ridcp

basicsr/archs/vgg_arch.py

+import os
+import torch
+from collections import OrderedDict
+from torch import nn as nn
+from torchvision.models import vgg as vgg
+
+from basicsr.utils.registry import ARCH_REGISTRY
+
+VGG_PRETRAIN_PATH = 'experiments/pretrained_models/vgg19-dcbb9e9d.pth'
+NAMES = {
+    'vgg11': [
+        'conv1_1', 'relu1_1', 'pool1', 'conv2_1', 'relu2_1', 'pool2', 'conv3_1', 'relu3_1', 'conv3_2', 'relu3_2',
+        'pool3', 'conv4_1', 'relu4_1', 'conv4_2', 'relu4_2', 'pool4', 'conv5_1', 'relu5_1', 'conv5_2', 'relu5_2',
+        'pool5'
+    ],
+    'vgg13': [
+        'conv1_1', 'relu1_1', 'conv1_2', 'relu1_2', 'pool1', 'conv2_1', 'relu2_1', 'conv2_2', 'relu2_2', 'pool2',
+        'conv3_1', 'relu3_1', 'conv3_2', 'relu3_2', 'pool3', 'conv4_1', 'relu4_1', 'conv4_2', 'relu4_2', 'pool4',
+        'conv5_1', 'relu5_1', 'conv5_2', 'relu5_2', 'pool5'
+    ],
+    'vgg16': [
+        'conv1_1', 'relu1_1', 'conv1_2', 'relu1_2', 'pool1', 'conv2_1', 'relu2_1', 'conv2_2', 'relu2_2', 'pool2',
+        'conv3_1', 'relu3_1', 'conv3_2', 'relu3_2', 'conv3_3', 'relu3_3', 'pool3', 'conv4_1', 'relu4_1', 'conv4_2',
+        'relu4_2', 'conv4_3', 'relu4_3', 'pool4', 'conv5_1', 'relu5_1', 'conv5_2', 'relu5_2', 'conv5_3', 'relu5_3',
+        'pool5'
+    ],
+    'vgg19': [
+        'conv1_1', 'relu1_1', 'conv1_2', 'relu1_2', 'pool1', 'conv2_1', 'relu2_1', 'conv2_2', 'relu2_2', 'pool2',
+        'conv3_1', 'relu3_1', 'conv3_2', 'relu3_2', 'conv3_3', 'relu3_3', 'conv3_4', 'relu3_4', 'pool3', 'conv4_1',
+        'relu4_1', 'conv4_2', 'relu4_2', 'conv4_3', 'relu4_3', 'conv4_4', 'relu4_4', 'pool4', 'conv5_1', 'relu5_1',
+        'conv5_2', 'relu5_2', 'conv5_3', 'relu5_3', 'conv5_4', 'relu5_4', 'pool5'
+    ]
+}
+
+
+def insert_bn(names):
+    """Insert bn layer after each conv.
+
+    Args:
+        names (list): The list of layer names.
+
+    Returns:
+        list: The list of layer names with bn layers.
+    """
+    names_bn = []
+    for name in names:
+        names_bn.append(name)
+        if 'conv' in name:
+            position = name.replace('conv', '')
+            names_bn.append('bn' + position)
+    return names_bn
+
+
+@ARCH_REGISTRY.register()
+class VGGFeatureExtractor(nn.Module):
+    """VGG network for feature extraction.
+
+    In this implementation, we allow users to choose whether use normalization
+    in the input feature and the type of vgg network. Note that the pretrained
+    path must fit the vgg type.
+
+    Args:
+        layer_name_list (list[str]): Forward function returns the corresponding
+            features according to the layer_name_list.
+            Example: {'relu1_1', 'relu2_1', 'relu3_1'}.
+        vgg_type (str): Set the type of vgg network. Default: 'vgg19'.
+        use_input_norm (bool): If True, normalize the input image. Importantly,
+            the input feature must in the range [0, 1]. Default: True.
+        range_norm (bool): If True, norm images with range [-1, 1] to [0, 1].
+            Default: False.
+        requires_grad (bool): If true, the parameters of VGG network will be
+            optimized. Default: False.
+        remove_pooling (bool): If true, the max pooling operations in VGG net
+            will be removed. Default: False.
+        pooling_stride (int): The stride of max pooling operation. Default: 2.
+    """
+
+    def __init__(self,
+                 layer_name_list,
+                 vgg_type='vgg19',
+                 use_input_norm=True,
+                 range_norm=False,
+                 requires_grad=False,
+                 remove_pooling=False,
+                 pooling_stride=2):
+        super(VGGFeatureExtractor, self).__init__()
+
+        self.layer_name_list = layer_name_list
+        self.use_input_norm = use_input_norm
+        self.range_norm = range_norm
+
+        self.names = NAMES[vgg_type.replace('_bn', '')]
+        if 'bn' in vgg_type:
+            self.names = insert_bn(self.names)
+
+        # only borrow layers that will be used to avoid unused params
+        max_idx = 0
+        for v in layer_name_list:
+            idx = self.names.index(v)
+            if idx > max_idx:
+                max_idx = idx
+
+        if os.path.exists(VGG_PRETRAIN_PATH):
+            vgg_net = getattr(vgg, vgg_type)(pretrained=False)
+            state_dict = torch.load(VGG_PRETRAIN_PATH, map_location=lambda storage, loc: storage)
+            vgg_net.load_state_dict(state_dict)
+        else:
+            vgg_net = getattr(vgg, vgg_type)(pretrained=True)
+
+        features = vgg_net.features[:max_idx + 1]
+
+        modified_net = OrderedDict()
+        for k, v in zip(self.names, features):
+            if 'pool' in k:
+                # if remove_pooling is true, pooling operation will be removed
+                if remove_pooling:
+                    continue
+                else:
+                    # in some cases, we may want to change the default stride
+                    modified_net[k] = nn.MaxPool2d(kernel_size=2, stride=pooling_stride)
+            else:
+                modified_net[k] = v
+
+        self.vgg_net = nn.Sequential(modified_net)
+
+        if not requires_grad:
+            self.vgg_net.eval()
+            for param in self.parameters():
+                param.requires_grad = False
+        else:
+            self.vgg_net.train()
+            for param in self.parameters():
+                param.requires_grad = True
+
+        if self.use_input_norm:
+            # the mean is for image with range [0, 1]
+            self.register_buffer('mean', torch.Tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1))
+            # the std is for image with range [0, 1]
+            self.register_buffer('std', torch.Tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1))
+
+    def forward(self, x):
+        """Forward function.
+
+        Args:
+            x (Tensor): Input tensor with shape (n, c, h, w).
+
+        Returns:
+            Tensor: Forward results.
+        """
+        if self.range_norm:
+            x = (x + 1) / 2
+        if self.use_input_norm:
+            x = (x - self.mean) / self.std
+
+        output = {}
+        for key, layer in self.vgg_net._modules.items():
+            x = layer(x)
+            if key in self.layer_name_list:
+                output[key] = x.clone()
+
+        return output

