add realesrgan.

README.md

@@ -14,15 +14,19 @@ S-Lab, Nanyang Technological University
 
 <img src="assets/network.jpg" width="800px"/>
 
+&emsp;
+[![GitHub Stars](https://img.shields.io/github/stars/sczhou/CodeFormer?style=social)](https://github.com/sczhou/CodeFormer) 
+Please help to star this repo if CodeFormer is helpful to your pothos or projects. Thanks! :hugs: 
+
 ### Updates
 
-- **2022.07.29**: New face detector with supporting `['YOLOv5', 'RetinaFace']`. :hugs: 
+- **2022.08.07**: Integrate Real-ESRGAN to support background image enhancement.
+- **2022.07.29**: New face detector with supporting `['YOLOv5', 'RetinaFace']`. 
 - **2022.07.17**: The Colab demo of CodeFormer is available now. <a href="https://colab.research.google.com/drive/1m52PNveE4PBhYrecj34cnpEeiHcC5LTb?usp=sharing"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="google colab logo"></a>
 - **2022.07.16**: Test code for face restoration is released. :blush:
 - **2022.06.21**: This repo is created.
 
 
-
 #### Face Restoration
 
 <img src="assets/restoration_result1.png" width="400px"/> <img src="assets/restoration_result2.png" width="400px"/>
@@ -83,6 +87,8 @@ You can put the testing images in the `inputs/TestWhole` folder. If you would li
 python inference_codeformer.py --w 0.5 --has_aligned --test_path [input folder]
 
 # For the whole images
+# If you want to enhance the background regions with Real-ESRGAN,
+# you can add '--bg_upsampler realesrgan' in the following command
 python inference_codeformer.py --w 0.7 --test_path [input folder]
 ```
 

basicsr/archs/arch_util.py

@@ -0,0 +1,318 @@
+import collections.abc
+import math
+import torch
+import torchvision
+import warnings
+from distutils.version import LooseVersion
+from itertools import repeat
+from torch import nn as nn
+from torch.nn import functional as F
+from torch.nn import init as init
+from torch.nn.modules.batchnorm import _BatchNorm
+
+from basicsr.ops.dcn import ModulatedDeformConvPack, modulated_deform_conv
+from basicsr.utils import get_root_logger
+
+
+@torch.no_grad()
+def default_init_weights(module_list, scale=1, bias_fill=0, **kwargs):
+    """Initialize network weights.
+
+    Args:
+        module_list (list[nn.Module] | nn.Module): Modules to be initialized.
+        scale (float): Scale initialized weights, especially for residual
+            blocks. Default: 1.
+        bias_fill (float): The value to fill bias. Default: 0
+        kwargs (dict): Other arguments for initialization function.
+    """
+    if not isinstance(module_list, list):
+        module_list = [module_list]
+    for module in module_list:
+        for m in module.modules():
+            if isinstance(m, nn.Conv2d):
+                init.kaiming_normal_(m.weight, **kwargs)
+                m.weight.data *= scale
+                if m.bias is not None:
+                    m.bias.data.fill_(bias_fill)
+            elif isinstance(m, nn.Linear):
+                init.kaiming_normal_(m.weight, **kwargs)
+                m.weight.data *= scale
+                if m.bias is not None:
+                    m.bias.data.fill_(bias_fill)
+            elif isinstance(m, _BatchNorm):
+                init.constant_(m.weight, 1)
+                if m.bias is not None:
+                    m.bias.data.fill_(bias_fill)
+
+
+def make_layer(basic_block, num_basic_block, **kwarg):
+    """Make layers by stacking the same blocks.
+
+    Args:
+        basic_block (nn.module): nn.module class for basic block.
+        num_basic_block (int): number of blocks.
+
+    Returns:
+        nn.Sequential: Stacked blocks in nn.Sequential.
+    """
+    layers = []
+    for _ in range(num_basic_block):
+        layers.append(basic_block(**kwarg))
+    return nn.Sequential(*layers)
+
+
+class ResidualBlockNoBN(nn.Module):
+    """Residual block without BN.
+
+    It has a style of:
+        ---Conv-ReLU-Conv-+-
+         |________________|
+
+    Args:
+        num_feat (int): Channel number of intermediate features.
+            Default: 64.
+        res_scale (float): Residual scale. Default: 1.
