Upload 14 files

realesrgan/__init__.py

@@ -0,0 +1,6 @@
+# flake8: noqa
+from .archs import *
+from .data import *
+from .models import *
+from .utils import *
+from .version import *

realesrgan/archs/__init__.py

@@ -0,0 +1,10 @@
+import importlib
+from basicsr.utils import scandir
+from os import path as osp
+
+# automatically scan and import arch modules for registry
+# scan all the files that end with '_arch.py' under the archs folder
+arch_folder = osp.dirname(osp.abspath(__file__))
+arch_filenames = [osp.splitext(osp.basename(v))[0] for v in scandir(arch_folder) if v.endswith('_arch.py')]
+# import all the arch modules
+_arch_modules = [importlib.import_module(f'realesrgan.archs.{file_name}') for file_name in arch_filenames]

realesrgan/archs/discriminator_arch.py

@@ -0,0 +1,67 @@
+from basicsr.utils.registry import ARCH_REGISTRY
+from torch import nn as nn
+from torch.nn import functional as F
+from torch.nn.utils import spectral_norm
+
+
+@ARCH_REGISTRY.register()
+class UNetDiscriminatorSN(nn.Module):
+    """Defines a U-Net discriminator with spectral normalization (SN)
+
+    It is used in Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data.
+
+    Arg:
+        num_in_ch (int): Channel number of inputs. Default: 3.
+        num_feat (int): Channel number of base intermediate features. Default: 64.
+        skip_connection (bool): Whether to use skip connections between U-Net. Default: True.
+    """
+
+    def __init__(self, num_in_ch, num_feat=64, skip_connection=True):
+        super(UNetDiscriminatorSN, self).__init__()
+        self.skip_connection = skip_connection
+        norm = spectral_norm
+        # the first convolution
+        self.conv0 = nn.Conv2d(num_in_ch, num_feat, kernel_size=3, stride=1, padding=1)
+        # downsample
+        self.conv1 = norm(nn.Conv2d(num_feat, num_feat * 2, 4, 2, 1, bias=False))
+        self.conv2 = norm(nn.Conv2d(num_feat * 2, num_feat * 4, 4, 2, 1, bias=False))
+        self.conv3 = norm(nn.Conv2d(num_feat * 4, num_feat * 8, 4, 2, 1, bias=False))
+        # upsample
+        self.conv4 = norm(nn.Conv2d(num_feat * 8, num_feat * 4, 3, 1, 1, bias=False))
+        self.conv5 = norm(nn.Conv2d(num_feat * 4, num_feat * 2, 3, 1, 1, bias=False))
+        self.conv6 = norm(nn.Conv2d(num_feat * 2, num_feat, 3, 1, 1, bias=False))
+        # extra convolutions
+        self.conv7 = norm(nn.Conv2d(num_feat, num_feat, 3, 1, 1, bias=False))
+        self.conv8 = norm(nn.Conv2d(num_feat, num_feat, 3, 1, 1, bias=False))
+        self.conv9 = nn.Conv2d(num_feat, 1, 3, 1, 1)
+
+    def forward(self, x):
+        # downsample
+        x0 = F.leaky_relu(self.conv0(x), negative_slope=0.2, inplace=True)
+        x1 = F.leaky_relu(self.conv1(x0), negative_slope=0.2, inplace=True)
+        x2 = F.leaky_relu(self.conv2(x1), negative_slope=0.2, inplace=True)
+        x3 = F.leaky_relu(self.conv3(x2), negative_slope=0.2, inplace=True)
+
+        # upsample
+        x3 = F.interpolate(x3, scale_factor=2, mode='bilinear', align_corners=False)
+        x4 = F.leaky_relu(self.conv4(x3), negative_slope=0.2, inplace=True)
+
+        if self.skip_connection:
+            x4 = x4 + x2
+        x4 = F.interpolate(x4, scale_factor=2, mode='bilinear', align_corners=False)
+        x5 = F.leaky_relu(self.conv5(x4), negative_slope=0.2, inplace=True)
+
+        if self.skip_connection:
+            x5 = x5 + x1
+        x5 = F.interpolate(x5, scale_factor=2, mode='bilinear', align_corners=False)
+        x6 = F.leaky_relu(self.conv6(x5), negative_slope=0.2, inplace=True)
+
+        if self.skip_connection:
+            x6 = x6 + x0
+
+        # extra convolutions
+        out = F.leaky_relu(self.conv7(x6), negative_slope=0.2, inplace=True)
+        out = F.leaky_relu(self.conv8(out), negative_slope=0.2, inplace=True)
+        out = self.conv9(out)
+
+        return out

realesrgan/archs/srvgg_arch.py

@@ -0,0 +1,69 @@
+from basicsr.utils.registry import ARCH_REGISTRY
+from torch import nn as nn
+from torch.nn import functional as F
+
+
+@ARCH_REGISTRY.register()
+class SRVGGNetCompact(nn.Module):
+    """A compact VGG-style network structure for super-resolution.
+
+    It is a compact network structure, which performs upsampling in the last layer and no convolution is
+    conducted on the HR feature space.
+
+    Args:
+        num_in_ch (int): Channel number of inputs. Default: 3.
+        num_out_ch (int): Channel number of outputs. Default: 3.
+        num_feat (int): Channel number of intermediate features. Default: 64.
+        num_conv (int): Number of convolution layers in the body network. Default: 16.
+        upscale (int): Upsampling factor. Default: 4.
+        act_type (str): Activation type, options: 'relu', 'prelu', 'leakyrelu'. Default: prelu.
+    """
+
+    def __init__(self, num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=16, upscale=4, act_type='prelu'):
+        super(SRVGGNetCompact, self).__init__()
+        self.num_in_ch = num_in_ch
+        self.num_out_ch = num_out_ch
+        self.num_feat = num_feat
+        self.num_conv = num_conv
+        self.upscale = upscale
+        self.act_type = act_type
+
+        self.body = nn.ModuleList()
+        # the first conv
+        self.body.append(nn.Conv2d(num_in_ch, num_feat, 3, 1, 1))
+        # the first activation
+        if act_type == 'relu':
+            activation = nn.ReLU(inplace=True)
+        elif act_type == 'prelu':
+            activation = nn.PReLU(num_parameters=num_feat)
+        elif act_type == 'leakyrelu':
+            activation = nn.LeakyReLU(negative_slope=0.1, inplace=True)
+        self.body.append(activation)
+
+        # the body structure
+        for _ in range(num_conv):
+            self.body.append(nn.Conv2d(num_feat, num_feat, 3, 1, 1))
+            # activation
+            if act_type == 'relu':
+                activation = nn.ReLU(inplace=True)
+            elif act_type == 'prelu':
+                activation = nn.PReLU(num_parameters=num_feat)
+            elif act_type == 'leakyrelu':
+                activation = nn.LeakyReLU(negative_slope=0.1, inplace=True)
+            self.body.append(activation)
+
+        # the last conv
+        self.body.append(nn.Conv2d(num_feat, num_out_ch * upscale * upscale, 3, 1, 1))
+        # upsample
+        self.upsampler = nn.PixelShuffle(upscale)
+
+    def forward(self, x):
+        out = x
+        for i in range(0, len(self.body)):
+            out = self.body[i](out)
+
+        out = self.upsampler(out)
+        # add the nearest upsampled image, so that the network learns the residual
+        base = F.interpolate(x, scale_factor=self.upscale, mode='nearest')
+        out += base
+        return out

