Upload 34 files

gfpgan/__init__.py

@@ -0,0 +1,7 @@
+# flake8: noqa
+from .archs import *
+from .data import *
+from .models import *
+from .utils import *
+
+# from .version import *

gfpgan/archs/__init__.py

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

gfpgan/archs/arcface_arch.py

@@ -0,0 +1,245 @@
+import torch.nn as nn
+from basicsr.utils.registry import ARCH_REGISTRY
+
+
+def conv3x3(inplanes, outplanes, stride=1):
+    """A simple wrapper for 3x3 convolution with padding.
+
+    Args:
+        inplanes (int): Channel number of inputs.
+        outplanes (int): Channel number of outputs.
+        stride (int): Stride in convolution. Default: 1.
+    """
+    return nn.Conv2d(inplanes, outplanes, kernel_size=3, stride=stride, padding=1, bias=False)
+
+
+class BasicBlock(nn.Module):
+    """Basic residual block used in the ResNetArcFace architecture.
+
+    Args:
+        inplanes (int): Channel number of inputs.
+        planes (int): Channel number of outputs.
+        stride (int): Stride in convolution. Default: 1.
+        downsample (nn.Module): The downsample module. Default: None.
+    """
+    expansion = 1  # output channel expansion ratio
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None):
+        super(BasicBlock, self).__init__()
+        self.conv1 = conv3x3(inplanes, planes, stride)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.relu = nn.ReLU(inplace=True)
+        self.conv2 = conv3x3(planes, planes)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class IRBlock(nn.Module):
+    """Improved residual block (IR Block) used in the ResNetArcFace architecture.
+
+    Args:
+        inplanes (int): Channel number of inputs.
+        planes (int): Channel number of outputs.
+        stride (int): Stride in convolution. Default: 1.
+        downsample (nn.Module): The downsample module. Default: None.
+        use_se (bool): Whether use the SEBlock (squeeze and excitation block). Default: True.
+    """
+    expansion = 1  # output channel expansion ratio
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None, use_se=True):
+        super(IRBlock, self).__init__()
+        self.bn0 = nn.BatchNorm2d(inplanes)
+        self.conv1 = conv3x3(inplanes, inplanes)
+        self.bn1 = nn.BatchNorm2d(inplanes)
+        self.prelu = nn.PReLU()
+        self.conv2 = conv3x3(inplanes, planes, stride)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.downsample = downsample
+        self.stride = stride
+        self.use_se = use_se
+        if self.use_se:
+            self.se = SEBlock(planes)
+
+    def forward(self, x):
+        residual = x
+        out = self.bn0(x)
+        out = self.conv1(out)
+        out = self.bn1(out)
+        out = self.prelu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        if self.use_se:
+            out = self.se(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.prelu(out)
+
+        return out
+
+
+class Bottleneck(nn.Module):
+    """Bottleneck block used in the ResNetArcFace architecture.
+
+    Args:
+        inplanes (int): Channel number of inputs.
+        planes (int): Channel number of outputs.
+        stride (int): Stride in convolution. Default: 1.
+        downsample (nn.Module): The downsample module. Default: None.
+    """
+    expansion = 4  # output channel expansion ratio
+
+    def __init__(self, inplanes, planes, stride=1, downsample=None):
+        super(Bottleneck, self).__init__()
+        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.conv3 = nn.Conv2d(planes, planes * self.expansion, kernel_size=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
+        self.relu = nn.ReLU(inplace=True)
+        self.downsample = downsample
+        self.stride = stride
+
+    def forward(self, x):
+        residual = x
+
+        out = self.conv1(x)
+        out = self.bn1(out)
+        out = self.relu(out)
+
+        out = self.conv2(out)
+        out = self.bn2(out)
+        out = self.relu(out)
+
+        out = self.conv3(out)
+        out = self.bn3(out)
+
+        if self.downsample is not None:
+            residual = self.downsample(x)
+
+        out += residual
+        out = self.relu(out)
+
+        return out
+
+
+class SEBlock(nn.Module):
+    """The squeeze-and-excitation block (SEBlock) used in the IRBlock.
+
+    Args:
+        channel (int): Channel number of inputs.
+        reduction (int): Channel reduction ration. Default: 16.
+    """
+
+    def __init__(self, channel, reduction=16):
+        super(SEBlock, self).__init__()
+        self.avg_pool = nn.AdaptiveAvgPool2d(1)  # pool to 1x1 without spatial information
+        self.fc = nn.Sequential(
+            nn.Linear(channel, channel // reduction), nn.PReLU(), nn.Linear(channel // reduction, channel),
+            nn.Sigmoid())
+
+    def forward(self, x):
+        b, c, _, _ = x.size()
+        y = self.avg_pool(x).view(b, c)
+        y = self.fc(y).view(b, c, 1, 1)
+        return x * y
+
+
+@ARCH_REGISTRY.register()
+class ResNetArcFace(nn.Module):
+    """ArcFace with ResNet architectures.