basicsr/data/__init__.py

+import importlib
+import numpy as np
+import random
+import torch
+import torch.utils.data
+from copy import deepcopy
+from functools import partial
+from os import path as osp
+
+from basicsr.data.prefetch_dataloader import PrefetchDataLoader
+from basicsr.utils import get_root_logger, scandir
+from basicsr.utils.dist_util import get_dist_info
+from basicsr.utils.registry import DATASET_REGISTRY
+
+__all__ = ['build_dataset', 'build_dataloader']
+
+# automatically scan and import dataset modules for registry
+# scan all the files under the data folder with '_dataset' in file names
+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'basicsr.data.{file_name}') for file_name in dataset_filenames]
+
+
+def build_dataset(dataset_opt):
+    """Build dataset from options.
+
+    Args:
+        dataset_opt (dict): Configuration for dataset. It must contain:
+            name (str): Dataset name.
+            type (str): Dataset type.
+    """
+    dataset_opt = deepcopy(dataset_opt)
+    dataset = DATASET_REGISTRY.get(dataset_opt['type'])(dataset_opt)
+    logger = get_root_logger()
+    logger.info(f'Dataset [{dataset.__class__.__name__}] - {dataset_opt["name"]} ' 'is built.')
+    return dataset
+
+
+def build_dataloader(dataset, dataset_opt, num_gpu=1, dist=False, sampler=None, seed=None):
+    """Build dataloader.
+
+    Args:
+        dataset (torch.utils.data.Dataset): Dataset.
+        dataset_opt (dict): Dataset options. It contains the following keys:
+            phase (str): 'train' or 'val'.
+            num_worker_per_gpu (int): Number of workers for each GPU.
+            batch_size_per_gpu (int): Training batch size for each GPU.
+        num_gpu (int): Number of GPUs. Used only in the train phase.
+            Default: 1.
+        dist (bool): Whether in distributed training. Used only in the train
+            phase. Default: False.
+        sampler (torch.utils.data.sampler): Data sampler. Default: None.
+        seed (int | None): Seed. Default: None
+    """
+    phase = dataset_opt['phase']
+    rank, _ = get_dist_info()
+    if phase == 'train':
+        if dist:  # distributed training
+            batch_size = dataset_opt['batch_size_per_gpu']
+            num_workers = dataset_opt['num_worker_per_gpu']
+        else:  # non-distributed training
+            multiplier = 1 if num_gpu == 0 else num_gpu
+            batch_size = dataset_opt['batch_size_per_gpu'] * multiplier
+            num_workers = dataset_opt['num_worker_per_gpu'] * multiplier
+        dataloader_args = dict(
+            dataset=dataset,
+            batch_size=batch_size,
+            shuffle=False,
+            num_workers=num_workers,
+            sampler=sampler,
+            drop_last=True)
+        if sampler is None:
+            dataloader_args['shuffle'] = True
+        dataloader_args['worker_init_fn'] = partial(
+            worker_init_fn, num_workers=num_workers, rank=rank, seed=seed) if seed is not None else None
+    elif phase in ['val', 'test']:  # validation
+        dataloader_args = dict(dataset=dataset, batch_size=1, shuffle=False, num_workers=0)
+    else:
+        raise ValueError(f'Wrong dataset phase: {phase}. ' "Supported ones are 'train', 'val' and 'test'.")
+
+    dataloader_args['pin_memory'] = dataset_opt.get('pin_memory', False)
+    dataloader_args['persistent_workers'] = dataset_opt.get('persistent_workers', False)
+
+    prefetch_mode = dataset_opt.get('prefetch_mode')
+    if prefetch_mode == 'cpu':  # CPUPrefetcher
+        num_prefetch_queue = dataset_opt.get('num_prefetch_queue', 1)
+        logger = get_root_logger()
+        logger.info(f'Use {prefetch_mode} prefetch dataloader: num_prefetch_queue = {num_prefetch_queue}')
+        return PrefetchDataLoader(num_prefetch_queue=num_prefetch_queue, **dataloader_args)
+    else:
+        # prefetch_mode=None: Normal dataloader
+        # prefetch_mode='cuda': dataloader for CUDAPrefetcher
+        return torch.utils.data.DataLoader(**dataloader_args)
+
+
+def worker_init_fn(worker_id, num_workers, rank, seed):
+    # Set the worker seed to num_workers * rank + worker_id + seed
+    worker_seed = num_workers * rank + worker_id + seed
+    np.random.seed(worker_seed)
+    random.seed(worker_seed)

basicsr/data/data_sampler.py

+import math
+import torch
+from torch.utils.data.sampler import Sampler
+
+
+class EnlargedSampler(Sampler):
+    """Sampler that restricts data loading to a subset of the dataset.
+
+    Modified from torch.utils.data.distributed.DistributedSampler
+    Support enlarging the dataset for iteration-based training, for saving
+    time when restart the dataloader after each epoch
+
+    Args:
+        dataset (torch.utils.data.Dataset): Dataset used for sampling.
+        num_replicas (int | None): Number of processes participating in
+            the training. It is usually the world_size.
+        rank (int | None): Rank of the current process within num_replicas.
+        ratio (int): Enlarging ratio. Default: 1.
+    """
+
+    def __init__(self, dataset, num_replicas, rank, ratio=1):
+        self.dataset = dataset
+        self.num_replicas = num_replicas
+        self.rank = rank
+        self.epoch = 0
+        self.num_samples = math.ceil(len(self.dataset) * ratio / self.num_replicas)
+        self.total_size = self.num_samples * self.num_replicas
+
+    def __iter__(self):
+        # deterministically shuffle based on epoch
+        g = torch.Generator()
+        g.manual_seed(self.epoch)
+        indices = torch.randperm(self.total_size, generator=g).tolist()
+
+        dataset_size = len(self.dataset)
+        indices = [v % dataset_size for v in indices]
+
+        # subsample
+        indices = indices[self.rank:self.total_size:self.num_replicas]
+        assert len(indices) == self.num_samples
+
+        return iter(indices)
+
+    def __len__(self):
+        return self.num_samples
+
+    def set_epoch(self, epoch):
+        self.epoch = epoch