realesrgan/data/__init__.py

@@ -0,0 +1,10 @@
+import importlib
+from basicsr.utils import scandir
+from os import path as osp
+
+# automatically scan and import dataset modules for registry
+# scan all the files that end with '_dataset.py' under the data folder
+data_folder = osp.dirname(osp.abspath(__file__))
+dataset_filenames = [osp.splitext(osp.basename(v))[0] for v in scandir(data_folder) if v.endswith('_dataset.py')]
+# import all the dataset modules
+_dataset_modules = [importlib.import_module(f'realesrgan.data.{file_name}') for file_name in dataset_filenames]

realesrgan/data/realesrgan_dataset.py

@@ -0,0 +1,192 @@
+import cv2
+import math
+import numpy as np
+import os
+import os.path as osp
+import random
+import time
+import torch
+from basicsr.data.degradations import circular_lowpass_kernel, random_mixed_kernels
+from basicsr.data.transforms import augment
+from basicsr.utils import FileClient, get_root_logger, imfrombytes, img2tensor
+from basicsr.utils.registry import DATASET_REGISTRY
+from torch.utils import data as data
+
+
+@DATASET_REGISTRY.register()
+class RealESRGANDataset(data.Dataset):
+    """Dataset used for Real-ESRGAN model:
+    Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data.
+
+    It loads gt (Ground-Truth) images, and augments them.
+    It also generates blur kernels and sinc kernels for generating low-quality images.
+    Note that the low-quality images are processed in tensors on GPUS for faster processing.
+
+    Args:
+        opt (dict): Config for train datasets. It contains the following keys:
+            dataroot_gt (str): Data root path for gt.
+            meta_info (str): Path for meta information file.
+            io_backend (dict): IO backend type and other kwarg.
+            use_hflip (bool): Use horizontal flips.
+            use_rot (bool): Use rotation (use vertical flip and transposing h and w for implementation).
+            Please see more options in the codes.
+    """
+
+    def __init__(self, opt):
+        super(RealESRGANDataset, self).__init__()
+        self.opt = opt
+        self.file_client = None
+        self.io_backend_opt = opt['io_backend']
+        self.gt_folder = opt['dataroot_gt']
+
+        # file client (lmdb io backend)
+        if self.io_backend_opt['type'] == 'lmdb':
+            self.io_backend_opt['db_paths'] = [self.gt_folder]
+            self.io_backend_opt['client_keys'] = ['gt']
+            if not self.gt_folder.endswith('.lmdb'):
+                raise ValueError(f"'dataroot_gt' should end with '.lmdb', but received {self.gt_folder}")
+            with open(osp.join(self.gt_folder, 'meta_info.txt')) as fin:
+                self.paths = [line.split('.')[0] for line in fin]
+        else:
+            # disk backend with meta_info
+            # Each line in the meta_info describes the relative path to an image
+            with open(self.opt['meta_info']) as fin:
+                paths = [line.strip().split(' ')[0] for line in fin]
+                self.paths = [os.path.join(self.gt_folder, v) for v in paths]
+
+        # blur settings for the first degradation
+        self.blur_kernel_size = opt['blur_kernel_size']
+        self.kernel_list = opt['kernel_list']
+        self.kernel_prob = opt['kernel_prob']  # a list for each kernel probability
+        self.blur_sigma = opt['blur_sigma']
+        self.betag_range = opt['betag_range']  # betag used in generalized Gaussian blur kernels
+        self.betap_range = opt['betap_range']  # betap used in plateau blur kernels
+        self.sinc_prob = opt['sinc_prob']  # the probability for sinc filters
+
+        # blur settings for the second degradation
+        self.blur_kernel_size2 = opt['blur_kernel_size2']
+        self.kernel_list2 = opt['kernel_list2']
+        self.kernel_prob2 = opt['kernel_prob2']
+        self.blur_sigma2 = opt['blur_sigma2']
+        self.betag_range2 = opt['betag_range2']
+        self.betap_range2 = opt['betap_range2']
+        self.sinc_prob2 = opt['sinc_prob2']
+
+        # a final sinc filter
+        self.final_sinc_prob = opt['final_sinc_prob']
+
+        self.kernel_range = [2 * v + 1 for v in range(3, 11)]  # kernel size ranges from 7 to 21
+        # TODO: kernel range is now hard-coded, should be in the configure file
+        self.pulse_tensor = torch.zeros(21, 21).float()  # convolving with pulse tensor brings no blurry effect
+        self.pulse_tensor[10, 10] = 1
+
+    def __getitem__(self, index):
+        if self.file_client is None:
+            self.file_client = FileClient(self.io_backend_opt.pop('type'), **self.io_backend_opt)
+
+        # -------------------------------- Load gt images -------------------------------- #
+        # Shape: (h, w, c); channel order: BGR; image range: [0, 1], float32.
+        gt_path = self.paths[index]
+        # avoid errors caused by high latency in reading files
+        retry = 3
+        while retry > 0:
+            try:
+                img_bytes = self.file_client.get(gt_path, 'gt')
+            except (IOError, OSError) as e:
+                logger = get_root_logger()
+                logger.warn(f'File client error: {e}, remaining retry times: {retry - 1}')
+                # change another file to read
+                index = random.randint(0, self.__len__())
+                gt_path = self.paths[index]
+                time.sleep(1)  # sleep 1s for occasional server congestion
+            else:
+                break
+            finally:
+                retry -= 1
+        img_gt = imfrombytes(img_bytes, float32=True)
+
+        # -------------------- Do augmentation for training: flip, rotation -------------------- #
+        img_gt = augment(img_gt, self.opt['use_hflip'], self.opt['use_rot'])
+
+        # crop or pad to 400
+        # TODO: 400 is hard-coded. You may change it accordingly
+        h, w = img_gt.shape[0:2]
+        crop_pad_size = 400
+        # pad
+        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)
+        # crop
+        if img_gt.shape[0] > crop_pad_size or img_gt.shape[1] > crop_pad_size:
+            h, w = img_gt.shape[0:2]
+            # randomly choose top and left coordinates
+            top = random.randint(0, h - crop_pad_size)
+            left = random.randint(0, w - crop_pad_size)
+            img_gt = img_gt[top:top + crop_pad_size, left:left + crop_pad_size, ...]
+
+        # ------------------------ Generate kernels (used in the first degradation) ------------------------ #
+        kernel_size = random.choice(self.kernel_range)
+        if np.random.uniform() < self.opt['sinc_prob']:
+            # this sinc filter setting is for kernels ranging from [7, 21]
+            if kernel_size < 13:
+                omega_c = np.random.uniform(np.pi / 3, np.pi)
+            else:
+                omega_c = np.random.uniform(np.pi / 5, np.pi)
+            kernel = circular_lowpass_kernel(omega_c, kernel_size, pad_to=False)
+        else:
+            kernel = random_mixed_kernels(
+                self.kernel_list,
+                self.kernel_prob,
+                kernel_size,
+                self.blur_sigma,
+                self.blur_sigma, [-math.pi, math.pi],
+                self.betag_range,
+                self.betap_range,
+                noise_range=None)
+        # pad kernel
+        pad_size = (21 - kernel_size) // 2
+        kernel = np.pad(kernel, ((pad_size, pad_size), (pad_size, pad_size)))
+
+        # ------------------------ Generate kernels (used in the second degradation) ------------------------ #
+        kernel_size = random.choice(self.kernel_range)
+        if np.random.uniform() < self.opt['sinc_prob2']:
+            if kernel_size < 13:
+                omega_c = np.random.uniform(np.pi / 3, np.pi)
+            else:
+                omega_c = np.random.uniform(np.pi / 5, np.pi)
+            kernel2 = circular_lowpass_kernel(omega_c, kernel_size, pad_to=False)
+        else:
+            kernel2 = random_mixed_kernels(
+                self.kernel_list2,
+                self.kernel_prob2,
+                kernel_size,
+                self.blur_sigma2,
+                self.blur_sigma2, [-math.pi, math.pi],
+                self.betag_range2,
+                self.betap_range2,
+                noise_range=None)
+
+        # pad kernel
+        pad_size = (21 - kernel_size) // 2
+        kernel2 = np.pad(kernel2, ((pad_size, pad_size), (pad_size, pad_size)))
+
+        # ------------------------------------- the final sinc kernel ------------------------------------- #
+        if np.random.uniform() < self.opt['final_sinc_prob']:
+            kernel_size = random.choice(self.kernel_range)
+            omega_c = np.random.uniform(np.pi / 3, np.pi)
+            sinc_kernel = circular_lowpass_kernel(omega_c, kernel_size, pad_to=21)
+            sinc_kernel = torch.FloatTensor(sinc_kernel)
+        else:
+            sinc_kernel = self.pulse_tensor
+
+        # BGR to RGB, HWC to CHW, numpy to tensor
+        img_gt = img2tensor([img_gt], bgr2rgb=True, float32=True)[0]
+        kernel = torch.FloatTensor(kernel)
+        kernel2 = torch.FloatTensor(kernel2)
+
+        return_d = {'gt': img_gt, 'kernel1': kernel, 'kernel2': kernel2, 'sinc_kernel': sinc_kernel, 'gt_path': gt_path}
+        return return_d
+
+    def __len__(self):
+        return len(self.paths)