+        pytorch_init (bool): If set to True, use pytorch default init,
+            otherwise, use default_init_weights. Default: False.
+    """
+
+    def __init__(self, num_feat=64, res_scale=1, pytorch_init=False):
+        super(ResidualBlockNoBN, self).__init__()
+        self.res_scale = res_scale
+        self.conv1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1, bias=True)
+        self.conv2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1, bias=True)
+        self.relu = nn.ReLU(inplace=True)
+
+        if not pytorch_init:
+            default_init_weights([self.conv1, self.conv2], 0.1)
+
+    def forward(self, x):
+        identity = x
+        out = self.conv2(self.relu(self.conv1(x)))
+        return identity + out * self.res_scale
+
+
+class Upsample(nn.Sequential):
+    """Upsample module.
+
+    Args:
+        scale (int): Scale factor. Supported scales: 2^n and 3.
+        num_feat (int): Channel number of intermediate features.
+    """
+
+    def __init__(self, scale, num_feat):
+        m = []
+        if (scale & (scale - 1)) == 0:  # scale = 2^n
+            for _ in range(int(math.log(scale, 2))):
+                m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
+                m.append(nn.PixelShuffle(2))
+        elif scale == 3:
+            m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
+            m.append(nn.PixelShuffle(3))
+        else:
+            raise ValueError(f'scale {scale} is not supported. Supported scales: 2^n and 3.')
+        super(Upsample, self).__init__(*m)
+
+
+def flow_warp(x, flow, interp_mode='bilinear', padding_mode='zeros', align_corners=True):
+    """Warp an image or feature map with optical flow.
+
+    Args:
+        x (Tensor): Tensor with size (n, c, h, w).
+        flow (Tensor): Tensor with size (n, h, w, 2), normal value.
+        interp_mode (str): 'nearest' or 'bilinear'. Default: 'bilinear'.
+        padding_mode (str): 'zeros' or 'border' or 'reflection'.
+            Default: 'zeros'.
+        align_corners (bool): Before pytorch 1.3, the default value is
+            align_corners=True. After pytorch 1.3, the default value is
+            align_corners=False. Here, we use the True as default.
+
+    Returns:
+        Tensor: Warped image or feature map.
+    """
+    assert x.size()[-2:] == flow.size()[1:3]
+    _, _, h, w = x.size()
+    # create mesh grid
+    grid_y, grid_x = torch.meshgrid(torch.arange(0, h).type_as(x), torch.arange(0, w).type_as(x))
+    grid = torch.stack((grid_x, grid_y), 2).float()  # W(x), H(y), 2
+    grid.requires_grad = False
+
+    vgrid = grid + flow
+    # scale grid to [-1,1]
+    vgrid_x = 2.0 * vgrid[:, :, :, 0] / max(w - 1, 1) - 1.0
+    vgrid_y = 2.0 * vgrid[:, :, :, 1] / max(h - 1, 1) - 1.0
+    vgrid_scaled = torch.stack((vgrid_x, vgrid_y), dim=3)
+    output = F.grid_sample(x, vgrid_scaled, mode=interp_mode, padding_mode=padding_mode, align_corners=align_corners)
+
+    # TODO, what if align_corners=False
+    return output
+
+
+def resize_flow(flow, size_type, sizes, interp_mode='bilinear', align_corners=False):
+    """Resize a flow according to ratio or shape.
+
+    Args:
+        flow (Tensor): Precomputed flow. shape [N, 2, H, W].
+        size_type (str): 'ratio' or 'shape'.
+        sizes (list[int | float]): the ratio for resizing or the final output
+            shape.
+            1) The order of ratio should be [ratio_h, ratio_w]. For
+            downsampling, the ratio should be smaller than 1.0 (i.e., ratio
+            < 1.0). For upsampling, the ratio should be larger than 1.0 (i.e.,
+            ratio > 1.0).
+            2) The order of output_size should be [out_h, out_w].
+        interp_mode (str): The mode of interpolation for resizing.
+            Default: 'bilinear'.
+        align_corners (bool): Whether align corners. Default: False.
+
+    Returns:
+        Tensor: Resized flow.
+    """
+    _, _, flow_h, flow_w = flow.size()
+    if size_type == 'ratio':
+        output_h, output_w = int(flow_h * sizes[0]), int(flow_w * sizes[1])
+    elif size_type == 'shape':
+        output_h, output_w = sizes[0], sizes[1]
+    else:
+        raise ValueError(f'Size type should be ratio or shape, but got type {size_type}.')