basicsr/data/data_util.py

+import cv2
+import numpy as np
+import torch
+import os
+from os import path as osp
+from torch.nn import functional as F
+
+from basicsr.data.transforms import mod_crop
+from basicsr.utils import img2tensor, scandir
+
+
+IMG_EXTENSIONS = [
+    '.jpg', '.JPG', '.jpeg', '.JPEG',
+    '.png', '.PNG', '.ppm', '.PPM', '.bmp', '.BMP',
+    '.tif', '.TIF', '.tiff', '.TIFF', '.npy'
+]
+
+
+def is_image_file(filename):
+    return any(filename.endswith(extension) for extension in IMG_EXTENSIONS)
+
+
+def make_dataset(dir, max_dataset_size=float("inf"), followlinks=True):
+    images = []
+    assert os.path.isdir(dir), '%s is not a valid directory' % dir
+
+    for root, _, fnames in sorted(os.walk(dir, followlinks=followlinks)):
+        for fname in fnames:
+            if is_image_file(fname):
+                path = os.path.join(root, fname)
+                images.append(path)
+    return images[:min(max_dataset_size, len(images))]
+
+
+def read_img_seq(path, require_mod_crop=False, scale=1, return_imgname=False):
+    """Read a sequence of images from a given folder path.
+
+    Args:
+        path (list[str] | str): List of image paths or image folder path.
+        require_mod_crop (bool): Require mod crop for each image.
+            Default: False.
+        scale (int): Scale factor for mod_crop. Default: 1.
+        return_imgname(bool): Whether return image names. Default False.
+
+    Returns:
+        Tensor: size (t, c, h, w), RGB, [0, 1].
+        list[str]: Returned image name list.
+    """
+    if isinstance(path, list):
+        img_paths = path
+    else:
+        img_paths = sorted(list(scandir(path, full_path=True)))
+    imgs = [cv2.imread(v).astype(np.float32) / 255. for v in img_paths]
+
+    if require_mod_crop:
+        imgs = [mod_crop(img, scale) for img in imgs]
+    imgs = img2tensor(imgs, bgr2rgb=True, float32=True)
+    imgs = torch.stack(imgs, dim=0)
+
+    if return_imgname:
+        imgnames = [osp.splitext(osp.basename(path))[0] for path in img_paths]
+        return imgs, imgnames
+    else:
+        return imgs
+
+
+def generate_frame_indices(crt_idx, max_frame_num, num_frames, padding='reflection'):
+    """Generate an index list for reading `num_frames` frames from a sequence
+    of images.
+
+    Args:
+        crt_idx (int): Current center index.
+        max_frame_num (int): Max number of the sequence of images (from 1).
+        num_frames (int): Reading num_frames frames.
+        padding (str): Padding mode, one of
+            'replicate' | 'reflection' | 'reflection_circle' | 'circle'
+            Examples: current_idx = 0, num_frames = 5
+            The generated frame indices under different padding mode:
+            replicate: [0, 0, 0, 1, 2]
+            reflection: [2, 1, 0, 1, 2]
+            reflection_circle: [4, 3, 0, 1, 2]
+            circle: [3, 4, 0, 1, 2]
+
+    Returns:
+        list[int]: A list of indices.
+    """
+    assert num_frames % 2 == 1, 'num_frames should be an odd number.'
+    assert padding in ('replicate', 'reflection', 'reflection_circle', 'circle'), f'Wrong padding mode: {padding}.'
+
+    max_frame_num = max_frame_num - 1  # start from 0
+    num_pad = num_frames // 2
+
+    indices = []
+    for i in range(crt_idx - num_pad, crt_idx + num_pad + 1):
+        if i < 0:
+            if padding == 'replicate':
+                pad_idx = 0
+            elif padding == 'reflection':
+                pad_idx = -i
+            elif padding == 'reflection_circle':
+                pad_idx = crt_idx + num_pad - i
+            else:
+                pad_idx = num_frames + i
+        elif i > max_frame_num:
+            if padding == 'replicate':
+                pad_idx = max_frame_num
+            elif padding == 'reflection':
+                pad_idx = max_frame_num * 2 - i
+            elif padding == 'reflection_circle':
+                pad_idx = (crt_idx - num_pad) - (i - max_frame_num)
+            else:
+                pad_idx = i - num_frames
+        else:
+            pad_idx = i
+        indices.append(pad_idx)
+    return indices
+
+
+def paired_paths_from_lmdb(folders, keys):
+    """Generate paired paths from lmdb files.
+
+    Contents of lmdb. Taking the `lq.lmdb` for example, the file structure is:
+
+    lq.lmdb
+    ├── data.mdb
+    ├── lock.mdb
+    ├── meta_info.txt
+
+    The data.mdb and lock.mdb are standard lmdb files and you can refer to
+    https://lmdb.readthedocs.io/en/release/ for more details.
+
+    The meta_info.txt is a specified txt file to record the meta information
+    of our datasets. It will be automatically created when preparing
+    datasets by our provided dataset tools.
+    Each line in the txt file records
+    1)image name (with extension),
+    2)image shape,
+    3)compression level, separated by a white space.
+    Example: `baboon.png (120,125,3) 1`
+
+    We use the image name without extension as the lmdb key.
+    Note that we use the same key for the corresponding lq and gt images.
+
+    Args:
+        folders (list[str]): A list of folder path. The order of list should
+            be [input_folder, gt_folder].
+        keys (list[str]): A list of keys identifying folders. The order should
+            be in consistent with folders, e.g., ['lq', 'gt'].
+            Note that this key is different from lmdb keys.
+
+    Returns:
+        list[str]: Returned path list.
+    """
+    assert len(folders) == 2, ('The len of folders should be 2 with [input_folder, gt_folder]. '
+                               f'But got {len(folders)}')
+    assert len(keys) == 2, f'The len of keys should be 2 with [input_key, gt_key]. But got {len(keys)}'
+    input_folder, gt_folder = folders
+    input_key, gt_key = keys
+
+    if not (input_folder.endswith('.lmdb') and gt_folder.endswith('.lmdb')):
+        raise ValueError(f'{input_key} folder and {gt_key} folder should both in lmdb '
+                         f'formats. But received {input_key}: {input_folder}; '
+                         f'{gt_key}: {gt_folder}')
+    # ensure that the two meta_info files are the same
+    with open(osp.join(input_folder, 'meta_info.txt')) as fin:
+        input_lmdb_keys = [line.split('.')[0] for line in fin]
+    with open(osp.join(gt_folder, 'meta_info.txt')) as fin:
+        gt_lmdb_keys = [line.split('.')[0] for line in fin]
+    if set(input_lmdb_keys) != set(gt_lmdb_keys):
+        raise ValueError(f'Keys in {input_key}_folder and {gt_key}_folder are different.')
+    else:
+        paths = []
+        for lmdb_key in sorted(input_lmdb_keys):
+            paths.append(dict([(f'{input_key}_path', lmdb_key), (f'{gt_key}_path', lmdb_key)]))
+        return paths
+
+
+def paired_paths_from_meta_info_file(folders, keys, meta_info_file, filename_tmpl):
+    """Generate paired paths from an meta information file.
+
+    Each line in the meta information file contains the image names and
+    image shape (usually for gt), separated by a white space.
+
+    Example of an meta information file:
+    ```
+    0001_s001.png (480,480,3)
+    0001_s002.png (480,480,3)
+    ```
+
+    Args:
+        folders (list[str]): A list of folder path. The order of list should
+            be [input_folder, gt_folder].
+        keys (list[str]): A list of keys identifying folders. The order should
+            be in consistent with folders, e.g., ['lq', 'gt'].
+        meta_info_file (str): Path to the meta information file.
+        filename_tmpl (str): Template for each filename. Note that the
+            template excludes the file extension. Usually the filename_tmpl is
+            for files in the input folder.
+
+    Returns:
+        list[str]: Returned path list.
+    """
+    assert len(folders) == 2, ('The len of folders should be 2 with [input_folder, gt_folder]. '
+                               f'But got {len(folders)}')
+    assert len(keys) == 2, f'The len of keys should be 2 with [input_key, gt_key]. But got {len(keys)}'
+    input_folder, gt_folder = folders
+    input_key, gt_key = keys
+
+    with open(meta_info_file, 'r') as fin:
+        gt_names = [line.strip().split(' ')[0] for line in fin]
+
+    paths = []
+    for gt_name in gt_names:
+        basename, ext = osp.splitext(osp.basename(gt_name))
+        input_name = f'{filename_tmpl.format(basename)}{ext}'
+        input_path = osp.join(input_folder, input_name)
+        gt_path = osp.join(gt_folder, gt_name)
+        paths.append(dict([(f'{input_key}_path', input_path), (f'{gt_key}_path', gt_path)]))
+    return paths
+
+
+def paired_paths_from_folder(folders, keys, filename_tmpl):
+    """Generate paired paths from folders.
+
+    Args:
+        folders (list[str]): A list of folder path. The order of list should
+            be [input_folder, gt_folder].
+        keys (list[str]): A list of keys identifying folders. The order should
+            be in consistent with folders, e.g., ['lq', 'gt'].
+        filename_tmpl (str): Template for each filename. Note that the
+            template excludes the file extension. Usually the filename_tmpl is
+            for files in the input folder.
+
+    Returns:
+        list[str]: Returned path list.
+    """
+    assert len(folders) == 2, ('The len of folders should be 2 with [input_folder, gt_folder]. '
+                               f'But got {len(folders)}')
+    assert len(keys) == 2, f'The len of keys should be 2 with [input_key, gt_key]. But got {len(keys)}'
+    input_folder, gt_folder = folders
+    input_key, gt_key = keys
+
+    input_paths = list(scandir(input_folder))
+    gt_paths = list(scandir(gt_folder))
+    assert len(input_paths) == len(gt_paths), (f'{input_key} and {gt_key} datasets have different number of images: '
+                                               f'{len(input_paths)}, {len(gt_paths)}.')
+    paths = []
+    for gt_path in gt_paths:
+        basename, ext = osp.splitext(osp.basename(gt_path))
+        input_name = f'{filename_tmpl.format(basename)}{ext}'
+        input_path = osp.join(input_folder, input_name)
+        assert input_name in input_paths, f'{input_name} is not in {input_key}_paths.'
+        gt_path = osp.join(gt_folder, gt_path)
+        paths.append(dict([(f'{input_key}_path', input_path), (f'{gt_key}_path', gt_path)]))
+    return paths
+
+
+def paths_from_folder(folder):
+    """Generate paths from folder.
+
+    Args:
+        folder (str): Folder path.
+
+    Returns:
+        list[str]: Returned path list.
+    """
+
+    paths = list(scandir(folder))
+    paths = [osp.join(folder, path) for path in paths]
+    return paths
+
+
+def paths_from_lmdb(folder):
+    """Generate paths from lmdb.
+
+    Args:
+        folder (str): Folder path.
+
+    Returns:
+        list[str]: Returned path list.
+    """
+    if not folder.endswith('.lmdb'):
+        raise ValueError(f'Folder {folder}folder should in lmdb format.')
+    with open(osp.join(folder, 'meta_info.txt')) as fin:
+        paths = [line.split('.')[0] for line in fin]
+    return paths
+
+
+def generate_gaussian_kernel(kernel_size=13, sigma=1.6):
+    """Generate Gaussian kernel used in `duf_downsample`.
+
+    Args:
+        kernel_size (int): Kernel size. Default: 13.
+        sigma (float): Sigma of the Gaussian kernel. Default: 1.6.
+
+    Returns:
+        np.array: The Gaussian kernel.
+    """
+    from scipy.ndimage import filters as filters
+    kernel = np.zeros((kernel_size, kernel_size))
+    # set element at the middle to one, a dirac delta
+    kernel[kernel_size // 2, kernel_size // 2] = 1
+    # gaussian-smooth the dirac, resulting in a gaussian filter
+    return filters.gaussian_filter(kernel, sigma)
+
+
+def duf_downsample(x, kernel_size=13, scale=4):
+    """Downsamping with Gaussian kernel used in the DUF official code.
+
+    Args:
+        x (Tensor): Frames to be downsampled, with shape (b, t, c, h, w).
+        kernel_size (int): Kernel size. Default: 13.
+        scale (int): Downsampling factor. Supported scale: (2, 3, 4).
+            Default: 4.
+
+    Returns:
+        Tensor: DUF downsampled frames.
+    """
+    assert scale in (2, 3, 4), f'Only support scale (2, 3, 4), but got {scale}.'
+
+    squeeze_flag = False
+    if x.ndim == 4:
+        squeeze_flag = True
+        x = x.unsqueeze(0)
+    b, t, c, h, w = x.size()
+    x = x.view(-1, 1, h, w)
+    pad_w, pad_h = kernel_size // 2 + scale * 2, kernel_size // 2 + scale * 2
+    x = F.pad(x, (pad_w, pad_w, pad_h, pad_h), 'reflect')
+
+    gaussian_filter = generate_gaussian_kernel(kernel_size, 0.4 * scale)
+    gaussian_filter = torch.from_numpy(gaussian_filter).type_as(x).unsqueeze(0).unsqueeze(0)
+    x = F.conv2d(x, gaussian_filter, stride=scale)
+    x = x[:, :, 2:-2, 2:-2]
+    x = x.view(b, t, c, x.size(2), x.size(3))
+    if squeeze_flag:
+        x = x.squeeze(0)
+    return x