realesrgan/data/realesrgan_paired_dataset.py

@@ -0,0 +1,108 @@
+import os
+from basicsr.data.data_util import paired_paths_from_folder, paired_paths_from_lmdb
+from basicsr.data.transforms import augment, paired_random_crop
+from basicsr.utils import FileClient, imfrombytes, img2tensor
+from basicsr.utils.registry import DATASET_REGISTRY
+from torch.utils import data as data
+from torchvision.transforms.functional import normalize
+
+
+@DATASET_REGISTRY.register()
+class RealESRGANPairedDataset(data.Dataset):
+    """Paired image dataset for image restoration.
+
+    Read LQ (Low Quality, e.g. LR (Low Resolution), blurry, noisy, etc) and GT image pairs.
+
+    There are three modes:
+    1. 'lmdb': Use lmdb files.
+        If opt['io_backend'] == lmdb.
+    2. 'meta_info': Use meta information file to generate paths.
+        If opt['io_backend'] != lmdb and opt['meta_info'] is not None.
+    3. 'folder': Scan folders to generate paths.
+        The rest.
+
+    Args:
+        opt (dict): Config for train datasets. It contains the following keys:
+            dataroot_gt (str): Data root path for gt.
+            dataroot_lq (str): Data root path for lq.
+            meta_info (str): Path for meta information file.
+            io_backend (dict): IO backend type and other kwarg.
+            filename_tmpl (str): Template for each filename. Note that the template excludes the file extension.
+                Default: '{}'.
+            gt_size (int): Cropped patched size for gt patches.
+            use_hflip (bool): Use horizontal flips.
+            use_rot (bool): Use rotation (use vertical flip and transposing h
+                and w for implementation).
+
+            scale (bool): Scale, which will be added automatically.
+            phase (str): 'train' or 'val'.
+    """
+
+    def __init__(self, opt):
+        super(RealESRGANPairedDataset, self).__init__()
+        self.opt = opt
+        self.file_client = None
+        self.io_backend_opt = opt['io_backend']
+        # mean and std for normalizing the input images
+        self.mean = opt['mean'] if 'mean' in opt else None
+        self.std = opt['std'] if 'std' in opt else None
+
+        self.gt_folder, self.lq_folder = opt['dataroot_gt'], opt['dataroot_lq']
+        self.filename_tmpl = opt['filename_tmpl'] if 'filename_tmpl' in opt else '{}'
+
+        # file client (lmdb io backend)
+        if self.io_backend_opt['type'] == 'lmdb':
+            self.io_backend_opt['db_paths'] = [self.lq_folder, self.gt_folder]
+            self.io_backend_opt['client_keys'] = ['lq', 'gt']
+            self.paths = paired_paths_from_lmdb([self.lq_folder, self.gt_folder], ['lq', 'gt'])
+        elif 'meta_info' in self.opt and self.opt['meta_info'] is not None:
+            # disk backend with meta_info
+            # Each line in the meta_info describes the relative path to an image
+            with open(self.opt['meta_info']) as fin:
+                paths = [line.strip() for line in fin]
+            self.paths = []
+            for path in paths:
+                gt_path, lq_path = path.split(', ')
+                gt_path = os.path.join(self.gt_folder, gt_path)
+                lq_path = os.path.join(self.lq_folder, lq_path)
+                self.paths.append(dict([('gt_path', gt_path), ('lq_path', lq_path)]))
+        else:
+            # disk backend
+            # it will scan the whole folder to get meta info
+            # it will be time-consuming for folders with too many files. It is recommended using an extra meta txt file
+            self.paths = paired_paths_from_folder([self.lq_folder, self.gt_folder], ['lq', 'gt'], self.filename_tmpl)
+
+    def __getitem__(self, index):
+        if self.file_client is None:
+            self.file_client = FileClient(self.io_backend_opt.pop('type'), **self.io_backend_opt)
+
+        scale = self.opt['scale']
+
+        # Load gt and lq images. Dimension order: HWC; channel order: BGR;
+        # image range: [0, 1], float32.
+        gt_path = self.paths[index]['gt_path']
+        img_bytes = self.file_client.get(gt_path, 'gt')
+        img_gt = imfrombytes(img_bytes, float32=True)
+        lq_path = self.paths[index]['lq_path']
+        img_bytes = self.file_client.get(lq_path, 'lq')
+        img_lq = imfrombytes(img_bytes, float32=True)
+
+        # augmentation for training
+        if self.opt['phase'] == 'train':
+            gt_size = self.opt['gt_size']
+            # random crop
+            img_gt, img_lq = paired_random_crop(img_gt, img_lq, gt_size, scale, gt_path)
+            # flip, rotation
+            img_gt, img_lq = augment([img_gt, img_lq], self.opt['use_hflip'], self.opt['use_rot'])
+
+        # BGR to RGB, HWC to CHW, numpy to tensor
+        img_gt, img_lq = img2tensor([img_gt, img_lq], bgr2rgb=True, float32=True)
+        # normalize
+        if self.mean is not None or self.std is not None:
+            normalize(img_lq, self.mean, self.std, inplace=True)
+            normalize(img_gt, self.mean, self.std, inplace=True)
+
+        return {'lq': img_lq, 'gt': img_gt, 'lq_path': lq_path, 'gt_path': gt_path}
+
+    def __len__(self):
+        return len(self.paths)