+
+    input_flow = flow.clone()
+    ratio_h = output_h / flow_h
+    ratio_w = output_w / flow_w
+    input_flow[:, 0, :, :] *= ratio_w
+    input_flow[:, 1, :, :] *= ratio_h
+    resized_flow = F.interpolate(
+        input=input_flow, size=(output_h, output_w), mode=interp_mode, align_corners=align_corners)
+    return resized_flow
+
+
+# TODO: may write a cpp file
+def pixel_unshuffle(x, scale):
+    """ Pixel unshuffle.
+
+    Args:
+        x (Tensor): Input feature with shape (b, c, hh, hw).
+        scale (int): Downsample ratio.
+
+    Returns:
+        Tensor: the pixel unshuffled feature.
+    """
+    b, c, hh, hw = x.size()
+    out_channel = c * (scale**2)
+    assert hh % scale == 0 and hw % scale == 0
+    h = hh // scale
+    w = hw // scale
+    x_view = x.view(b, c, h, scale, w, scale)
+    return x_view.permute(0, 1, 3, 5, 2, 4).reshape(b, out_channel, h, w)
+
+
+class DCNv2Pack(ModulatedDeformConvPack):
+    """Modulated deformable conv for deformable alignment.
+
+    Different from the official DCNv2Pack, which generates offsets and masks
+    from the preceding features, this DCNv2Pack takes another different
+    features to generate offsets and masks.
+
+    Ref:
+        Delving Deep into Deformable Alignment in Video Super-Resolution.
+    """
+
+    def forward(self, x, feat):
+        out = self.conv_offset(feat)
+        o1, o2, mask = torch.chunk(out, 3, dim=1)
+        offset = torch.cat((o1, o2), dim=1)
+        mask = torch.sigmoid(mask)
+
+        offset_absmean = torch.mean(torch.abs(offset))
+        if offset_absmean > 50:
+            logger = get_root_logger()
+            logger.warning(f'Offset abs mean is {offset_absmean}, larger than 50.')
+
+        if LooseVersion(torchvision.__version__) >= LooseVersion('0.9.0'):
+            return torchvision.ops.deform_conv2d(x, offset, self.weight, self.bias, self.stride, self.padding,
+                                                 self.dilation, mask)
+        else:
+            return modulated_deform_conv(x, offset, mask, self.weight, self.bias, self.stride, self.padding,
+                                         self.dilation, self.groups, self.deformable_groups)
+
+
+def _no_grad_trunc_normal_(tensor, mean, std, a, b):
+    # From: https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/weight_init.py
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1. + math.erf(x / math.sqrt(2.))) / 2.
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn(
+            'mean is more than 2 std from [a, b] in nn.init.trunc_normal_. '
+            'The distribution of values may be incorrect.',
+            stacklevel=2)
+
+    with torch.no_grad():
+        # Values are generated by using a truncated uniform distribution and
+        # then using the inverse CDF for the normal distribution.
+        # Get upper and lower cdf values
+        low = norm_cdf((a - mean) / std)
+        up = norm_cdf((b - mean) / std)
+
+        # Uniformly fill tensor with values from [low, up], then translate to
+        # [2l-1, 2u-1].
+        tensor.uniform_(2 * low - 1, 2 * up - 1)
+
+        # Use inverse cdf transform for normal distribution to get truncated
+        # standard normal
+        tensor.erfinv_()
+
+        # Transform to proper mean, std
+        tensor.mul_(std * math.sqrt(2.))
+        tensor.add_(mean)
+
+        # Clamp to ensure it's in the proper range
+        tensor.clamp_(min=a, max=b)
+        return tensor
+
+
+def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
+    r"""Fills the input Tensor with values drawn from a truncated
+    normal distribution.
+
+    From: https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/weight_init.py
+
+    The values are effectively drawn from the
+    normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \leq \text{mean} \leq b`.