basicsr/data/haze_online_dataset.py

+import os
+import cv2
+import random
+import numpy as np
+from torch.utils import data as data
+from scipy import ndimage
+import scipy
+import scipy.stats as ss
+from scipy.interpolate import interp2d
+from scipy.linalg import orth
+
+from basicsr.data.transforms import augment, paired_random_crop
+from basicsr.utils import FileClient, img2tensor
+from basicsr.utils.registry import DATASET_REGISTRY
+
+from .data_util import make_dataset
+
+def uint2single(img):
+    return np.float32(img/255.)
+
+def single2uint(img):
+    return np.uint8((img.clip(0, 1)*255.).round())
+
+def random_resize(img, scale_factor=1.):
+    return cv2.resize(img, None, fx=scale_factor, fy=scale_factor, interpolation=cv2.INTER_CUBIC)
+
+def add_Gaussian_noise(img, noise_level1=2, noise_level2=25):
+    noise_level = random.randint(noise_level1, noise_level2)
+    rnum = np.random.rand()
+    if rnum > 0.6:   # add color Gaussian noise
+        img += np.random.normal(0, noise_level/255.0, img.shape).astype(np.float32)
+    elif rnum < 0.4: # add grayscale Gaussian noise
+        img += np.random.normal(0, noise_level/255.0, (*img.shape[:2], 1)).astype(np.float32)
+    else:            # add  noise
+        L = noise_level2/255.
+        D = np.diag(np.random.rand(3))
+        U = orth(np.random.rand(3,3))
+        conv = np.dot(np.dot(np.transpose(U), D), U)
+        img += np.random.multivariate_normal([0,0,0], np.abs(L**2*conv), img.shape[:2]).astype(np.float32)
+    img = np.clip(img, 0.0, 1.0)
+    return img
+
+def add_JPEG_noise(img):
+    quality_factor = random.randint(30, 95)
+    img = cv2.cvtColor(single2uint(img), cv2.COLOR_RGB2BGR)
+    result, encimg = cv2.imencode('.jpg', img, [int(cv2.IMWRITE_JPEG_QUALITY), quality_factor])
+    img = cv2.imdecode(encimg, 1)
+    img = cv2.cvtColor(uint2single(img), cv2.COLOR_BGR2RGB)
+    return img
+
+@DATASET_REGISTRY.register()
+class HazeOnlineDataset(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_file': Use meta information file to generate paths.
+        If opt['io_backend'] != lmdb and opt['meta_info_file'] 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_file (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_flip (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(HazeOnlineDataset, self).__init__()
+        self.opt = opt
+        # file client (io backend)
+        self.file_client = None
+        self.io_backend_opt = opt['io_backend']
+
+        self.gt_folder = opt['dataroot_gt']
+        self.depth_folder = opt['dataroot_depth']
+        self.gt_paths = make_dataset(self.gt_folder)
+        self.depth_paths = make_dataset(self.depth_folder)
+        self.beta_range = opt['beta_range']
+        self.A_range = opt['A_range']
+        self.color_p = opt['color_p']
+        self.color_range = opt['color_range']
+
+    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.gt_paths[index]
+        img_gt = cv2.imread(gt_path).astype(np.float32) / 255.0
+
+        depth_path = os.path.join(self.depth_folder, gt_path.split('/')[-1].split('.')[0] + '.npy')
+        img_depth = np.load(depth_path)
+        img_depth = (img_depth - img_depth.min()) / (img_depth.max() - img_depth.min())
+
+        beta = np.random.rand(1) * (self.beta_range[1] - self.beta_range[0]) + self.beta_range[0]
+        t = np.exp(-(1- img_depth) * 2.0 * beta)
+        t = t[:, :, np.newaxis]
+        
+        A = np.random.rand(1) * (self.A_range[1] - self.A_range[0]) + self.A_range[0]
+        if np.random.rand(1) < self.color_p:
+            A_random = np.random.rand(3) * (self.color_range[1] - self.color_range[0]) + self.color_range[0]
+            A = A + A_random
+        
+        
+        img_lq = img_gt.copy()
+        # adjust luminance
+        if np.random.rand(1) < 0.5:
+            img_lq = np.power(img_lq, np.random.rand(1) * 1.5 + 1.5)
+        # add gaussian noise
+        if np.random.rand(1) < 0.5:
+            img_lq = add_Gaussian_noise(img_lq)
+
+        # add haze
+        img_lq = img_lq * t + A * (1 - t)
+
+        # add JPEG noise
+        if np.random.rand(1) < 0.5:
+            img_lq = add_JPEG_noise(img_lq)
+
+        if img_gt.shape[-1] > 3:
+            img_gt = img_gt[:, :, :3]
+            img_lq = img_lq[:, :, :3]
+        # augmentation for training
+        if self.opt['phase'] == 'train':
+            input_gt_size = np.min(img_gt.shape[:2])
+            input_lq_size = np.min(img_lq.shape[:2])
+            scale = input_gt_size // input_lq_size
+            gt_size = self.opt['gt_size']
+
+            if self.opt['use_resize_crop']:
+                # random resize
+                if input_gt_size > gt_size:
+                    input_gt_random_size = random.randint(gt_size, input_gt_size)
+                    input_gt_random_size = input_gt_random_size - input_gt_random_size % scale # make sure divisible by scale 
+                    resize_factor = input_gt_random_size / input_gt_size
+                else:
+                    resize_factor = (gt_size+1) / input_gt_size
+                img_gt = random_resize(img_gt, resize_factor)
+                img_lq = random_resize(img_lq, resize_factor)
+                t = random_resize(t, resize_factor)
+
+                # random crop
+                img_gt, img_lq = paired_random_crop(img_gt, img_lq, gt_size, input_gt_size // input_lq_size,
+                                               gt_path)
+
+            # flip, rotation
+            img_gt, img_lq = augment([img_gt, img_lq], self.opt['use_flip'],
+                                     self.opt['use_rot'])
+
+
+        if self.opt['phase'] != 'train':
+            crop_eval_size = self.opt.get('crop_eval_size', None)
+            if crop_eval_size:
+                input_gt_size = img_gt.shape[0]
+                input_lq_size = img_lq.shape[0]
+                scale = input_gt_size // input_lq_size
+                img_gt, img_lq = paired_random_crop(img_gt, img_lq, crop_eval_size, input_gt_size // input_lq_size,
+                                               gt_path)
+
+        # TODO: color space transform
+        # BGR to RGB, HWC to CHW, numpy to tensor
+        img_gt, img_lq = img2tensor([img_gt, img_lq], bgr2rgb=True, float32=True)
+
+        return {
+            'lq': img_lq,
+            'gt': img_gt,
+            'lq_path': gt_path,
+            'gt_path': gt_path
+        }
+
+    def __len__(self):
+        return len(self.gt_paths)

basicsr/data/prefetch_dataloader.py

+import queue as Queue
+import threading
+import torch
+from torch.utils.data import DataLoader
+
+
+class PrefetchGenerator(threading.Thread):
+    """A general prefetch generator.
+
+    Ref:
+    https://stackoverflow.com/questions/7323664/python-generator-pre-fetch
+
+    Args:
+        generator: Python generator.
+        num_prefetch_queue (int): Number of prefetch queue.
+    """
+
+    def __init__(self, generator, num_prefetch_queue):
+        threading.Thread.__init__(self)
+        self.queue = Queue.Queue(num_prefetch_queue)
+        self.generator = generator
+        self.daemon = True
+        self.start()
+
+    def run(self):
+        for item in self.generator:
+            self.queue.put(item)
+        self.queue.put(None)
+
+    def __next__(self):
+        next_item = self.queue.get()
+        if next_item is None:
+            raise StopIteration
+        return next_item
+
+    def __iter__(self):
+        return self
+
+
+class PrefetchDataLoader(DataLoader):
+    """Prefetch version of dataloader.
+
+    Ref:
+    https://github.com/IgorSusmelj/pytorch-styleguide/issues/5#
+
+    TODO:
+    Need to test on single gpu and ddp (multi-gpu). There is a known issue in
+    ddp.
+
+    Args:
+        num_prefetch_queue (int): Number of prefetch queue.
+        kwargs (dict): Other arguments for dataloader.
+    """
+
+    def __init__(self, num_prefetch_queue, **kwargs):
+        self.num_prefetch_queue = num_prefetch_queue
+        super(PrefetchDataLoader, self).__init__(**kwargs)
+
+    def __iter__(self):
+        return PrefetchGenerator(super().__iter__(), self.num_prefetch_queue)
+
+
+class CPUPrefetcher():
+    """CPU prefetcher.
+
+    Args:
+        loader: Dataloader.
+    """
+
+    def __init__(self, loader):
+        self.ori_loader = loader
+        self.loader = iter(loader)
+
+    def next(self):
+        try:
+            return next(self.loader)
+        except StopIteration:
+            return None
+
+    def reset(self):
+        self.loader = iter(self.ori_loader)
+
+
+class CUDAPrefetcher():
+    """CUDA prefetcher.
+
+    Ref:
+    https://github.com/NVIDIA/apex/issues/304#
+
+    It may consums more GPU memory.
+
+    Args:
+        loader: Dataloader.
+        opt (dict): Options.
+    """
+
+    def __init__(self, loader, opt):
+        self.ori_loader = loader
+        self.loader = iter(loader)
+        self.opt = opt
+        self.stream = torch.cuda.Stream()
+        self.device = torch.device('cuda' if opt['num_gpu'] != 0 else 'cpu')
+        self.preload()
+
+    def preload(self):
+        try:
+            self.batch = next(self.loader)  # self.batch is a dict
+        except StopIteration:
+            self.batch = None
+            return None
+        # put tensors to gpu
+        with torch.cuda.stream(self.stream):
+            for k, v in self.batch.items():
+                if torch.is_tensor(v):
+                    self.batch[k] = self.batch[k].to(device=self.device, non_blocking=True)
+
+    def next(self):
+        torch.cuda.current_stream().wait_stream(self.stream)
+        batch = self.batch
+        self.preload()
+        return batch
+
+    def reset(self):
+        self.loader = iter(self.ori_loader)
+        self.preload()