+
+    Ref: ArcFace: Additive Angular Margin Loss for Deep Face Recognition.
+
+    Args:
+        block (str): Block used in the ArcFace architecture.
+        layers (tuple(int)): Block numbers in each layer.
+        use_se (bool): Whether use the SEBlock (squeeze and excitation block). Default: True.
+    """
+
+    def __init__(self, block, layers, use_se=True):
+        if block == 'IRBlock':
+            block = IRBlock
+        self.inplanes = 64
+        self.use_se = use_se
+        super(ResNetArcFace, self).__init__()
+
+        self.conv1 = nn.Conv2d(1, 64, kernel_size=3, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(64)
+        self.prelu = nn.PReLU()
+        self.maxpool = nn.MaxPool2d(kernel_size=2, stride=2)
+        self.layer1 = self._make_layer(block, 64, layers[0])
+        self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
+        self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
+        self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
+        self.bn4 = nn.BatchNorm2d(512)
+        self.dropout = nn.Dropout()
+        self.fc5 = nn.Linear(512 * 8 * 8, 512)
+        self.bn5 = nn.BatchNorm1d(512)
+
+        # initialization
+        for m in self.modules():
+            if isinstance(m, nn.Conv2d):
+                nn.init.xavier_normal_(m.weight)
+            elif isinstance(m, nn.BatchNorm2d) or isinstance(m, nn.BatchNorm1d):
+                nn.init.constant_(m.weight, 1)
+                nn.init.constant_(m.bias, 0)
+            elif isinstance(m, nn.Linear):
+                nn.init.xavier_normal_(m.weight)
+                nn.init.constant_(m.bias, 0)
+
+    def _make_layer(self, block, planes, num_blocks, stride=1):
+        downsample = None
+        if stride != 1 or self.inplanes != planes * block.expansion:
+            downsample = nn.Sequential(
+                nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False),
+                nn.BatchNorm2d(planes * block.expansion),
+            )
+        layers = []
+        layers.append(block(self.inplanes, planes, stride, downsample, use_se=self.use_se))
+        self.inplanes = planes
+        for _ in range(1, num_blocks):
+            layers.append(block(self.inplanes, planes, use_se=self.use_se))
+
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.prelu(x)
+        x = self.maxpool(x)
+
+        x = self.layer1(x)
+        x = self.layer2(x)
+        x = self.layer3(x)
+        x = self.layer4(x)
+        x = self.bn4(x)
+        x = self.dropout(x)
+        x = x.view(x.size(0), -1)
+        x = self.fc5(x)
+        x = self.bn5(x)
+
+        return x

gfpgan/archs/gfpgan_bilinear_arch.py

@@ -0,0 +1,312 @@
+import math
+import random
+import torch
+from basicsr.utils.registry import ARCH_REGISTRY
+from torch import nn
+
+from .gfpganv1_arch import ResUpBlock
+from .stylegan2_bilinear_arch import (ConvLayer, EqualConv2d, EqualLinear, ResBlock, ScaledLeakyReLU,
+                                      StyleGAN2GeneratorBilinear)
+
+
+class StyleGAN2GeneratorBilinearSFT(StyleGAN2GeneratorBilinear):
+    """StyleGAN2 Generator with SFT modulation (Spatial Feature Transform).
+
+    It is the bilinear version. It does not use the complicated UpFirDnSmooth function that is not friendly for
+    deployment. It can be easily converted to the clean version: StyleGAN2GeneratorCSFT.
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        num_style_feat (int): Channel number of style features. Default: 512.
+        num_mlp (int): Layer number of MLP style layers. Default: 8.
+        channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
+        lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
+        narrow (float): The narrow ratio for channels. Default: 1.
+        sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
+    """
+
+    def __init__(self,
+                 out_size,
+                 num_style_feat=512,
+                 num_mlp=8,
+                 channel_multiplier=2,
+                 lr_mlp=0.01,
+                 narrow=1,
+                 sft_half=False):
+        super(StyleGAN2GeneratorBilinearSFT, self).__init__(
+            out_size,
+            num_style_feat=num_style_feat,
+            num_mlp=num_mlp,
+            channel_multiplier=channel_multiplier,
+            lr_mlp=lr_mlp,
+            narrow=narrow)
+        self.sft_half = sft_half
+
+    def forward(self,
+                styles,
+                conditions,
+                input_is_latent=False,
+                noise=None,
+                randomize_noise=True,
+                truncation=1,
+                truncation_latent=None,
+                inject_index=None,
+                return_latents=False):
+        """Forward function for StyleGAN2GeneratorBilinearSFT.
+
+        Args:
+            styles (list[Tensor]): Sample codes of styles.