realesrgan/models/__init__.py

@@ -0,0 +1,10 @@
+import importlib
+from basicsr.utils import scandir
+from os import path as osp
+
+# automatically scan and import model modules for registry
+# scan all the files that end with '_model.py' under the model folder
+model_folder = osp.dirname(osp.abspath(__file__))
+model_filenames = [osp.splitext(osp.basename(v))[0] for v in scandir(model_folder) if v.endswith('_model.py')]
+# import all the model modules
+_model_modules = [importlib.import_module(f'realesrgan.models.{file_name}') for file_name in model_filenames]

realesrgan/models/realesrgan_model.py

@@ -0,0 +1,258 @@
+import numpy as np
+import random
+import torch
+from basicsr.data.degradations import random_add_gaussian_noise_pt, random_add_poisson_noise_pt
+from basicsr.data.transforms import paired_random_crop
+from basicsr.models.srgan_model import SRGANModel
+from basicsr.utils import DiffJPEG, USMSharp
+from basicsr.utils.img_process_util import filter2D
+from basicsr.utils.registry import MODEL_REGISTRY
+from collections import OrderedDict
+from torch.nn import functional as F
+
+
+@MODEL_REGISTRY.register()
+class RealESRGANModel(SRGANModel):
+    """RealESRGAN Model for Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data.
+
+    It mainly performs:
+    1. randomly synthesize LQ images in GPU tensors
+    2. optimize the networks with GAN training.
+    """
+
+    def __init__(self, opt):
+        super(RealESRGANModel, self).__init__(opt)
+        self.jpeger = DiffJPEG(differentiable=False).cuda()  # simulate JPEG compression artifacts
+        self.usm_sharpener = USMSharp().cuda()  # do usm sharpening
+        self.queue_size = opt.get('queue_size', 180)
+
+    @torch.no_grad()
+    def _dequeue_and_enqueue(self):
+        """It is the training pair pool for increasing the diversity in a batch.
+
+        Batch processing limits the diversity of synthetic degradations in a batch. For example, samples in a
+        batch could not have different resize scaling factors. Therefore, we employ this training pair pool
+        to increase the degradation diversity in a batch.
+        """
+        # initialize
+        b, c, h, w = self.lq.size()
+        if not hasattr(self, 'queue_lr'):
+            assert self.queue_size % b == 0, f'queue size {self.queue_size} should be divisible by batch size {b}'
+            self.queue_lr = torch.zeros(self.queue_size, c, h, w).cuda()
+            _, c, h, w = self.gt.size()
+            self.queue_gt = torch.zeros(self.queue_size, c, h, w).cuda()
+            self.queue_ptr = 0
+        if self.queue_ptr == self.queue_size:  # the pool is full
+            # do dequeue and enqueue
+            # shuffle
+            idx = torch.randperm(self.queue_size)
+            self.queue_lr = self.queue_lr[idx]
+            self.queue_gt = self.queue_gt[idx]
+            # get first b samples
+            lq_dequeue = self.queue_lr[0:b, :, :, :].clone()
+            gt_dequeue = self.queue_gt[0:b, :, :, :].clone()
+            # update the queue
+            self.queue_lr[0:b, :, :, :] = self.lq.clone()
+            self.queue_gt[0:b, :, :, :] = self.gt.clone()
+
+            self.lq = lq_dequeue
+            self.gt = gt_dequeue
+        else:
+            # only do enqueue
+            self.queue_lr[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.lq.clone()
+            self.queue_gt[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.gt.clone()
+            self.queue_ptr = self.queue_ptr + b
+
+    @torch.no_grad()
+    def feed_data(self, data):
+        """Accept data from dataloader, and then add two-order degradations to obtain LQ images.
+        """
+        if self.is_train and self.opt.get('high_order_degradation', True):
+            # training data synthesis
+            self.gt = data['gt'].to(self.device)
+            self.gt_usm = self.usm_sharpener(self.gt)
+
+            self.kernel1 = data['kernel1'].to(self.device)
+            self.kernel2 = data['kernel2'].to(self.device)
+            self.sinc_kernel = data['sinc_kernel'].to(self.device)
+
+            ori_h, ori_w = self.gt.size()[2:4]
+
+            # ----------------------- The first degradation process ----------------------- #
+            # blur
+            out = filter2D(self.gt_usm, self.kernel1)
+            # random resize
+            updown_type = random.choices(['up', 'down', 'keep'], self.opt['resize_prob'])[0]
+            if updown_type == 'up':
+                scale = np.random.uniform(1, self.opt['resize_range'][1])
+            elif updown_type == 'down':
+                scale = np.random.uniform(self.opt['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 = self.opt['gray_noise_prob']
+            if np.random.uniform() < self.opt['gaussian_noise_prob']:
+                out = random_add_gaussian_noise_pt(
+                    out, sigma_range=self.opt['noise_range'], clip=True, rounds=False, gray_prob=gray_noise_prob)
+            else:
+                out = random_add_poisson_noise_pt(
+                    out,
+                    scale_range=self.opt['poisson_scale_range'],
+                    gray_prob=gray_noise_prob,
+                    clip=True,
+                    rounds=False)
+            # JPEG compression
+            jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range'])
+            out = torch.clamp(out, 0, 1)  # clamp to [0, 1], otherwise JPEGer will result in unpleasant artifacts
+            out = self.jpeger(out, quality=jpeg_p)
+
+            # ----------------------- The second degradation process ----------------------- #
+            # blur
+            if np.random.uniform() < self.opt['second_blur_prob']:
+                out = filter2D(out, self.kernel2)
+            # random resize
+            updown_type = random.choices(['up', 'down', 'keep'], self.opt['resize_prob2'])[0]
+            if updown_type == 'up':