+
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+
+    Examples:
+        >>> w = torch.empty(3, 5)
+        >>> nn.init.trunc_normal_(w)
+    """
+    return _no_grad_trunc_normal_(tensor, mean, std, a, b)
+
+
+# From PyTorch
+def _ntuple(n):
+
+    def parse(x):
+        if isinstance(x, collections.abc.Iterable):
+            return x
+        return tuple(repeat(x, n))
+
+    return parse
+
+
+to_1tuple = _ntuple(1)
+to_2tuple = _ntuple(2)
+to_3tuple = _ntuple(3)
+to_4tuple = _ntuple(4)
+to_ntuple = _ntuple

basicsr/archs/rrdbnet_arch.py

@@ -0,0 +1,119 @@
+import torch
+from torch import nn as nn
+from torch.nn import functional as F
+
+from basicsr.utils.registry import ARCH_REGISTRY
+from .arch_util import default_init_weights, make_layer, pixel_unshuffle
+
+
+class ResidualDenseBlock(nn.Module):
+    """Residual Dense Block.
+
+    Used in RRDB block in ESRGAN.
+
+    Args:
+        num_feat (int): Channel number of intermediate features.
+        num_grow_ch (int): Channels for each growth.
+    """
+
+    def __init__(self, num_feat=64, num_grow_ch=32):
+        super(ResidualDenseBlock, self).__init__()
+        self.conv1 = nn.Conv2d(num_feat, num_grow_ch, 3, 1, 1)
+        self.conv2 = nn.Conv2d(num_feat + num_grow_ch, num_grow_ch, 3, 1, 1)
+        self.conv3 = nn.Conv2d(num_feat + 2 * num_grow_ch, num_grow_ch, 3, 1, 1)
+        self.conv4 = nn.Conv2d(num_feat + 3 * num_grow_ch, num_grow_ch, 3, 1, 1)
+        self.conv5 = nn.Conv2d(num_feat + 4 * num_grow_ch, num_feat, 3, 1, 1)
+
+        self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
+
+        # initialization
+        default_init_weights([self.conv1, self.conv2, self.conv3, self.conv4, self.conv5], 0.1)
+
+    def forward(self, x):
+        x1 = self.lrelu(self.conv1(x))
+        x2 = self.lrelu(self.conv2(torch.cat((x, x1), 1)))
+        x3 = self.lrelu(self.conv3(torch.cat((x, x1, x2), 1)))
+        x4 = self.lrelu(self.conv4(torch.cat((x, x1, x2, x3), 1)))
+        x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), 1))
+        # Emperically, we use 0.2 to scale the residual for better performance
+        return x5 * 0.2 + x
+
+
+class RRDB(nn.Module):
+    """Residual in Residual Dense Block.
+
+    Used in RRDB-Net in ESRGAN.
+
+    Args:
+        num_feat (int): Channel number of intermediate features.
+        num_grow_ch (int): Channels for each growth.
+    """
+
+    def __init__(self, num_feat, num_grow_ch=32):
+        super(RRDB, self).__init__()
+        self.rdb1 = ResidualDenseBlock(num_feat, num_grow_ch)
+        self.rdb2 = ResidualDenseBlock(num_feat, num_grow_ch)
+        self.rdb3 = ResidualDenseBlock(num_feat, num_grow_ch)
+
+    def forward(self, x):
+        out = self.rdb1(x)
+        out = self.rdb2(out)
+        out = self.rdb3(out)
+        # Emperically, we use 0.2 to scale the residual for better performance
+        return out * 0.2 + x
+
+
+@ARCH_REGISTRY.register()
+class RRDBNet(nn.Module):
+    """Networks consisting of Residual in Residual Dense Block, which is used
+    in ESRGAN.
+
+    ESRGAN: Enhanced Super-Resolution Generative Adversarial Networks.
+
+    We extend ESRGAN for scale x2 and scale x1.
+    Note: This is one option for scale 1, scale 2 in RRDBNet.
+    We first employ the pixel-unshuffle (an inverse operation of pixelshuffle to reduce the spatial size
+    and enlarge the channel size before feeding inputs into the main ESRGAN architecture.
+
+    Args:
+        num_in_ch (int): Channel number of inputs.
+        num_out_ch (int): Channel number of outputs.
+        num_feat (int): Channel number of intermediate features.
+            Default: 64
+        num_block (int): Block number in the trunk network. Defaults: 23
+        num_grow_ch (int): Channels for each growth. Default: 32.