+            conditions (list[Tensor]): SFT conditions to generators.
+            input_is_latent (bool): Whether input is latent style. Default: False.
+            noise (Tensor | None): Input noise or None. Default: None.
+            randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
+            truncation (float): The truncation ratio. Default: 1.
+            truncation_latent (Tensor | None): The truncation latent tensor. Default: None.
+            inject_index (int | None): The injection index for mixing noise. Default: None.
+            return_latents (bool): Whether to return style latents. Default: False.
+        """
+        # style codes -> latents with Style MLP layer
+        if not input_is_latent:
+            styles = [self.style_mlp(s) for s in styles]
+        # noises
+        if noise is None:
+            if randomize_noise:
+                noise = [None] * self.num_layers  # for each style conv layer
+            else:  # use the stored noise
+                noise = [getattr(self.noises, f'noise{i}') for i in range(self.num_layers)]
+        # style truncation
+        if truncation < 1:
+            style_truncation = []
+            for style in styles:
+                style_truncation.append(truncation_latent + truncation * (style - truncation_latent))
+            styles = style_truncation
+        # get style latents with injection
+        if len(styles) == 1:
+            inject_index = self.num_latent
+
+            if styles[0].ndim < 3:
+                # repeat latent code for all the layers
+                latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            else:  # used for encoder with different latent code for each layer
+                latent = styles[0]
+        elif len(styles) == 2:  # mixing noises
+            if inject_index is None:
+                inject_index = random.randint(1, self.num_latent - 1)
+            latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            latent2 = styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
+            latent = torch.cat([latent1, latent2], 1)
+
+        # main generation
+        out = self.constant_input(latent.shape[0])
+        out = self.style_conv1(out, latent[:, 0], noise=noise[0])
+        skip = self.to_rgb1(out, latent[:, 1])
+
+        i = 1
+        for conv1, conv2, noise1, noise2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], noise[1::2],
+                                                        noise[2::2], self.to_rgbs):
+            out = conv1(out, latent[:, i], noise=noise1)
+
+            # the conditions may have fewer levels
+            if i < len(conditions):
+                # SFT part to combine the conditions
+                if self.sft_half:  # only apply SFT to half of the channels
+                    out_same, out_sft = torch.split(out, int(out.size(1) // 2), dim=1)
+                    out_sft = out_sft * conditions[i - 1] + conditions[i]
+                    out = torch.cat([out_same, out_sft], dim=1)
+                else:  # apply SFT to all the channels
+                    out = out * conditions[i - 1] + conditions[i]
+
+            out = conv2(out, latent[:, i + 1], noise=noise2)
+            skip = to_rgb(out, latent[:, i + 2], skip)  # feature back to the rgb space
+            i += 2
+
+        image = skip
+
+        if return_latents:
+            return image, latent
+        else:
+            return image, None
+
+
+@ARCH_REGISTRY.register()
+class GFPGANBilinear(nn.Module):
+    """The GFPGAN architecture: Unet + StyleGAN2 decoder with SFT.
+
+    It is the bilinear version and it does not use the complicated UpFirDnSmooth function that is not friendly for
+    deployment. It can be easily converted to the clean version: GFPGANv1Clean.
+
+
+    Ref: GFP-GAN: Towards Real-World Blind Face Restoration with Generative Facial Prior.
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        num_style_feat (int): Channel number of style features. Default: 512.
+        channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
+        decoder_load_path (str): The path to the pre-trained decoder model (usually, the StyleGAN2). Default: None.
+        fix_decoder (bool): Whether to fix the decoder. Default: True.
+
+        num_mlp (int): Layer number of MLP style layers. Default: 8.
+        lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
+        input_is_latent (bool): Whether input is latent style. Default: False.
+        different_w (bool): Whether to use different latent w for different layers. Default: False.
+        narrow (float): The narrow ratio for channels. Default: 1.