+                scale = np.random.uniform(1, self.opt['resize_range2'][1])
+            elif updown_type == 'down':
+                scale = np.random.uniform(self.opt['resize_range2'][0], 1)
+            else:
+                scale = 1
+            mode = random.choice(['area', 'bilinear', 'bicubic'])
+            out = F.interpolate(
+                out, size=(int(ori_h / self.opt['scale'] * scale), int(ori_w / self.opt['scale'] * scale)), mode=mode)
+            # add noise
+            gray_noise_prob = self.opt['gray_noise_prob2']
+            if np.random.uniform() < self.opt['gaussian_noise_prob2']:
+                out = random_add_gaussian_noise_pt(
+                    out, sigma_range=self.opt['noise_range2'], clip=True, rounds=False, gray_prob=gray_noise_prob)
+            else:
+                out = random_add_poisson_noise_pt(
+                    out,
+                    scale_range=self.opt['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 np.random.uniform() < 0.5:
+                # resize back + the final sinc filter
+                mode = random.choice(['area', 'bilinear', 'bicubic'])
+                out = F.interpolate(out, size=(ori_h // self.opt['scale'], ori_w // self.opt['scale']), mode=mode)
+                out = filter2D(out, self.sinc_kernel)
+                # JPEG compression
+                jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range2'])
+                out = torch.clamp(out, 0, 1)
+                out = self.jpeger(out, quality=jpeg_p)
+            else:
+                # JPEG compression
+                jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range2'])
+                out = torch.clamp(out, 0, 1)
+                out = self.jpeger(out, quality=jpeg_p)
+                # resize back + the final sinc filter
+                mode = random.choice(['area', 'bilinear', 'bicubic'])
+                out = F.interpolate(out, size=(ori_h // self.opt['scale'], ori_w // self.opt['scale']), mode=mode)
+                out = filter2D(out, self.sinc_kernel)
+
+            # clamp and round
+            self.lq = torch.clamp((out * 255.0).round(), 0, 255) / 255.
+
+            # random crop
+            gt_size = self.opt['gt_size']
+            (self.gt, self.gt_usm), self.lq = paired_random_crop([self.gt, self.gt_usm], self.lq, gt_size,
+                                                                 self.opt['scale'])
+
+            # training pair pool
+            self._dequeue_and_enqueue()
+            # sharpen self.gt again, as we have changed the self.gt with self._dequeue_and_enqueue
+            self.gt_usm = self.usm_sharpener(self.gt)
+            self.lq = self.lq.contiguous()  # for the warning: grad and param do not obey the gradient layout contract
+        else:
+            # for paired training or validation
+            self.lq = data['lq'].to(self.device)
+            if 'gt' in data:
+                self.gt = data['gt'].to(self.device)
+                self.gt_usm = self.usm_sharpener(self.gt)
+
+    def nondist_validation(self, dataloader, current_iter, tb_logger, save_img):
+        # do not use the synthetic process during validation
+        self.is_train = False
+        super(RealESRGANModel, self).nondist_validation(dataloader, current_iter, tb_logger, save_img)
+        self.is_train = True
+
+    def optimize_parameters(self, current_iter):
+        # usm sharpening
+        l1_gt = self.gt_usm
+        percep_gt = self.gt_usm
+        gan_gt = self.gt_usm
+        if self.opt['l1_gt_usm'] is False:
+            l1_gt = self.gt
+        if self.opt['percep_gt_usm'] is False:
+            percep_gt = self.gt
+        if self.opt['gan_gt_usm'] is False:
+            gan_gt = self.gt
+
+        # optimize net_g
+        for p in self.net_d.parameters():
+            p.requires_grad = False
+
+        self.optimizer_g.zero_grad()
+        self.output = self.net_g(self.lq)
+
+        l_g_total = 0
+        loss_dict = OrderedDict()
+        if (current_iter % self.net_d_iters == 0 and current_iter > self.net_d_init_iters):
+            # pixel loss
+            if self.cri_pix:
+                l_g_pix = self.cri_pix(self.output, l1_gt)
+                l_g_total += l_g_pix
+                loss_dict['l_g_pix'] = l_g_pix
+            # perceptual loss
+            if self.cri_perceptual:
+                l_g_percep, l_g_style = self.cri_perceptual(self.output, percep_gt)
+                if l_g_percep is not None:
+                    l_g_total += l_g_percep
+                    loss_dict['l_g_percep'] = l_g_percep
+                if l_g_style is not None:
+                    l_g_total += l_g_style
+                    loss_dict['l_g_style'] = l_g_style
+            # gan loss
+            fake_g_pred = self.net_d(self.output)
+            l_g_gan = self.cri_gan(fake_g_pred, True, is_disc=False)
+            l_g_total += l_g_gan
+            loss_dict['l_g_gan'] = l_g_gan
+
+            l_g_total.backward()
+            self.optimizer_g.step()
+
+        # optimize net_d
+        for p in self.net_d.parameters():
+            p.requires_grad = True
+
+        self.optimizer_d.zero_grad()
+        # real
+        real_d_pred = self.net_d(gan_gt)
+        l_d_real = self.cri_gan(real_d_pred, True, is_disc=True)
+        loss_dict['l_d_real'] = l_d_real
+        loss_dict['out_d_real'] = torch.mean(real_d_pred.detach())
+        l_d_real.backward()
+        # fake
+        fake_d_pred = self.net_d(self.output.detach().clone())  # clone for pt1.9
+        l_d_fake = self.cri_gan(fake_d_pred, False, is_disc=True)
+        loss_dict['l_d_fake'] = l_d_fake
+        loss_dict['out_d_fake'] = torch.mean(fake_d_pred.detach())
+        l_d_fake.backward()
+        self.optimizer_d.step()
+
+        if self.ema_decay > 0:
+            self.model_ema(decay=self.ema_decay)
+
+        self.log_dict = self.reduce_loss_dict(loss_dict)