+    """
+
+    def __init__(self, num_in_ch, num_out_ch, scale=4, num_feat=64, num_block=23, num_grow_ch=32):
+        super(RRDBNet, self).__init__()
+        self.scale = scale
+        if scale == 2:
+            num_in_ch = num_in_ch * 4
+        elif scale == 1:
+            num_in_ch = num_in_ch * 16
+        self.conv_first = nn.Conv2d(num_in_ch, num_feat, 3, 1, 1)
+        self.body = make_layer(RRDB, num_block, num_feat=num_feat, num_grow_ch=num_grow_ch)
+        self.conv_body = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        # upsample
+        self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+
+        self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
+
+    def forward(self, x):
+        if self.scale == 2:
+            feat = pixel_unshuffle(x, scale=2)
+        elif self.scale == 1:
+            feat = pixel_unshuffle(x, scale=4)
+        else:
+            feat = x
+        feat = self.conv_first(feat)
+        body_feat = self.conv_body(self.body(feat))
+        feat = feat + body_feat
+        # upsample
+        feat = self.lrelu(self.conv_up1(F.interpolate(feat, scale_factor=2, mode='nearest')))
+        feat = self.lrelu(self.conv_up2(F.interpolate(feat, scale_factor=2, mode='nearest')))
+        out = self.conv_last(self.lrelu(self.conv_hr(feat)))
+        return out

basicsr/utils/realesrgan_utils.py

@@ -0,0 +1,293 @@
+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,
+                 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 the model_path starts with https, it will first download models to the folder: realesrgan/weights
+        if model_path.startswith('https://'):
+            model_path = load_file_from_url(
+                url=model_path, model_dir=os.path.join('weights/realesrgan'), 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 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.')

inference_codeformer.py

@@ -1,3 +1,4 @@
+# Modified by Shangchen Zhou from: https://github.com/TencentARC/GFPGAN/blob/master/inference_gfpgan.py
 import os
 import cv2
 import argparse
@@ -20,15 +21,41 @@ if __name__ == '__main__':
     parser.add_argument('--has_aligned', action='store_true', help='Input are cropped and aligned faces')
     parser.add_argument('--only_center_face', action='store_true', help='Only restore the center face')
     parser.add_argument('--draw_box', action='store_true')
+    parser.add_argument('--bg_upsampler', type=str, default='realesrgan', help='background upsampler. Default: realesrgan')
+    parser.add_argument('--bg_tile', type=int, default=400, help='Tile size for background sampler. Default: 400')
 
     args = parser.parse_args()
+
+    # ------------------------ input & output ------------------------
     if args.test_path.endswith('/'):  # solve when path ends with /
         args.test_path = args.test_path[:-1]
 
     w = args.w
     result_root = f'results/{os.path.basename(args.test_path)}_{w}'
 
-    # set up the Network
+    # ------------------ set up background upsampler ------------------
+    if args.bg_upsampler == 'realesrgan':
+        if not torch.cuda.is_available():  # CPU
+            import warnings
+            warnings.warn('The unoptimized RealESRGAN is slow on CPU. We do not use it. '
+                          'If you really want to use it, please modify the corresponding codes.')
+            bg_upsampler = None
+        else:
+            from basicsr.archs.rrdbnet_arch import RRDBNet
+            from basicsr.utils.realesrgan_utils import RealESRGANer
+            model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=2)
+            bg_upsampler = RealESRGANer(
+                scale=2,
+                model_path='https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.1/RealESRGAN_x2plus.pth',
+                model=model,
+                tile=args.bg_tile,
+                tile_pad=10,
+                pre_pad=0,
+                half=True)  # need to set False in CPU mode
+    else:
+        bg_upsampler = None
+
+    # ------------------ set up CodeFormer restorer -------------------
     net = ARCH_REGISTRY.get('CodeFormer')(dim_embd=512, codebook_size=1024, n_head=8, n_layers=9, 
                                             connect_list=['32', '64', '128', '256']).to(device)
 
@@ -98,7 +125,12 @@ if __name__ == '__main__':
 
         # paste_back
         if not args.has_aligned:
-            bg_img = None
+            # upsample the background
+            if bg_upsampler is not None:
+                # Now only support RealESRGAN for upsampling background
+                bg_img = bg_upsampler.enhance(img, outscale=args.upscale)[0]
+            else:
+                bg_img = None
             face_helper.get_inverse_affine(None)
             # paste each restored face to the input image
             restored_img = face_helper.paste_faces_to_input_image(upsample_img=bg_img, draw_box=args.draw_box)