+        sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
+    """
+
+    def __init__(
+            self,
+            out_size,
+            num_style_feat=512,
+            channel_multiplier=1,
+            decoder_load_path=None,
+            fix_decoder=True,
+            # for stylegan decoder
+            num_mlp=8,
+            lr_mlp=0.01,
+            input_is_latent=False,
+            different_w=False,
+            narrow=1,
+            sft_half=False):
+
+        super(GFPGANBilinear, self).__init__()
+        self.input_is_latent = input_is_latent
+        self.different_w = different_w
+        self.num_style_feat = num_style_feat
+
+        unet_narrow = narrow * 0.5  # by default, use a half of input channels
+        channels = {
+            '4': int(512 * unet_narrow),
+            '8': int(512 * unet_narrow),
+            '16': int(512 * unet_narrow),
+            '32': int(512 * unet_narrow),
+            '64': int(256 * channel_multiplier * unet_narrow),
+            '128': int(128 * channel_multiplier * unet_narrow),
+            '256': int(64 * channel_multiplier * unet_narrow),
+            '512': int(32 * channel_multiplier * unet_narrow),
+            '1024': int(16 * channel_multiplier * unet_narrow)
+        }
+
+        self.log_size = int(math.log(out_size, 2))
+        first_out_size = 2**(int(math.log(out_size, 2)))
+
+        self.conv_body_first = ConvLayer(3, channels[f'{first_out_size}'], 1, bias=True, activate=True)
+
+        # downsample
+        in_channels = channels[f'{first_out_size}']
+        self.conv_body_down = nn.ModuleList()
+        for i in range(self.log_size, 2, -1):
+            out_channels = channels[f'{2**(i - 1)}']
+            self.conv_body_down.append(ResBlock(in_channels, out_channels))
+            in_channels = out_channels
+
+        self.final_conv = ConvLayer(in_channels, channels['4'], 3, bias=True, activate=True)
+
+        # upsample
+        in_channels = channels['4']
+        self.conv_body_up = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            out_channels = channels[f'{2**i}']
+            self.conv_body_up.append(ResUpBlock(in_channels, out_channels))
+            in_channels = out_channels
+
+        # to RGB
+        self.toRGB = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            self.toRGB.append(EqualConv2d(channels[f'{2**i}'], 3, 1, stride=1, padding=0, bias=True, bias_init_val=0))
+
+        if different_w:
+            linear_out_channel = (int(math.log(out_size, 2)) * 2 - 2) * num_style_feat
+        else:
+            linear_out_channel = num_style_feat
+
+        self.final_linear = EqualLinear(
+            channels['4'] * 4 * 4, linear_out_channel, bias=True, bias_init_val=0, lr_mul=1, activation=None)
+
+        # the decoder: stylegan2 generator with SFT modulations
+        self.stylegan_decoder = StyleGAN2GeneratorBilinearSFT(
+            out_size=out_size,
+            num_style_feat=num_style_feat,
+            num_mlp=num_mlp,
+            channel_multiplier=channel_multiplier,
+            lr_mlp=lr_mlp,
+            narrow=narrow,
+            sft_half=sft_half)
+
+        # load pre-trained stylegan2 model if necessary
+        if decoder_load_path:
+            self.stylegan_decoder.load_state_dict(
+                torch.load(decoder_load_path, map_location=lambda storage, loc: storage)['params_ema'])
+        # fix decoder without updating params
+        if fix_decoder:
+            for _, param in self.stylegan_decoder.named_parameters():
+                param.requires_grad = False
+
+        # for SFT modulations (scale and shift)
+        self.condition_scale = nn.ModuleList()
+        self.condition_shift = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            out_channels = channels[f'{2**i}']
+            if sft_half:
+                sft_out_channels = out_channels
+            else:
+                sft_out_channels = out_channels * 2
+            self.condition_scale.append(
+                nn.Sequential(
+                    EqualConv2d(out_channels, out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0),
+                    ScaledLeakyReLU(0.2),
+                    EqualConv2d(out_channels, sft_out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=1)))
+            self.condition_shift.append(
+                nn.Sequential(
+                    EqualConv2d(out_channels, out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0),
+                    ScaledLeakyReLU(0.2),
+                    EqualConv2d(out_channels, sft_out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0)))
+
+    def forward(self, x, return_latents=False, return_rgb=True, randomize_noise=True):
+        """Forward function for GFPGANBilinear.
+
+        Args:
+            x (Tensor): Input images.
+            return_latents (bool): Whether to return style latents. Default: False.
+            return_rgb (bool): Whether return intermediate rgb images. Default: True.
+            randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
+        """
+        conditions = []
+        unet_skips = []
+        out_rgbs = []
+
+        # encoder
+        feat = self.conv_body_first(x)
+        for i in range(self.log_size - 2):
+            feat = self.conv_body_down[i](feat)
+            unet_skips.insert(0, feat)
+
+        feat = self.final_conv(feat)
+
+        # style code
+        style_code = self.final_linear(feat.view(feat.size(0), -1))
+        if self.different_w:
+            style_code = style_code.view(style_code.size(0), -1, self.num_style_feat)
+
+        # decode
+        for i in range(self.log_size - 2):
+            # add unet skip
+            feat = feat + unet_skips[i]
+            # ResUpLayer
+            feat = self.conv_body_up[i](feat)
+            # generate scale and shift for SFT layers
+            scale = self.condition_scale[i](feat)
+            conditions.append(scale.clone())
+            shift = self.condition_shift[i](feat)
+            conditions.append(shift.clone())
+            # generate rgb images
+            if return_rgb:
+                out_rgbs.append(self.toRGB[i](feat))
+
+        # decoder
+        image, _ = self.stylegan_decoder([style_code],
+                                         conditions,
+                                         return_latents=return_latents,
+                                         input_is_latent=self.input_is_latent,
+                                         randomize_noise=randomize_noise)
+
+        return image, out_rgbs

gfpgan/archs/gfpganv1_arch.py

@@ -0,0 +1,439 @@
+import math
+import random
+import torch
+from basicsr.archs.stylegan2_arch import (ConvLayer, EqualConv2d, EqualLinear, ResBlock, ScaledLeakyReLU,
+                                          StyleGAN2Generator)
+from basicsr.ops.fused_act import FusedLeakyReLU
+from basicsr.utils.registry import ARCH_REGISTRY
+from torch import nn
+from torch.nn import functional as F
+
+
+class StyleGAN2GeneratorSFT(StyleGAN2Generator):
+    """StyleGAN2 Generator with SFT modulation (Spatial Feature Transform).