basicsr/data/transforms.py

+import cv2
+import random
+import torch
+
+
+def mod_crop(img, scale):
+    """Mod crop images, used during testing.
+
+    Args:
+        img (ndarray): Input image.
+        scale (int): Scale factor.
+
+    Returns:
+        ndarray: Result image.
+    """
+    img = img.copy()
+    if img.ndim in (2, 3):
+        h, w = img.shape[0], img.shape[1]
+        h_remainder, w_remainder = h % scale, w % scale
+        img = img[:h - h_remainder, :w - w_remainder, ...]
+    else:
+        raise ValueError(f'Wrong img ndim: {img.ndim}.')
+    return img
+
+
+def paired_random_crop(img_gts, img_lqs, gt_patch_size, scale, gt_path=None):
+    # if gt_path == '../dataset/OTS/clear_images/2578.jpg':
+    #     print('here')
+    #     print(img_lqs.shape)
+    #     print(img_gts.shape)
+    """Paired random crop. Support Numpy array and Tensor inputs.
+
+    It crops lists of lq and gt images with corresponding locations.
+
+    Args:
+        img_gts (list[ndarray] | ndarray | list[Tensor] | Tensor): GT images. Note that all images
+            should have the same shape. If the input is an ndarray, it will
+            be transformed to a list containing itself.
+        img_lqs (list[ndarray] | ndarray): LQ images. Note that all images
+            should have the same shape. If the input is an ndarray, it will
+            be transformed to a list containing itself.
+        gt_patch_size (int): GT patch size.
+        scale (int): Scale factor.
+        gt_path (str): Path to ground-truth. Default: None.
+
+    Returns:
+        list[ndarray] | ndarray: GT images and LQ images. If returned results
+            only have one element, just return ndarray.
+    """
+
+    if not isinstance(img_gts, list):
+        img_gts = [img_gts]
+    if not isinstance(img_lqs, list):
+        img_lqs = [img_lqs]
+
+    # determine input type: Numpy array or Tensor
+    input_type = 'Tensor' if torch.is_tensor(img_gts[0]) else 'Numpy'
+
+    if input_type == 'Tensor':
+        h_lq, w_lq = img_lqs[0].size()[-2:]
+        h_gt, w_gt = img_gts[0].size()[-2:]
+    else:
+        h_lq, w_lq = img_lqs[0].shape[0:2]
+        h_gt, w_gt = img_gts[0].shape[0:2]
+    lq_patch_size = gt_patch_size // scale
+
+    if h_gt != h_lq * scale or w_gt != w_lq * scale:
+        raise ValueError(f'Scale mismatches. GT ({h_gt}, {w_gt}) is not {scale}x ',
+                         f'multiplication of LQ ({h_lq}, {w_lq}).')
+    if h_lq < lq_patch_size or w_lq < lq_patch_size:
+        raise ValueError(f'LQ ({h_lq}, {w_lq}) is smaller than patch size '
+                         f'({lq_patch_size}, {lq_patch_size}). '
+                         f'Please remove {gt_path}.')
+
+    # randomly choose top and left coordinates for lq patch
+    top = random.randint(0, h_lq - lq_patch_size)
+    left = random.randint(0, w_lq - lq_patch_size)
+
+    # crop lq patch
+    if input_type == 'Tensor':
+        img_lqs = [v[:, :, top:top + lq_patch_size, left:left + lq_patch_size] for v in img_lqs]
+    else:
+        img_lqs = [v[top:top + lq_patch_size, left:left + lq_patch_size, ...] for v in img_lqs]
+
+    # crop corresponding gt patch
+    top_gt, left_gt = int(top * scale), int(left * scale)
+    if input_type == 'Tensor':
+        img_gts = [v[:, :, top_gt:top_gt + gt_patch_size, left_gt:left_gt + gt_patch_size] for v in img_gts]
+    else:
+        img_gts = [v[top_gt:top_gt + gt_patch_size, left_gt:left_gt + gt_patch_size, ...] for v in img_gts]
+    if len(img_gts) == 1:
+        img_gts = img_gts[0]
+    if len(img_lqs) == 1:
+        img_lqs = img_lqs[0]
+    return img_gts, img_lqs
+
+
+def augment(imgs, hflip=True, rotation=True, flows=None, return_status=False):
+    """Augment: horizontal flips OR rotate (0, 90, 180, 270 degrees).
+
+    We use vertical flip and transpose for rotation implementation.
+    All the images in the list use the same augmentation.
+
+    Args:
+        imgs (list[ndarray] | ndarray): Images to be augmented. If the input
+            is an ndarray, it will be transformed to a list.
+        hflip (bool): Horizontal flip. Default: True.
+        rotation (bool): Ratotation. Default: True.
+        flows (list[ndarray]: Flows to be augmented. If the input is an
+            ndarray, it will be transformed to a list.
+            Dimension is (h, w, 2). Default: None.
+        return_status (bool): Return the status of flip and rotation.
+            Default: False.
+
+    Returns:
+        list[ndarray] | ndarray: Augmented images and flows. If returned
+            results only have one element, just return ndarray.
+
+    """
+    hflip = hflip and random.random() < 0.5
+    vflip = rotation and random.random() < 0.5
+    rot90 = rotation and random.random() < 0.5
+
+    def _augment(img):
+        if hflip:  # horizontal
+            cv2.flip(img, 1, img)
+        if vflip:  # vertical
+            cv2.flip(img, 0, img)
+        if rot90:
+            img = img.transpose(1, 0, 2)
+        return img
+
+    def _augment_flow(flow):
+        if hflip:  # horizontal
+            cv2.flip(flow, 1, flow)
+            flow[:, :, 0] *= -1
+        if vflip:  # vertical
+            cv2.flip(flow, 0, flow)
+            flow[:, :, 1] *= -1
+        if rot90:
+            flow = flow.transpose(1, 0, 2)
+            flow = flow[:, :, [1, 0]]
+        return flow
+
+    if not isinstance(imgs, list):
+        imgs = [imgs]
+    imgs = [_augment(img) for img in imgs]
+    if len(imgs) == 1:
+        imgs = imgs[0]
+
+    if flows is not None:
+        if not isinstance(flows, list):
+            flows = [flows]
+        flows = [_augment_flow(flow) for flow in flows]
+        if len(flows) == 1:
+            flows = flows[0]
+        return imgs, flows
+    else:
+        if return_status:
+            return imgs, (hflip, vflip, rot90)
+        else:
+            return imgs
+
+
+def img_rotate(img, angle, center=None, scale=1.0):
+    """Rotate image.
+
+    Args:
+        img (ndarray): Image to be rotated.
+        angle (float): Rotation angle in degrees. Positive values mean
+            counter-clockwise rotation.
+        center (tuple[int]): Rotation center. If the center is None,
+            initialize it as the center of the image. Default: None.
+        scale (float): Isotropic scale factor. Default: 1.0.
+    """
+    (h, w) = img.shape[:2]
+
+    if center is None:
+        center = (w // 2, h // 2)
+
+    matrix = cv2.getRotationMatrix2D(center, angle, scale)
+    rotated_img = cv2.warpAffine(img, matrix, (w, h))
+    return rotated_img