realesrgan/models/realesrnet_model.py

@@ -0,0 +1,188 @@
+import numpy as np
+import random
+import torch
+from basicsr.data.degradations import random_add_gaussian_noise_pt, random_add_poisson_noise_pt
+from basicsr.data.transforms import paired_random_crop
+from basicsr.models.sr_model import SRModel
+from basicsr.utils import DiffJPEG, USMSharp
+from basicsr.utils.img_process_util import filter2D
+from basicsr.utils.registry import MODEL_REGISTRY
+from torch.nn import functional as F
+
+
+@MODEL_REGISTRY.register()
+class RealESRNetModel(SRModel):
+    """RealESRNet Model for Real-ESRGAN: Training Real-World Blind Super-Resolution with Pure Synthetic Data.
+
+    It is trained without GAN losses.
+    It mainly performs:
+    1. randomly synthesize LQ images in GPU tensors
+    2. optimize the networks with GAN training.
+    """
+
+    def __init__(self, opt):
+        super(RealESRNetModel, self).__init__(opt)
+        self.jpeger = DiffJPEG(differentiable=False).cuda()  # simulate JPEG compression artifacts
+        self.usm_sharpener = USMSharp().cuda()  # do usm sharpening
+        self.queue_size = opt.get('queue_size', 180)
+
+    @torch.no_grad()
+    def _dequeue_and_enqueue(self):
+        """It is the training pair pool for increasing the diversity in a batch.
+
+        Batch processing limits the diversity of synthetic degradations in a batch. For example, samples in a
+        batch could not have different resize scaling factors. Therefore, we employ this training pair pool
+        to increase the degradation diversity in a batch.
+        """
+        # initialize
+        b, c, h, w = self.lq.size()
+        if not hasattr(self, 'queue_lr'):
+            assert self.queue_size % b == 0, f'queue size {self.queue_size} should be divisible by batch size {b}'
+            self.queue_lr = torch.zeros(self.queue_size, c, h, w).cuda()
+            _, c, h, w = self.gt.size()
+            self.queue_gt = torch.zeros(self.queue_size, c, h, w).cuda()
+            self.queue_ptr = 0
+        if self.queue_ptr == self.queue_size:  # the pool is full
+            # do dequeue and enqueue
+            # shuffle
+            idx = torch.randperm(self.queue_size)
+            self.queue_lr = self.queue_lr[idx]
+            self.queue_gt = self.queue_gt[idx]
+            # get first b samples
+            lq_dequeue = self.queue_lr[0:b, :, :, :].clone()
+            gt_dequeue = self.queue_gt[0:b, :, :, :].clone()
+            # update the queue
+            self.queue_lr[0:b, :, :, :] = self.lq.clone()
+            self.queue_gt[0:b, :, :, :] = self.gt.clone()
+
+            self.lq = lq_dequeue
+            self.gt = gt_dequeue
+        else:
+            # only do enqueue
+            self.queue_lr[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.lq.clone()
+            self.queue_gt[self.queue_ptr:self.queue_ptr + b, :, :, :] = self.gt.clone()
+            self.queue_ptr = self.queue_ptr + b
+
+    @torch.no_grad()
+    def feed_data(self, data):
+        """Accept data from dataloader, and then add two-order degradations to obtain LQ images.
+        """
+        if self.is_train and self.opt.get('high_order_degradation', True):
+            # training data synthesis
+            self.gt = data['gt'].to(self.device)
+            # USM sharpen the GT images
+            if self.opt['gt_usm'] is True:
+                self.gt = self.usm_sharpener(self.gt)
+
+            self.kernel1 = data['kernel1'].to(self.device)
+            self.kernel2 = data['kernel2'].to(self.device)
+            self.sinc_kernel = data['sinc_kernel'].to(self.device)
+
+            ori_h, ori_w = self.gt.size()[2:4]
+
+            # ----------------------- The first degradation process ----------------------- #
+            # blur
+            out = filter2D(self.gt, self.kernel1)
+            # random resize
+            updown_type = random.choices(['up', 'down', 'keep'], self.opt['resize_prob'])[0]
+            if updown_type == 'up':
+                scale = np.random.uniform(1, self.opt['resize_range'][1])
+            elif updown_type == 'down':
+                scale = np.random.uniform(self.opt['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 = self.opt['gray_noise_prob']
+            if np.random.uniform() < self.opt['gaussian_noise_prob']:
+                out = random_add_gaussian_noise_pt(
+                    out, sigma_range=self.opt['noise_range'], clip=True, rounds=False, gray_prob=gray_noise_prob)
+            else:
+                out = random_add_poisson_noise_pt(
+                    out,
+                    scale_range=self.opt['poisson_scale_range'],
+                    gray_prob=gray_noise_prob,
+                    clip=True,
+                    rounds=False)
+            # JPEG compression
+            jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range'])
+            out = torch.clamp(out, 0, 1)  # clamp to [0, 1], otherwise JPEGer will result in unpleasant artifacts
+            out = self.jpeger(out, quality=jpeg_p)
+
+            # ----------------------- The second degradation process ----------------------- #
+            # blur
+            if np.random.uniform() < self.opt['second_blur_prob']:
+                out = filter2D(out, self.kernel2)
+            # random resize
+            updown_type = random.choices(['up', 'down', 'keep'], self.opt['resize_prob2'])[0]
+            if updown_type == 'up':
+                scale = np.random.uniform(1, self.opt['resize_range2'][1])
+            elif updown_type == 'down':
+                scale = np.random.uniform(self.opt['resize_range2'][0], 1)
+            else:
+                scale = 1
+            mode = random.choice(['area', 'bilinear', 'bicubic'])
+            out = F.interpolate(
+                out, size=(int(ori_h / self.opt['scale'] * scale), int(ori_w / self.opt['scale'] * scale)), mode=mode)
+            # add noise
+            gray_noise_prob = self.opt['gray_noise_prob2']
+            if np.random.uniform() < self.opt['gaussian_noise_prob2']:
+                out = random_add_gaussian_noise_pt(
+                    out, sigma_range=self.opt['noise_range2'], clip=True, rounds=False, gray_prob=gray_noise_prob)
+            else:
+                out = random_add_poisson_noise_pt(
+                    out,
+                    scale_range=self.opt['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 np.random.uniform() < 0.5:
+                # resize back + the final sinc filter
+                mode = random.choice(['area', 'bilinear', 'bicubic'])
+                out = F.interpolate(out, size=(ori_h // self.opt['scale'], ori_w // self.opt['scale']), mode=mode)
+                out = filter2D(out, self.sinc_kernel)
+                # JPEG compression
+                jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range2'])
+                out = torch.clamp(out, 0, 1)
+                out = self.jpeger(out, quality=jpeg_p)
+            else:
+                # JPEG compression
+                jpeg_p = out.new_zeros(out.size(0)).uniform_(*self.opt['jpeg_range2'])
+                out = torch.clamp(out, 0, 1)
+                out = self.jpeger(out, quality=jpeg_p)
+                # resize back + the final sinc filter
+                mode = random.choice(['area', 'bilinear', 'bicubic'])
+                out = F.interpolate(out, size=(ori_h // self.opt['scale'], ori_w // self.opt['scale']), mode=mode)
+                out = filter2D(out, self.sinc_kernel)
+
+            # clamp and round
+            self.lq = torch.clamp((out * 255.0).round(), 0, 255) / 255.
+
+            # random crop
+            gt_size = self.opt['gt_size']
+            self.gt, self.lq = paired_random_crop(self.gt, self.lq, gt_size, self.opt['scale'])
+
+            # training pair pool
+            self._dequeue_and_enqueue()
+            self.lq = self.lq.contiguous()  # for the warning: grad and param do not obey the gradient layout contract
+        else:
+            # for paired training or validation
+            self.lq = data['lq'].to(self.device)
+            if 'gt' in data:
+                self.gt = data['gt'].to(self.device)
+                self.gt_usm = self.usm_sharpener(self.gt)
+
+    def nondist_validation(self, dataloader, current_iter, tb_logger, save_img):
+        # do not use the synthetic process during validation
+        self.is_train = False
+        super(RealESRNetModel, self).nondist_validation(dataloader, current_iter, tb_logger, save_img)
+        self.is_train = True