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        num_style_feat (int): Channel number of style features. Default: 512.
+        num_mlp (int): Layer number of MLP style layers. Default: 8.
+        channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
+        resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. A cross production will be
+            applied to extent 1D resample kernel to 2D resample kernel. Default: (1, 3, 3, 1).
+        lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
+        narrow (float): The narrow ratio for channels. Default: 1.
+        sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
+    """
+
+    def __init__(self,
+                 out_size,
+                 num_style_feat=512,
+                 num_mlp=8,
+                 channel_multiplier=2,
+                 resample_kernel=(1, 3, 3, 1),
+                 lr_mlp=0.01,
+                 narrow=1,
+                 sft_half=False):
+        super(StyleGAN2GeneratorSFT, self).__init__(
+            out_size,
+            num_style_feat=num_style_feat,
+            num_mlp=num_mlp,
+            channel_multiplier=channel_multiplier,
+            resample_kernel=resample_kernel,
+            lr_mlp=lr_mlp,
+            narrow=narrow)
+        self.sft_half = sft_half
+
+    def forward(self,
+                styles,
+                conditions,
+                input_is_latent=False,
+                noise=None,
+                randomize_noise=True,
+                truncation=1,
+                truncation_latent=None,
+                inject_index=None,
+                return_latents=False):
+        """Forward function for StyleGAN2GeneratorSFT.
+
+        Args:
+            styles (list[Tensor]): Sample codes of styles.
+            conditions (list[Tensor]): SFT conditions to generators.
+            input_is_latent (bool): Whether input is latent style. Default: False.
+            noise (Tensor | None): Input noise or None. Default: None.
+            randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
+            truncation (float): The truncation ratio. Default: 1.
+            truncation_latent (Tensor | None): The truncation latent tensor. Default: None.
+            inject_index (int | None): The injection index for mixing noise. Default: None.
+            return_latents (bool): Whether to return style latents. Default: False.
+        """
+        # style codes -> latents with Style MLP layer
+        if not input_is_latent:
+            styles = [self.style_mlp(s) for s in styles]
+        # noises
+        if noise is None:
+            if randomize_noise:
+                noise = [None] * self.num_layers  # for each style conv layer
+            else:  # use the stored noise
+                noise = [getattr(self.noises, f'noise{i}') for i in range(self.num_layers)]
+        # style truncation
+        if truncation < 1:
+            style_truncation = []
+            for style in styles:
+                style_truncation.append(truncation_latent + truncation * (style - truncation_latent))
+            styles = style_truncation
+        # get style latents with injection
+        if len(styles) == 1:
+            inject_index = self.num_latent
+
+            if styles[0].ndim < 3:
+                # repeat latent code for all the layers
+                latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            else:  # used for encoder with different latent code for each layer
+                latent = styles[0]
+        elif len(styles) == 2:  # mixing noises
+            if inject_index is None:
+                inject_index = random.randint(1, self.num_latent - 1)
+            latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            latent2 = styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
+            latent = torch.cat([latent1, latent2], 1)
+
+        # main generation
+        out = self.constant_input(latent.shape[0])
+        out = self.style_conv1(out, latent[:, 0], noise=noise[0])
+        skip = self.to_rgb1(out, latent[:, 1])
+
+        i = 1
+        for conv1, conv2, noise1, noise2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], noise[1::2],
+                                                        noise[2::2], self.to_rgbs):
+            out = conv1(out, latent[:, i], noise=noise1)
+
+            # the conditions may have fewer levels
+            if i < len(conditions):
+                # SFT part to combine the conditions
+                if self.sft_half:  # only apply SFT to half of the channels
+                    out_same, out_sft = torch.split(out, int(out.size(1) // 2), dim=1)
+                    out_sft = out_sft * conditions[i - 1] + conditions[i]
+                    out = torch.cat([out_same, out_sft], dim=1)
+                else:  # apply SFT to all the channels
+                    out = out * conditions[i - 1] + conditions[i]
+
+            out = conv2(out, latent[:, i + 1], noise=noise2)
+            skip = to_rgb(out, latent[:, i + 2], skip)  # feature back to the rgb space
+            i += 2
+
+        image = skip
+
+        if return_latents:
+            return image, latent
+        else:
+            return image, None
+
+
+class ConvUpLayer(nn.Module):
+    """Convolutional upsampling layer. It uses bilinear upsampler + Conv.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+        kernel_size (int): Size of the convolving kernel.