basicsr/losses/__init__.py

+from copy import deepcopy
+
+from basicsr.utils import get_root_logger
+from basicsr.utils.registry import LOSS_REGISTRY
+from .losses import (CharbonnierLoss, GANLoss, L1Loss, MSELoss, PerceptualLoss, WeightedTVLoss, g_path_regularize,
+                     gradient_penalty_loss, r1_penalty)
+
+__all__ = [
+    'L1Loss', 'MSELoss', 'CharbonnierLoss', 'WeightedTVLoss', 'PerceptualLoss', 'GANLoss', 'gradient_penalty_loss',
+    'r1_penalty', 'g_path_regularize'
+]
+
+
+def build_loss(opt):
+    """Build loss from options.
+
+    Args:
+        opt (dict): Configuration. It must contain:
+            type (str): Model type.
+    """
+    opt = deepcopy(opt)
+    loss_type = opt.pop('type')
+    loss = LOSS_REGISTRY.get(loss_type)(**opt)
+    logger = get_root_logger()
+    logger.info(f'Loss [{loss.__class__.__name__}] is created.')
+    return loss

basicsr/losses/loss_util.py

+import functools
+from torch.nn import functional as F
+
+
+def reduce_loss(loss, reduction):
+    """Reduce loss as specified.
+
+    Args:
+        loss (Tensor): Elementwise loss tensor.
+        reduction (str): Options are 'none', 'mean' and 'sum'.
+
+    Returns:
+        Tensor: Reduced loss tensor.
+    """
+    reduction_enum = F._Reduction.get_enum(reduction)
+    # none: 0, elementwise_mean:1, sum: 2
+    if reduction_enum == 0:
+        return loss
+    elif reduction_enum == 1:
+        return loss.mean()
+    else:
+        return loss.sum()
+
+
+def weight_reduce_loss(loss, weight=None, reduction='mean'):
+    """Apply element-wise weight and reduce loss.
+
+    Args:
+        loss (Tensor): Element-wise loss.
+        weight (Tensor): Element-wise weights. Default: None.
+        reduction (str): Same as built-in losses of PyTorch. Options are
+            'none', 'mean' and 'sum'. Default: 'mean'.
+
+    Returns:
+        Tensor: Loss values.
+    """
+    # if weight is specified, apply element-wise weight
+    if weight is not None:
+        assert weight.dim() == loss.dim()
+        assert weight.size(1) == 1 or weight.size(1) == loss.size(1)
+        loss = loss * weight
+
+    # if weight is not specified or reduction is sum, just reduce the loss
+    if weight is None or reduction == 'sum':
+        loss = reduce_loss(loss, reduction)
+    # if reduction is mean, then compute mean over weight region
+    elif reduction == 'mean':
+        if weight.size(1) > 1:
+            weight = weight.sum()
+        else:
+            weight = weight.sum() * loss.size(1)
+        loss = loss.sum() / weight
+
+    return loss
+
+
+def weighted_loss(loss_func):
+    """Create a weighted version of a given loss function.
+
+    To use this decorator, the loss function must have the signature like
+    `loss_func(pred, target, **kwargs)`. The function only needs to compute
+    element-wise loss without any reduction. This decorator will add weight
+    and reduction arguments to the function. The decorated function will have
+    the signature like `loss_func(pred, target, weight=None, reduction='mean',
+    **kwargs)`.
+
+    :Example:
+
+    >>> import torch
+    >>> @weighted_loss
+    >>> def l1_loss(pred, target):
+    >>>     return (pred - target).abs()
+
+    >>> pred = torch.Tensor([0, 2, 3])
+    >>> target = torch.Tensor([1, 1, 1])
+    >>> weight = torch.Tensor([1, 0, 1])
+
+    >>> l1_loss(pred, target)
+    tensor(1.3333)
+    >>> l1_loss(pred, target, weight)
+    tensor(1.5000)
+    >>> l1_loss(pred, target, reduction='none')
+    tensor([1., 1., 2.])
+    >>> l1_loss(pred, target, weight, reduction='sum')
+    tensor(3.)
+    """
+
+    @functools.wraps(loss_func)
+    def wrapper(pred, target, weight=None, reduction='mean', **kwargs):
+        # get element-wise loss
+        loss = loss_func(pred, target, **kwargs)
+        loss = weight_reduce_loss(loss, weight, reduction)
+        return loss
+
+    return wrapper

