop_border:-crop_border, crop_border:-crop_border, ...]
+
+    if test_y_channel:
+        img1 = to_y_channel(img1)
+        img2 = to_y_channel(img2)
+
+    ssims = []
+    for i in range(img1.shape[2]):
+        ssims.append(_ssim(img1[..., i], img2[..., i]))
+    return np.array(ssims).mean()
+
+
+def _blocking_effect_factor(im):
+    block_size = 8
+
+    block_horizontal_positions = torch.arange(7, im.shape[3] - 1, 8)
+    block_vertical_positions = torch.arange(7, im.shape[2] - 1, 8)
+
+    horizontal_block_difference = (
+                (im[:, :, :, block_horizontal_positions] - im[:, :, :, block_horizontal_positions + 1]) ** 2).sum(
+        3).sum(2).sum(1)
+    vertical_block_difference = (
+                (im[:, :, block_vertical_positions, :] - im[:, :, block_vertical_positions + 1, :]) ** 2).sum(3).sum(
+        2).sum(1)
+
+    nonblock_horizontal_positions = np.setdiff1d(torch.arange(0, im.shape[3] - 1), block_horizontal_positions)
+    nonblock_vertical_positions = np.setdiff1d(torch.arange(0, im.shape[2] - 1), block_vertical_positions)
+
+    horizontal_nonblock_difference = (
+                (im[:, :, :, nonblock_horizontal_positions] - im[:, :, :, nonblock_horizontal_positions + 1]) ** 2).sum(
+        3).sum(2).sum(1)
+    vertical_nonblock_difference = (
+                (im[:, :, nonblock_vertical_positions, :] - im[:, :, nonblock_vertical_positions + 1, :]) ** 2).sum(
+        3).sum(2).sum(1)
+
+    n_boundary_horiz = im.shape[2] * (im.shape[3] // block_size - 1)
+    n_boundary_vert = im.shape[3] * (im.shape[2] // block_size - 1)
+    boundary_difference = (horizontal_block_difference + vertical_block_difference) / (
+                n_boundary_horiz + n_boundary_vert)
+
+    n_nonboundary_horiz = im.shape[2] * (im.shape[3] - 1) - n_boundary_horiz
+    n_nonboundary_vert = im.shape[3] * (im.shape[2] - 1) - n_boundary_vert
+    nonboundary_difference = (horizontal_nonblock_difference + vertical_nonblock_difference) / (
+                n_nonboundary_horiz + n_nonboundary_vert)
+
+    scaler = np.log2(block_size) / np.log2(min([im.shape[2], im.shape[3]]))
+    bef = scaler * (boundary_difference - nonboundary_difference)
+
+    bef[boundary_difference <= nonboundary_difference] = 0
+    return bef
+
+
+def calculate_psnrb(img1, img2, crop_border, input_order='HWC', test_y_channel=False):
+    """Calculate PSNR-B (Peak Signal-to-Noise Ratio).
+    Ref: Quality assessment of deblocked images, for JPEG image deblocking evaluation
+    # https://gitlab.com/Queuecumber/quantization-guided-ac/-/blob/master/metrics/psnrb.py
+    Args:
+        img1 (ndarray): Images with range [0, 255].
+        img2 (ndarray): Images with range [0, 255].
+        crop_border (int): Cropped pixels in each edge of an image. These
+            pixels are not involved in the PSNR calculation.
+        input_order (str): Whether the input order is 'HWC' or 'CHW'.
+            Default: 'HWC'.
+        test_y_channel (bool): Test on Y channel of YCbCr. Default: False.
+    Returns:
+        float: psnr result.
+    """
+
+    assert img1.shape == img2.shape, (f'Image shapes are differnet: {img1.shape}, {img2.shape}.')
+    if input_order not in ['HWC', 'CHW']:
+        raise ValueError(f'Wrong input_order {input_order}. Supported input_orders are ' '"HWC" and "CHW"')
+    img1 = reorder_image(img1, input_order=input_order)
+    img2 = reorder_image(img2, input_order=input_order)
+    img1 = img1.astype(np.float64)
+    img2 = img2.astype(np.float64)
+
+    if crop_border != 0:
+        img1 = img1[crop_border:-crop_border, crop_border:-crop_border, ...]
+        img2 = img2[crop_border:-crop_border, crop_border:-crop_border, ...]
+
+    if test_y_channel:
+        img1 = to_y_channel(img1)
+        img2 = to_y_channel(img2)
+
+    # follow https://gitlab.com/Queuecumber/quantization-guided-ac/-/blob/master/metrics/psnrb.py
+    img1 = torch.from_numpy(img1).permute(2, 0, 1).unsqueeze(0) / 255.
+    img2 = torch.from_numpy(img2).permute(2, 0, 1).unsqueeze(0) / 255.