+        stride (int): Stride of the convolution. Default: 1
+        padding (int): Zero-padding added to both sides of the input. Default: 0.
+        bias (bool): If ``True``, adds a learnable bias to the output. Default: ``True``.
+        bias_init_val (float): Bias initialized value. Default: 0.
+        activate (bool): Whether use activateion. Default: True.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 stride=1,
+                 padding=0,
+                 bias=True,
+                 bias_init_val=0,
+                 activate=True):
+        super(ConvUpLayer, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.stride = stride
+        self.padding = padding
+        # self.scale is used to scale the convolution weights, which is related to the common initializations.
+        self.scale = 1 / math.sqrt(in_channels * kernel_size**2)
+
+        self.weight = nn.Parameter(torch.randn(out_channels, in_channels, kernel_size, kernel_size))
+
+        if bias and not activate:
+            self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val))
+        else:
+            self.register_parameter('bias', None)
+
+        # activation
+        if activate:
+            if bias:
+                self.activation = FusedLeakyReLU(out_channels)
+            else:
+                self.activation = ScaledLeakyReLU(0.2)
+        else:
+            self.activation = None
+
+    def forward(self, x):
+        # bilinear upsample
+        out = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=False)
+        # conv
+        out = F.conv2d(
+            out,
+            self.weight * self.scale,
+            bias=self.bias,
+            stride=self.stride,
+            padding=self.padding,
+        )
+        # activation
+        if self.activation is not None:
+            out = self.activation(out)
+        return out
+
+
+class ResUpBlock(nn.Module):
+    """Residual block with upsampling.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+    """
+
+    def __init__(self, in_channels, out_channels):
+        super(ResUpBlock, self).__init__()
+
+        self.conv1 = ConvLayer(in_channels, in_channels, 3, bias=True, activate=True)
+        self.conv2 = ConvUpLayer(in_channels, out_channels, 3, stride=1, padding=1, bias=True, activate=True)
+        self.skip = ConvUpLayer(in_channels, out_channels, 1, bias=False, activate=False)
+
+    def forward(self, x):
+        out = self.conv1(x)
+        out = self.conv2(out)
+        skip = self.skip(x)
+        out = (out + skip) / math.sqrt(2)
+        return out
+
+
+@ARCH_REGISTRY.register()
+class GFPGANv1(nn.Module):
+    """The GFPGAN architecture: Unet + StyleGAN2 decoder with SFT.
+
+    Ref: GFP-GAN: Towards Real-World Blind Face Restoration with Generative Facial Prior.
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        num_style_feat (int): Channel number of style features. Default: 512.
+        channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
+        resample_kernel (list[int]): A list indicating the 1D resample kernel magnitude. A cross production will be
+            applied to extent 1D resample kernel to 2D resample kernel. Default: (1, 3, 3, 1).
+        decoder_load_path (str): The path to the pre-trained decoder model (usually, the StyleGAN2). Default: None.
+        fix_decoder (bool): Whether to fix the decoder. Default: True.
+
+        num_mlp (int): Layer number of MLP style layers. Default: 8.
+        lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
+        input_is_latent (bool): Whether input is latent style. Default: False.
+        different_w (bool): Whether to use different latent w for different layers. Default: False.
+        narrow (float): The narrow ratio for channels. Default: 1.