realesrgan/train.py

@@ -0,0 +1,11 @@
+# flake8: noqa
+import os.path as osp
+from basicsr.train import train_pipeline
+
+import realesrgan.archs
+import realesrgan.data
+import realesrgan.models
+
+if __name__ == '__main__':
+    root_path = osp.abspath(osp.join(__file__, osp.pardir, osp.pardir))
+    train_pipeline(root_path)

realesrgan/utils.py

@@ -0,0 +1,313 @@
+import cv2
+import math
+import numpy as np
+import os
+import queue
+import threading
+import torch
+from basicsr.utils.download_util import load_file_from_url
+from torch.nn import functional as F
+
+ROOT_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+
+
+class RealESRGANer():
+    """A helper class for upsampling images with RealESRGAN.
+
+    Args:
+        scale (int): Upsampling scale factor used in the networks. It is usually 2 or 4.
+        model_path (str): The path to the pretrained model. It can be urls (will first download it automatically).
+        model (nn.Module): The defined network. Default: None.
+        tile (int): As too large images result in the out of GPU memory issue, so this tile option will first crop
+            input images into tiles, and then process each of them. Finally, they will be merged into one image.
+            0 denotes for do not use tile. Default: 0.
+        tile_pad (int): The pad size for each tile, to remove border artifacts. Default: 10.
+        pre_pad (int): Pad the input images to avoid border artifacts. Default: 10.
+        half (float): Whether to use half precision during inference. Default: False.
+    """
+
+    def __init__(self,
+                 scale,
+                 model_path,
+                 dni_weight=None,
+                 model=None,
+                 tile=0,
+                 tile_pad=10,
+                 pre_pad=10,
+                 half=False,
+                 device=None,
+                 gpu_id=None):
+        self.scale = scale
+        self.tile_size = tile
+        self.tile_pad = tile_pad
+        self.pre_pad = pre_pad
+        self.mod_scale = None
+        self.half = half
+
+        # initialize model
+        if gpu_id:
+            self.device = torch.device(
+                f'cuda:{gpu_id}' if torch.cuda.is_available() else 'cpu') if device is None else device
+        else:
+            self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') if device is None else device
+
+        if isinstance(model_path, list):
+            # dni
+            assert len(model_path) == len(dni_weight), 'model_path and dni_weight should have the save length.'
+            loadnet = self.dni(model_path[0], model_path[1], dni_weight)
+        else:
+            # if the model_path starts with https, it will first download models to the folder: weights
+            if model_path.startswith('https://'):
+                model_path = load_file_from_url(
+                    url=model_path, model_dir=os.path.join(ROOT_DIR, 'weights'), progress=True, file_name=None)
+            loadnet = torch.load(model_path, map_location=torch.device('cpu'))
+
+        # prefer to use params_ema
+        if 'params_ema' in loadnet:
+            keyname = 'params_ema'
+        else:
+            keyname = 'params'
+        model.load_state_dict(loadnet[keyname], strict=True)
+
+        model.eval()
+        self.model = model.to(self.device)
+        if self.half:
+            self.model = self.model.half()
+
+    def dni(self, net_a, net_b, dni_weight, key='params', loc='cpu'):
+        """Deep network interpolation.
+
+        ``Paper: Deep Network Interpolation for Continuous Imagery Effect Transition``
+        """
+        net_a = torch.load(net_a, map_location=torch.device(loc))
+        net_b = torch.load(net_b, map_location=torch.device(loc))
+        for k, v_a in net_a[key].items():
+            net_a[key][k] = dni_weight[0] * v_a + dni_weight[1] * net_b[key][k]
+        return net_a
+
+    def pre_process(self, img):
+        """Pre-process, such as pre-pad and mod pad, so that the images can be divisible
+        """
+        img = torch.from_numpy(np.transpose(img, (2, 0, 1))).float()
+        self.img = img.unsqueeze(0).to(self.device)
+        if self.half:
+            self.img = self.img.half()
+
+        # pre_pad
+        if self.pre_pad != 0:
+            self.img = F.pad(self.img, (0, self.pre_pad, 0, self.pre_pad), 'reflect')
+        # mod pad for divisible borders
+        if self.scale == 2:
+            self.mod_scale = 2
+        elif self.scale == 1:
+            self.mod_scale = 4
+        if self.mod_scale is not None:
+            self.mod_pad_h, self.mod_pad_w = 0, 0
+            _, _, h, w = self.img.size()
+            if (h % self.mod_scale != 0):
+                self.mod_pad_h = (self.mod_scale - h % self.mod_scale)
+            if (w % self.mod_scale != 0):
+                self.mod_pad_w = (self.mod_scale - w % self.mod_scale)
+            self.img = F.pad(self.img, (0, self.mod_pad_w, 0, self.mod_pad_h), 'reflect')
+
+    def process(self):
+        # model inference
+        self.output = self.model(self.img)
+
+    def tile_process(self):
+        """It will first crop input images to tiles, and then process each tile.
+        Finally, all the processed tiles are merged into one images.
+
+        Modified from: https://github.com/ata4/esrgan-launcher
+        """
+        batch, channel, height, width = self.img.shape
+        output_height = height * self.scale
+        output_width = width * self.scale
+        output_shape = (batch, channel, output_height, output_width)
+
+        # start with black image
+        self.output = self.img.new_zeros(output_shape)
+        tiles_x = math.ceil(width / self.tile_size)
+        tiles_y = math.ceil(height / self.tile_size)
+
+        # loop over all tiles
+        for y in range(tiles_y):
+            for x in range(tiles_x):
+                # extract tile from input image
+                ofs_x = x * self.tile_size
+                ofs_y = y * self.tile_size
+                # input tile area on total image
+                input_start_x = ofs_x
+                input_end_x = min(ofs_x + self.tile_size, width)
+                input_start_y = ofs_y
+                input_end_y = min(ofs_y + self.tile_size, height)
+
+                # input tile area on total image with padding
+                input_start_x_pad = max(input_start_x - self.tile_pad, 0)
+                input_end_x_pad = min(input_end_x + self.tile_pad, width)
+                input_start_y_pad = max(input_start_y - self.tile_pad, 0)
+                input_end_y_pad = min(input_end_y + self.tile_pad, height)
+
+                # input tile dimensions
+                input_tile_width = input_end_x - input_start_x