basicsr/losses/losses.py

+import math
+import torch
+from torch import autograd as autograd
+from torch import nn as nn
+from torch.nn import functional as F
+
+from basicsr.archs.vgg_arch import VGGFeatureExtractor
+from basicsr.utils.registry import LOSS_REGISTRY
+from .loss_util import weighted_loss
+
+import pyiqa
+
+_reduction_modes = ['none', 'mean', 'sum']
+
+@weighted_loss
+def l1_loss(pred, target):
+    return F.l1_loss(pred, target, reduction='none')
+
+
+@weighted_loss
+def mse_loss(pred, target):
+    return F.mse_loss(pred, target, reduction='none')
+
+
+@weighted_loss
+def charbonnier_loss(pred, target, eps=1e-12):
+    return torch.sqrt((pred - target)**2 + eps)
+
+
+@LOSS_REGISTRY.register()
+class LPIPSLoss(nn.Module):
+    """LPIPS loss with vgg backbone.
+    """
+    def __init__(self, loss_weight = 1.0):
+        super(LPIPSLoss, self).__init__()
+        self.model = pyiqa.create_metric('lpips-vgg', as_loss=True)
+        self.loss_weight = loss_weight
+
+    def forward(self, x, gt):
+        return self.model(x, gt) * self.loss_weight, None
+
+
+@LOSS_REGISTRY.register()
+class L1Loss(nn.Module):
+    """L1 (mean absolute error, MAE) loss.
+
+    Args:
+        loss_weight (float): Loss weight for L1 loss. Default: 1.0.
+        reduction (str): Specifies the reduction to apply to the output.
+            Supported choices are 'none' | 'mean' | 'sum'. Default: 'mean'.
+    """
+
+    def __init__(self, loss_weight=1.0, reduction='mean'):
+        super(L1Loss, self).__init__()
+        if reduction not in ['none', 'mean', 'sum']:
+            raise ValueError(f'Unsupported reduction mode: {reduction}. Supported ones are: {_reduction_modes}')
+
+        self.loss_weight = loss_weight
+        self.reduction = reduction
+
+    def forward(self, pred, target, weight=None, **kwargs):
+        """
+        Args:
+            pred (Tensor): of shape (N, C, H, W). Predicted tensor.
+            target (Tensor): of shape (N, C, H, W). Ground truth tensor.
+            weight (Tensor, optional): of shape (N, C, H, W). Element-wise weights. Default: None.
+        """
+        return self.loss_weight * l1_loss(pred, target, weight, reduction=self.reduction)
+
+
+@LOSS_REGISTRY.register()
+class MSELoss(nn.Module):
+    """MSE (L2) loss.
+
+    Args:
+        loss_weight (float): Loss weight for MSE loss. Default: 1.0.
+        reduction (str): Specifies the reduction to apply to the output.
+            Supported choices are 'none' | 'mean' | 'sum'. Default: 'mean'.
+    """
+
+    def __init__(self, loss_weight=1.0, reduction='mean'):
+        super(MSELoss, self).__init__()
+        if reduction not in ['none', 'mean', 'sum']:
+            raise ValueError(f'Unsupported reduction mode: {reduction}. Supported ones are: {_reduction_modes}')
+
+        self.loss_weight = loss_weight
+        self.reduction = reduction
+
+    def forward(self, pred, target, weight=None, **kwargs):
+        """
+        Args:
+            pred (Tensor): of shape (N, C, H, W). Predicted tensor.
+            target (Tensor): of shape (N, C, H, W). Ground truth tensor.
+            weight (Tensor, optional): of shape (N, C, H, W). Element-wise weights. Default: None.
+        """
+        return self.loss_weight * mse_loss(pred, target, weight, reduction=self.reduction)
+
+
+@LOSS_REGISTRY.register()
+class CharbonnierLoss(nn.Module):
+    """Charbonnier loss (one variant of Robust L1Loss, a differentiable
+    variant of L1Loss).
+
+    Described in "Deep Laplacian Pyramid Networks for Fast and Accurate
+        Super-Resolution".
+
+    Args:
+        loss_weight (float): Loss weight for L1 loss. Default: 1.0.
+        reduction (str): Specifies the reduction to apply to the output.
+            Supported choices are 'none' | 'mean' | 'sum'. Default: 'mean'.
+        eps (float): A value used to control the curvature near zero. Default: 1e-12.
+    """
+
+    def __init__(self, loss_weight=1.0, reduction='mean', eps=1e-12):
+        super(CharbonnierLoss, self).__init__()
+        if reduction not in ['none', 'mean', 'sum']:
+            raise ValueError(f'Unsupported reduction mode: {reduction}. Supported ones are: {_reduction_modes}')
+
+        self.loss_weight = loss_weight
+        self.reduction = reduction
+        self.eps = eps
+
+    def forward(self, pred, target, weight=None, **kwargs):
+        """
+        Args:
+            pred (Tensor): of shape (N, C, H, W). Predicted tensor.
+            target (Tensor): of shape (N, C, H, W). Ground truth tensor.
+            weight (Tensor, optional): of shape (N, C, H, W). Element-wise weights. Default: None.
+        """
+        return self.loss_weight * charbonnier_loss(pred, target, weight, eps=self.eps, reduction=self.reduction)
+
+
+@LOSS_REGISTRY.register()
+class WeightedTVLoss(L1Loss):
+    """Weighted TV loss.
+
+    Args:
+        loss_weight (float): Loss weight. Default: 1.0.
+    """
+
+    def __init__(self, loss_weight=1.0, reduction='mean'):
+        if reduction not in ['mean', 'sum']:
+            raise ValueError(f'Unsupported reduction mode: {reduction}. Supported ones are: mean | sum')
+        super(WeightedTVLoss, self).__init__(loss_weight=loss_weight, reduction=reduction)
+
+    def forward(self, pred, weight=None):
+        if weight is None:
+            y_weight = None
+            x_weight = None
+        else:
+            y_weight = weight[:, :, :-1, :]
+            x_weight = weight[:, :, :, :-1]
+
+        y_diff = super().forward(pred[:, :, :-1, :], pred[:, :, 1:, :], weight=y_weight)
+        x_diff = super().forward(pred[:, :, :, :-1], pred[:, :, :, 1:], weight=x_weight)
+
+        loss = x_diff + y_diff
+
+        return loss
+
+
+@LOSS_REGISTRY.register()
+class PerceptualLoss(nn.Module):
+    """Perceptual loss with commonly used style loss.
+
+    Args:
+        layer_weights (dict): The weight for each layer of vgg feature.
+            Here is an example: {'conv5_4': 1.}, which means the conv5_4
+            feature layer (before relu5_4) will be extracted with weight
+            1.0 in calculating losses.
+        vgg_type (str): The type of vgg network used as feature extractor.
+            Default: 'vgg19'.
+        use_input_norm (bool):  If True, normalize the input image in vgg.
+            Default: True.
+        range_norm (bool): If True, norm images with range [-1, 1] to [0, 1].
+            Default: False.
+        perceptual_weight (float): If `perceptual_weight > 0`, the perceptual
+            loss will be calculated and the loss will multiplied by the
+            weight. Default: 1.0.
+        style_weight (float): If `style_weight > 0`, the style loss will be
+            calculated and the loss will multiplied by the weight.
+            Default: 0.
+        criterion (str): Criterion used for perceptual loss. Default: 'l1'.
+    """
+
+    def __init__(self,
+                 layer_weights,
+                 vgg_type='vgg19',
+                 use_input_norm=True,
+                 range_norm=False,
+                 perceptual_weight=1.0,
+                 style_weight=0.,
+                 criterion='l1'):
+        super(PerceptualLoss, self).__init__()
+        self.perceptual_weight = perceptual_weight
+        self.style_weight = style_weight
+        self.layer_weights = layer_weights
+        self.vgg = VGGFeatureExtractor(
+            layer_name_list=list(layer_weights.keys()),
+            vgg_type=vgg_type,
+            use_input_norm=use_input_norm,
+            range_norm=range_norm)
+
+        self.criterion_type = criterion
+        if self.criterion_type == 'l1':
+            self.criterion = torch.nn.L1Loss()
+        elif self.criterion_type == 'l2':
+            self.criterion = torch.nn.L2loss()
+        elif self.criterion_type == 'fro':
+            self.criterion = None
+        else:
+            raise NotImplementedError(f'{criterion} criterion has not been supported.')
+
+    def forward(self, x, gt):
+        """Forward function.
+
+        Args:
+            x (Tensor): Input tensor with shape (n, c, h, w).
+            gt (Tensor): Ground-truth tensor with shape (n, c, h, w).
+
+        Returns:
+            Tensor: Forward results.
+        """
+        # extract vgg features
+        x_features = self.vgg(x)
+        gt_features = self.vgg(gt.detach())
+
+        # calculate perceptual loss
+        if self.perceptual_weight > 0:
+            percep_loss = 0
+            for k in x_features.keys():
+                if self.criterion_type == 'fro':
+                    percep_loss += torch.norm(x_features[k] - gt_features[k], p='fro') * self.layer_weights[k]
+                else:
+                    percep_loss += self.criterion(x_features[k], gt_features[k]) * self.layer_weights[k]
+            percep_loss *= self.perceptual_weight
+        else:
+            percep_loss = None
+
+        # calculate style loss
+        if self.style_weight > 0:
+            style_loss = 0
+            for k in x_features.keys():
+                if self.criterion_type == 'fro':
+                    style_loss += torch.norm(
+                        self._gram_mat(x_features[k]) - self._gram_mat(gt_features[k]), p='fro') * self.layer_weights[k]
+                else:
+                    style_loss += self.criterion(self._gram_mat(x_features[k]), self._gram_mat(
+                        gt_features[k])) * self.layer_weights[k]
+            style_loss *= self.style_weight
+        else:
+            style_loss = None
+
+        return percep_loss, style_loss
+
+    def _gram_mat(self, x):
+        """Calculate Gram matrix.
+
+        Args:
+            x (torch.Tensor): Tensor with shape of (n, c, h, w).
+
+        Returns:
+            torch.Tensor: Gram matrix.
+        """
+        n, c, h, w = x.size()
+        features = x.view(n, c, w * h)
+        features_t = features.transpose(1, 2)
+        gram = features.bmm(features_t) / (c * h * w)
+        return gram
+
+
+@LOSS_REGISTRY.register()
+class GANLoss(nn.Module):
+    """Define GAN loss.
+
+    Args:
+        gan_type (str): Support 'vanilla', 'lsgan', 'wgan', 'hinge'.
+        real_label_val (float): The value for real label. Default: 1.0.
+        fake_label_val (float): The value for fake label. Default: 0.0.
+        loss_weight (float): Loss weight. Default: 1.0.
+            Note that loss_weight is only for generators; and it is always 1.0
+            for discriminators.
+    """
+
+    def __init__(self, gan_type, real_label_val=1.0, fake_label_val=0.0, loss_weight=1.0):
+        super(GANLoss, self).__init__()
+        self.gan_type = gan_type
+        self.loss_weight = loss_weight
+        self.real_label_val = real_label_val
+        self.fake_label_val = fake_label_val
+
+        if self.gan_type == 'vanilla':
+            self.loss = nn.BCEWithLogitsLoss()
+        elif self.gan_type == 'lsgan':
+            self.loss = nn.MSELoss()
+        elif self.gan_type == 'wgan':
+            self.loss = self._wgan_loss
+        elif self.gan_type == 'wgan_softplus':
+            self.loss = self._wgan_softplus_loss
+        elif self.gan_type == 'hinge':
+            self.loss = nn.ReLU()
+        else:
+            raise NotImplementedError(f'GAN type {self.gan_type} is not implemented.')
+
+    def _wgan_loss(self, input, target):
+        """wgan loss.
+
+        Args:
+            input (Tensor): Input tensor.
+            target (bool): Target label.
+
+        Returns:
+            Tensor: wgan loss.
+        """
+        return -input.mean() if target else input.mean()
+
+    def _wgan_softplus_loss(self, input, target):
+        """wgan loss with soft plus. softplus is a smooth approximation to the
+        ReLU function.
+
+        In StyleGAN2, it is called:
+            Logistic loss for discriminator;
+            Non-saturating loss for generator.
+
+        Args:
+            input (Tensor): Input tensor.
+            target (bool): Target label.
+
+        Returns:
+            Tensor: wgan loss.
+        """
+        return F.softplus(-input).mean() if target else F.softplus(input).mean()
+
+    def get_target_label(self, input, target_is_real):
+        """Get target label.
+
+        Args:
+            input (Tensor): Input tensor.
+            target_is_real (bool): Whether the target is real or fake.
+
+        Returns:
+            (bool | Tensor): Target tensor. Return bool for wgan, otherwise,
+                return Tensor.
+        """
+
+        if self.gan_type in ['wgan', 'wgan_softplus']:
+            return target_is_real
+        target_val = (self.real_label_val if target_is_real else self.fake_label_val)
+        return input.new_ones(input.size()) * target_val
+
+    def forward(self, input, target_is_real, is_disc=False):
+        """
+        Args:
+            input (Tensor): The input for the loss module, i.e., the network
+                prediction.
+            target_is_real (bool): Whether the targe is real or fake.
+            is_disc (bool): Whether the loss for discriminators or not.
+                Default: False.
+
+        Returns:
+            Tensor: GAN loss value.
+        """
+        target_label = self.get_target_label(input, target_is_real)
+        if self.gan_type == 'hinge':
+            if is_disc:  # for discriminators in hinge-gan
+                input = -input if target_is_real else input
+                loss = self.loss(1 + input).mean()
+            else:  # for generators in hinge-gan
+                loss = -input.mean()
+        else:  # other gan types
+            loss = self.loss(input, target_label)
+
+        # loss_weight is always 1.0 for discriminators
+        return loss if is_disc else loss * self.loss_weight
+
+
+@LOSS_REGISTRY.register()
+class MultiScaleGANLoss(GANLoss):
+    """
+    MultiScaleGANLoss accepts a list of predictions
+    """
+
+    def __init__(self, gan_type, real_label_val=1.0, fake_label_val=0.0, loss_weight=1.0):
+        super(MultiScaleGANLoss, self).__init__(gan_type, real_label_val, fake_label_val, loss_weight)
+
+    def forward(self, input, target_is_real, is_disc=False):
+        """
+        The input is a list of tensors, or a list of (a list of tensors)
+        """
+        if isinstance(input, list):
+            loss = 0
+            for pred_i in input:
+                if isinstance(pred_i, list):
+                    # Only compute GAN loss for the last layer
+                    # in case of multiscale feature matching
+                    pred_i = pred_i[-1]
+                # Safe operation: 0-dim tensor calling self.mean() does nothing
+                loss_tensor = super().forward(pred_i, target_is_real, is_disc).mean()
+                loss += loss_tensor
+            return loss / len(input)
+        else:
+            return super().forward(input, target_is_real, is_disc)
+
+
+def r1_penalty(real_pred, real_img):
+    """R1 regularization for discriminator. The core idea is to
+        penalize the gradient on real data alone: when the
+        generator distribution produces the true data distribution
+        and the discriminator is equal to 0 on the data manifold, the
+        gradient penalty ensures that the discriminator cannot create
+        a non-zero gradient orthogonal to the data manifold without
+        suffering a loss in the GAN game.
+
+        Ref:
+        Eq. 9 in Which training methods for GANs do actually converge.
+        """
+    grad_real = autograd.grad(outputs=real_pred.sum(), inputs=real_img, create_graph=True)[0]
+    grad_penalty = grad_real.pow(2).view(grad_real.shape[0], -1).sum(1).mean()
+    return grad_penalty
+
+
+def g_path_regularize(fake_img, latents, mean_path_length, decay=0.01):
+    noise = torch.randn_like(fake_img) / math.sqrt(fake_img.shape[2] * fake_img.shape[3])
+    grad = autograd.grad(outputs=(fake_img * noise).sum(), inputs=latents, create_graph=True)[0]
+    path_lengths = torch.sqrt(grad.pow(2).sum(2).mean(1))
+
+    path_mean = mean_path_length + decay * (path_lengths.mean() - mean_path_length)
+
+    path_penalty = (path_lengths - path_mean).pow(2).mean()
+
+    return path_penalty, path_lengths.detach().mean(), path_mean.detach()
+
+
+def gradient_penalty_loss(discriminator, real_data, fake_data, weight=None):
+    """Calculate gradient penalty for wgan-gp.
+
+    Args:
+        discriminator (nn.Module): Network for the discriminator.
+        real_data (Tensor): Real input data.
+        fake_data (Tensor): Fake input data.
+        weight (Tensor): Weight tensor. Default: None.
+
+    Returns:
+        Tensor: A tensor for gradient penalty.
+    """
+
+    batch_size = real_data.size(0)
+    alpha = real_data.new_tensor(torch.rand(batch_size, 1, 1, 1))
+
+    # interpolate between real_data and fake_data
+    interpolates = alpha * real_data + (1. - alpha) * fake_data
+    interpolates = autograd.Variable(interpolates, requires_grad=True)
+
+    disc_interpolates = discriminator(interpolates)
+    gradients = autograd.grad(
+        outputs=disc_interpolates,
+        inputs=interpolates,
+        grad_outputs=torch.ones_like(disc_interpolates),
+        create_graph=True,
+        retain_graph=True,
+        only_inputs=True)[0]
+
+    if weight is not None:
+        gradients = gradients * weight
+
+    gradients_penalty = ((gradients.norm(2, dim=1) - 1)**2).mean()
+    if weight is not None:
+        gradients_penalty /= torch.mean(weight)
+
+    return gradients_penalty
+
+
+@LOSS_REGISTRY.register()
+class GANFeatLoss(nn.Module):
+    """Define feature matching loss for gans
+
+    Args:
+        criterion (str): Support 'l1', 'l2', 'charbonnier'.
+        loss_weight (float): Loss weight. Default: 1.0.
+        reduction (str): Specifies the reduction to apply to the output.
+            Supported choices are 'none' | 'mean' | 'sum'. Default: 'mean'.
+    """
+
+    def __init__(self, criterion='l1', loss_weight=1.0, reduction='mean'):
+        super(GANFeatLoss, self).__init__()
+        if criterion == 'l1':
+            self.loss_op = L1Loss(loss_weight, reduction)
+        elif criterion == 'l2':
+            self.loss_op = MSELoss(loss_weight, reduction)
+        elif criterion == 'charbonnier':
+            self.loss_op = CharbonnierLoss(loss_weight, reduction)
+        else:
+            raise ValueError(f'Unsupported loss mode: {criterion}. Supported ones are: l1|l2|charbonnier')
+
+        self.loss_weight = loss_weight
+
+    def forward(self, pred_fake, pred_real):
+        num_d = len(pred_fake)
+        loss = 0
+        for i in range(num_d):  # for each discriminator
+            # last output is the final prediction, exclude it
+            num_intermediate_outputs = len(pred_fake[i]) - 1
+            for j in range(num_intermediate_outputs):  # for each layer output
+                unweighted_loss = self.loss_op(pred_fake[i][j], pred_real[i][j].detach())
+                loss += unweighted_loss / num_d
+        return loss * self.loss_weight