+        sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
+    """
+
+    def __init__(
+            self,
+            out_size,
+            num_style_feat=512,
+            channel_multiplier=1,
+            resample_kernel=(1, 3, 3, 1),
+            decoder_load_path=None,
+            fix_decoder=True,
+            # for stylegan decoder
+            num_mlp=8,
+            lr_mlp=0.01,
+            input_is_latent=False,
+            different_w=False,
+            narrow=1,
+            sft_half=False):
+
+        super(GFPGANv1, self).__init__()
+        self.input_is_latent = input_is_latent
+        self.different_w = different_w
+        self.num_style_feat = num_style_feat
+
+        unet_narrow = narrow * 0.5  # by default, use a half of input channels
+        channels = {
+            '4': int(512 * unet_narrow),
+            '8': int(512 * unet_narrow),
+            '16': int(512 * unet_narrow),
+            '32': int(512 * unet_narrow),
+            '64': int(256 * channel_multiplier * unet_narrow),
+            '128': int(128 * channel_multiplier * unet_narrow),
+            '256': int(64 * channel_multiplier * unet_narrow),
+            '512': int(32 * channel_multiplier * unet_narrow),
+            '1024': int(16 * channel_multiplier * unet_narrow)
+        }
+
+        self.log_size = int(math.log(out_size, 2))
+        first_out_size = 2**(int(math.log(out_size, 2)))
+
+        self.conv_body_first = ConvLayer(3, channels[f'{first_out_size}'], 1, bias=True, activate=True)
+
+        # downsample
+        in_channels = channels[f'{first_out_size}']
+        self.conv_body_down = nn.ModuleList()
+        for i in range(self.log_size, 2, -1):
+            out_channels = channels[f'{2**(i - 1)}']
+            self.conv_body_down.append(ResBlock(in_channels, out_channels, resample_kernel))
+            in_channels = out_channels
+
+        self.final_conv = ConvLayer(in_channels, channels['4'], 3, bias=True, activate=True)
+
+        # upsample
+        in_channels = channels['4']
+        self.conv_body_up = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            out_channels = channels[f'{2**i}']
+            self.conv_body_up.append(ResUpBlock(in_channels, out_channels))
+            in_channels = out_channels
+
+        # to RGB
+        self.toRGB = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            self.toRGB.append(EqualConv2d(channels[f'{2**i}'], 3, 1, stride=1, padding=0, bias=True, bias_init_val=0))
+
+        if different_w:
+            linear_out_channel = (int(math.log(out_size, 2)) * 2 - 2) * num_style_feat
+        else:
+            linear_out_channel = num_style_feat
+
+        self.final_linear = EqualLinear(
+            channels['4'] * 4 * 4, linear_out_channel, bias=True, bias_init_val=0, lr_mul=1, activation=None)
+
+        # the decoder: stylegan2 generator with SFT modulations
+        self.stylegan_decoder = StyleGAN2GeneratorSFT(
+            out_size=out_size,
+            num_style_feat=num_style_feat,
+            num_mlp=num_mlp,
+            channel_multiplier=channel_multiplier,
+            resample_kernel=resample_kernel,
+            lr_mlp=lr_mlp,
+            narrow=narrow,
+            sft_half=sft_half)
+
+        # load pre-trained stylegan2 model if necessary
+        if decoder_load_path:
+            self.stylegan_decoder.load_state_dict(
+                torch.load(decoder_load_path, map_location=lambda storage, loc: storage)['params_ema'])
+        # fix decoder without updating params
+        if fix_decoder:
+            for _, param in self.stylegan_decoder.named_parameters():
+                param.requires_grad = False
+
+        # for SFT modulations (scale and shift)
+        self.condition_scale = nn.ModuleList()
+        self.condition_shift = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            out_channels = channels[f'{2**i}']
+            if sft_half:
+                sft_out_channels = out_channels
+            else:
+                sft_out_channels = out_channels * 2
+            self.condition_scale.append(
+                nn.Sequential(
+                    EqualConv2d(out_channels, out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0),
+                    ScaledLeakyReLU(0.2),
+                    EqualConv2d(out_channels, sft_out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=1)))
+            self.condition_shift.append(
+                nn.Sequential(
+                    EqualConv2d(out_channels, out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0),
+                    ScaledLeakyReLU(0.2),
+                    EqualConv2d(out_channels, sft_out_channels, 3, stride=1, padding=1, bias=True, bias_init_val=0)))
+
+    def forward(self, x, return_latents=False, return_rgb=True, randomize_noise=True, **kwargs):
+        """Forward function for GFPGANv1.
+
+        Args:
+            x (Tensor): Input images.
+            return_latents (bool): Whether to return style latents. Default: False.
+            return_rgb (bool): Whether return intermediate rgb images. Default: True.
+            randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
+        """
+        conditions = []
+        unet_skips = []
+        out_rgbs = []
+
+        # encoder
+        feat = self.conv_body_first(x)
+        for i in range(self.log_size - 2):
+            feat = self.conv_body_down[i](feat)
+            unet_skips.insert(0, feat)
+
+        feat = self.final_conv(feat)
+
+        # style code
+        style_code = self.final_linear(feat.view(feat.size(0), -1))
+        if self.different_w:
+            style_code = style_code.view(style_code.size(0), -1, self.num_style_feat)
+
+        # decode
+        for i in range(self.log_size - 2):
+            # add unet skip
+            feat = feat + unet_skips[i]
+            # ResUpLayer
+            feat = self.conv_body_up[i](feat)
+            # generate scale and shift for SFT layers
+            scale = self.condition_scale[i](feat)
+            conditions.append(scale.clone())
+            shift = self.condition_shift[i](feat)
+            conditions.append(shift.clone())
+            # generate rgb images
+            if return_rgb:
+                out_rgbs.append(self.toRGB[i](feat))
+
+        # decoder
+        image, _ = self.stylegan_decoder([style_code],
+                                         conditions,
+                                         return_latents=return_latents,
+                                         input_is_latent=self.input_is_latent,
+                                         randomize_noise=randomize_noise)
+
+        return image, out_rgbs
+
+
+@ARCH_REGISTRY.register()
+class FacialComponentDiscriminator(nn.Module):
+    """Facial component (eyes, mouth, noise) discriminator used in GFPGAN.