+                input_tile_height = input_end_y - input_start_y
+                tile_idx = y * tiles_x + x + 1
+                input_tile = self.img[:, :, input_start_y_pad:input_end_y_pad, input_start_x_pad:input_end_x_pad]
+
+                # upscale tile
+                try:
+                    with torch.no_grad():
+                        output_tile = self.model(input_tile)
+                except RuntimeError as error:
+                    print('Error', error)
+                print(f'\tTile {tile_idx}/{tiles_x * tiles_y}')
+
+                # output tile area on total image
+                output_start_x = input_start_x * self.scale
+                output_end_x = input_end_x * self.scale
+                output_start_y = input_start_y * self.scale
+                output_end_y = input_end_y * self.scale
+
+                # output tile area without padding
+                output_start_x_tile = (input_start_x - input_start_x_pad) * self.scale
+                output_end_x_tile = output_start_x_tile + input_tile_width * self.scale
+                output_start_y_tile = (input_start_y - input_start_y_pad) * self.scale
+                output_end_y_tile = output_start_y_tile + input_tile_height * self.scale
+
+                # put tile into output image
+                self.output[:, :, output_start_y:output_end_y,
+                            output_start_x:output_end_x] = output_tile[:, :, output_start_y_tile:output_end_y_tile,
+                                                                       output_start_x_tile:output_end_x_tile]
+
+    def post_process(self):
+        # remove extra pad
+        if self.mod_scale is not None:
+            _, _, h, w = self.output.size()
+            self.output = self.output[:, :, 0:h - self.mod_pad_h * self.scale, 0:w - self.mod_pad_w * self.scale]
+        # remove prepad
+        if self.pre_pad != 0:
+            _, _, h, w = self.output.size()
+            self.output = self.output[:, :, 0:h - self.pre_pad * self.scale, 0:w - self.pre_pad * self.scale]
+        return self.output
+
+    @torch.no_grad()
+    def enhance(self, img, outscale=None, alpha_upsampler='realesrgan'):
+        h_input, w_input = img.shape[0:2]
+        # img: numpy
+        img = img.astype(np.float32)
+        if np.max(img) > 256:  # 16-bit image
+            max_range = 65535
+            print('\tInput is a 16-bit image')
+        else:
+            max_range = 255
+        img = img / max_range
+        if len(img.shape) == 2:  # gray image
+            img_mode = 'L'
+            img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB)
+        elif img.shape[2] == 4:  # RGBA image with alpha channel
+            img_mode = 'RGBA'
+            alpha = img[:, :, 3]
+            img = img[:, :, 0:3]
+            img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+            if alpha_upsampler == 'realesrgan':
+                alpha = cv2.cvtColor(alpha, cv2.COLOR_GRAY2RGB)
+        else:
+            img_mode = 'RGB'
+            img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+
+        # ------------------- process image (without the alpha channel) ------------------- #
+        self.pre_process(img)
+        if self.tile_size > 0:
+            self.tile_process()
+        else:
+            self.process()
+        output_img = self.post_process()
+        output_img = output_img.data.squeeze().float().cpu().clamp_(0, 1).numpy()
+        output_img = np.transpose(output_img[[2, 1, 0], :, :], (1, 2, 0))
+        if img_mode == 'L':
+            output_img = cv2.cvtColor(output_img, cv2.COLOR_BGR2GRAY)
+
+        # ------------------- process the alpha channel if necessary ------------------- #
+        if img_mode == 'RGBA':
+            if alpha_upsampler == 'realesrgan':
+                self.pre_process(alpha)
+                if self.tile_size > 0:
+                    self.tile_process()
+                else:
+                    self.process()
+                output_alpha = self.post_process()
+                output_alpha = output_alpha.data.squeeze().float().cpu().clamp_(0, 1).numpy()
+                output_alpha = np.transpose(output_alpha[[2, 1, 0], :, :], (1, 2, 0))
+                output_alpha = cv2.cvtColor(output_alpha, cv2.COLOR_BGR2GRAY)
+            else:  # use the cv2 resize for alpha channel
+                h, w = alpha.shape[0:2]
+                output_alpha = cv2.resize(alpha, (w * self.scale, h * self.scale), interpolation=cv2.INTER_LINEAR)
+
+            # merge the alpha channel
+            output_img = cv2.cvtColor(output_img, cv2.COLOR_BGR2BGRA)
+            output_img[:, :, 3] = output_alpha
+
+        # ------------------------------ return ------------------------------ #
+        if max_range == 65535:  # 16-bit image
+            output = (output_img * 65535.0).round().astype(np.uint16)
+        else:
+            output = (output_img * 255.0).round().astype(np.uint8)
+
+        if outscale is not None and outscale != float(self.scale):
+            output = cv2.resize(
+                output, (
+                    int(w_input * outscale),
+                    int(h_input * outscale),
+                ), interpolation=cv2.INTER_LANCZOS4)
+
+        return output, img_mode
+
+
+class PrefetchReader(threading.Thread):
+    """Prefetch images.
+
+    Args:
+        img_list (list[str]): A image list of image paths to be read.
+        num_prefetch_queue (int): Number of prefetch queue.
+    """
+
+    def __init__(self, img_list, num_prefetch_queue):
+        super().__init__()
+        self.que = queue.Queue(num_prefetch_queue)
+        self.img_list = img_list
+
+    def run(self):
+        for img_path in self.img_list:
+            img = cv2.imread(img_path, cv2.IMREAD_UNCHANGED)
+            self.que.put(img)
+
+        self.que.put(None)
+
+    def __next__(self):
+        next_item = self.que.get()
+        if next_item is None:
+            raise StopIteration
+        return next_item
+
+    def __iter__(self):
+        return self
+
+
+class IOConsumer(threading.Thread):
+
+    def __init__(self, opt, que, qid):
+        super().__init__()
+        self._queue = que
+        self.qid = qid
+        self.opt = opt
+
+    def run(self):
+        while True:
+            msg = self._queue.get()
+            if isinstance(msg, str) and msg == 'quit':
+                break
+
+            output = msg['output']
+            save_path = msg['save_path']
+            cv2.imwrite(save_path, output)
+        print(f'IO worker {self.qid} is done.')

realesrgan/version.py

@@ -0,0 +1,5 @@
+# GENERATED VERSION FILE
+# TIME: Thu Oct 12 15:29:54 2023
+__version__ = '0.3.0'
+__gitsha__ = '5ca1078'
+version_info = (0, 3, 0)