basicsr/metrics/calculate_fid_folder.py

+import argparse
+import math
+import numpy as np
+import torch
+from torch.utils.data import DataLoader
+
+from basicsr.data import build_dataset
+from basicsr.metrics.fid import calculate_fid, extract_inception_features, load_patched_inception_v3
+
+
+def calculate_fid_folder():
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument('folder', type=str, help='Path to the folder.')
+    parser.add_argument('--fid_stats', type=str, help='Path to the dataset fid statistics.')
+    parser.add_argument('--batch_size', type=int, default=64)
+    parser.add_argument('--num_sample', type=int, default=50000)
+    parser.add_argument('--num_workers', type=int, default=4)
+    parser.add_argument('--backend', type=str, default='disk', help='io backend for dataset. Option: disk, lmdb')
+    args = parser.parse_args()
+
+    # inception model
+    inception = load_patched_inception_v3(device)
+
+    # create dataset
+    opt = {}
+    opt['name'] = 'SingleImageDataset'
+    opt['type'] = 'SingleImageDataset'
+    opt['dataroot_lq'] = args.folder
+    opt['io_backend'] = dict(type=args.backend)
+    opt['mean'] = [0.5, 0.5, 0.5]
+    opt['std'] = [0.5, 0.5, 0.5]
+    dataset = build_dataset(opt)
+
+    # create dataloader
+    data_loader = DataLoader(
+        dataset=dataset,
+        batch_size=args.batch_size,
+        shuffle=False,
+        num_workers=args.num_workers,
+        sampler=None,
+        drop_last=False)
+    args.num_sample = min(args.num_sample, len(dataset))
+    total_batch = math.ceil(args.num_sample / args.batch_size)
+
+    def data_generator(data_loader, total_batch):
+        for idx, data in enumerate(data_loader):
+            if idx >= total_batch:
+                break
+            else:
+                yield data['lq']
+
+    features = extract_inception_features(data_generator(data_loader, total_batch), inception, total_batch, device)
+    features = features.numpy()
+    total_len = features.shape[0]
+    features = features[:args.num_sample]
+    print(f'Extracted {total_len} features, use the first {features.shape[0]} features to calculate stats.')
+
+    sample_mean = np.mean(features, 0)
+    sample_cov = np.cov(features, rowvar=False)
+
+    # load the dataset stats
+    stats = torch.load(args.fid_stats)
+    real_mean = stats['mean']
+    real_cov = stats['cov']
+
+    # calculate FID metric
+    fid = calculate_fid(sample_mean, sample_cov, real_mean, real_cov)
+    print('fid:', fid)
+
+
+if __name__ == '__main__':
+    calculate_fid_folder()