+
+    total = 0
+    for c in range(img1.shape[1]):
+        mse = torch.nn.functional.mse_loss(img1[:, c:c + 1, :, :], img2[:, c:c + 1, :, :], reduction='none')
+        bef = _blocking_effect_factor(img1[:, c:c + 1, :, :])
+
+        mse = mse.view(mse.shape[0], -1).mean(1)
+        total += 10 * torch.log10(1 / (mse + bef))
+
+    return float(total) / img1.shape[1]
+
+
+def reorder_image(img, input_order='HWC'):
+    """Reorder images to 'HWC' order.
+    If the input_order is (h, w), return (h, w, 1);
+    If the input_order is (c, h, w), return (h, w, c);
+    If the input_order is (h, w, c), return as it is.
+    Args:
+        img (ndarray): Input image.
+        input_order (str): Whether the input order is 'HWC' or 'CHW'.
+            If the input image shape is (h, w), input_order will not have
+            effects. Default: 'HWC'.
+    Returns:
+        ndarray: reordered image.
+    """
+
+    if input_order not in ['HWC', 'CHW']:
+        raise ValueError(f'Wrong input_order {input_order}. Supported input_orders are ' "'HWC' and 'CHW'")
+    if len(img.shape) == 2:
+        img = img[..., None]
+    if input_order == 'CHW':
+        img = img.transpose(1, 2, 0)
+    return img
+
+
+def to_y_channel(img):
+    """Change to Y channel of YCbCr.
+    Args:
+        img (ndarray): Images with range [0, 255].
+    Returns:
+        (ndarray): Images with range [0, 255] (float type) without round.
+    """
+    img = img.astype(np.float32) / 255.
+    if img.ndim == 3 and img.shape[2] == 3:
+        img = bgr2ycbcr(img, y_only=True)
+        img = img[..., None]
+    return img * 255.
+
+
+def _convert_input_type_range(img):
+    """Convert the type and range of the input image.
+    It converts the input image to np.float32 type and range of [0, 1].
+    It is mainly used for pre-processing the input image in colorspace
+    convertion functions such as rgb2ycbcr and ycbcr2rgb.
+    Args:
+        img (ndarray): The input image. It accepts:
+            1. np.uint8 type with range [0, 255];
+            2. np.float32 type with range [0, 1].
+    Returns:
+        (ndarray): The converted image with type of np.float32 and range of
+            [0, 1].
+    """
+    img_type = img.dtype
+    img = img.astype(np.float32)
+    if img_type == np.float32:
+        pass
+    elif img_type == np.uint8:
+        img /= 255.
+    else:
+        raise TypeError('The img type should be np.float32 or np.uint8, ' f'but got {img_type}')
+    return img
+
+
+def _convert_output_type_range(img, dst_type):
+    """Convert the type and range of the image according to dst_type.
+    It converts the image to desired type and range. If `dst_type` is np.uint8,
+    images will be converted to np.uint8 type with range [0, 255]. If
+    `dst_type` is np.float32, it converts the image to np.float32 type with
+    range [0, 1].
+    It is mainly used for post-processing images in colorspace convertion
+    functions such as rgb2ycbcr and ycbcr2rgb.
+    Args:
+        img (ndarray): The image to be converted with np.float32 type and
+            range [0, 255].
+        dst_type (np.uint8 | np.float32): If dst_type is np.uint8, it
+            converts the image to np.uint8 type with range [0, 255]. If
+            dst_type is np.float32, it converts the image to np.float32 type
+            with range [0, 1].
+    Returns:
+        (ndarray): The converted image with desired type and range.
+    """
+    if dst_type not in (np.uint8, np.float32):
+        raise TypeError('The dst_type should be np.float32 or np.uint8, ' f'but got {dst_type}')
+    if dst_type == np.uint8:
+        img = img.round()
+    else:
+        img /= 255.
+    return img.astype(dst_type)
+
+
+def bgr2ycbcr(img, y_only=False):
+    """Convert a BGR image to YCbCr image.
+    The bgr version of rgb2ycbcr.
+    It implements the ITU-R BT.601 conversion for standard-definition
+    television. See more details in
+    https://en.wikipedia.org/wiki/YCbCr#ITU-R_BT.601_conversion.
+    It differs from a similar function in cv2.cvtColor: `BGR <-> YCrCb`.
+    In OpenCV, it implements a JPEG conversion. See more details in
+    https://en.wikipedia.org/wiki/YCbCr#JPEG_conversion.
+    Args:
+        img (ndarray): The input image. It accepts:
+            1. np.uint8 type with range [0, 255];
+            2. np.float32 type with range [0, 1].
+        y_only (bool): Whether to only return Y channel. Default: False.
+    Returns:
+        ndarray: The converted YCbCr image. The output image has the same type
+            and range as input image.
+    """
+    img_type = img.dtype
+    img = _convert_input_type_range(img)
+    if y_only:
+        out_img = np.dot(img, [24.966, 128.553, 65.481]) + 16.0
+    else:
+        out_img = np.matmul(
+            img, [[24.966, 112.0, -18.214], [128.553, -74.203, -93.786], [65.481, -37.797, 112.0]]) + [16, 128, 128]
+    out_img = _convert_output_type_range(out_img, img_type)
+    return out_img