+    """
+
+    def __init__(self):
+        super(FacialComponentDiscriminator, self).__init__()
+        # It now uses a VGG-style architectrue with fixed model size
+        self.conv1 = ConvLayer(3, 64, 3, downsample=False, resample_kernel=(1, 3, 3, 1), bias=True, activate=True)
+        self.conv2 = ConvLayer(64, 128, 3, downsample=True, resample_kernel=(1, 3, 3, 1), bias=True, activate=True)
+        self.conv3 = ConvLayer(128, 128, 3, downsample=False, resample_kernel=(1, 3, 3, 1), bias=True, activate=True)
+        self.conv4 = ConvLayer(128, 256, 3, downsample=True, resample_kernel=(1, 3, 3, 1), bias=True, activate=True)
+        self.conv5 = ConvLayer(256, 256, 3, downsample=False, resample_kernel=(1, 3, 3, 1), bias=True, activate=True)
+        self.final_conv = ConvLayer(256, 1, 3, bias=True, activate=False)
+
+    def forward(self, x, return_feats=False, **kwargs):
+        """Forward function for FacialComponentDiscriminator.
+
+        Args:
+            x (Tensor): Input images.
+            return_feats (bool): Whether to return intermediate features. Default: False.
+        """
+        feat = self.conv1(x)
+        feat = self.conv3(self.conv2(feat))
+        rlt_feats = []
+        if return_feats:
+            rlt_feats.append(feat.clone())
+        feat = self.conv5(self.conv4(feat))
+        if return_feats:
+            rlt_feats.append(feat.clone())
+        out = self.final_conv(feat)
+
+        if return_feats:
+            return out, rlt_feats
+        else:
+            return out, None

gfpgan/archs/gfpganv1_clean_arch.py

@@ -0,0 +1,324 @@
+import math
+import random
+import torch
+from basicsr.utils.registry import ARCH_REGISTRY
+from torch import nn
+from torch.nn import functional as F
+
+from .stylegan2_clean_arch import StyleGAN2GeneratorClean
+
+
+class StyleGAN2GeneratorCSFT(StyleGAN2GeneratorClean):
+    """StyleGAN2 Generator with SFT modulation (Spatial Feature Transform).
+
+    It is the clean version without custom compiled CUDA extensions used in StyleGAN2.
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        num_style_feat (int): Channel number of style features. Default: 512.
+        num_mlp (int): Layer number of MLP style layers. Default: 8.
+        channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
+        narrow (float): The narrow ratio for channels. Default: 1.
+        sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
+    """
+
+    def __init__(self, out_size, num_style_feat=512, num_mlp=8, channel_multiplier=2, narrow=1, sft_half=False):
+        super(StyleGAN2GeneratorCSFT, self).__init__(
+            out_size,
+            num_style_feat=num_style_feat,
+            num_mlp=num_mlp,
+            channel_multiplier=channel_multiplier,
+            narrow=narrow)
+        self.sft_half = sft_half
+
+    def forward(self,
+                styles,
+                conditions,
+                input_is_latent=False,
+                noise=None,
+                randomize_noise=True,
+                truncation=1,
+                truncation_latent=None,
+                inject_index=None,
+                return_latents=False):
+        """Forward function for StyleGAN2GeneratorCSFT.
+
+        Args:
+            styles (list[Tensor]): Sample codes of styles.
+            conditions (list[Tensor]): SFT conditions to generators.
+            input_is_latent (bool): Whether input is latent style. Default: False.
+            noise (Tensor | None): Input noise or None. Default: None.
+            randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
+            truncation (float): The truncation ratio. Default: 1.
+            truncation_latent (Tensor | None): The truncation latent tensor. Default: None.
+            inject_index (int | None): The injection index for mixing noise. Default: None.
+            return_latents (bool): Whether to return style latents. Default: False.
+        """
+        # style codes -> latents with Style MLP layer
+        if not input_is_latent:
+            styles = [self.style_mlp(s) for s in styles]
+        # noises
+        if noise is None:
+            if randomize_noise:
+                noise = [None] * self.num_layers  # for each style conv layer
+            else:  # use the stored noise
+                noise = [getattr(self.noises, f'noise{i}') for i in range(self.num_layers)]
+        # style truncation
+        if truncation < 1:
+            style_truncation = []
+            for style in styles:
+                style_truncation.append(truncation_latent + truncation * (style - truncation_latent))
+            styles = style_truncation
+        # get style latents with injection
+        if len(styles) == 1:
+            inject_index = self.num_latent
+
+            if styles[0].ndim < 3:
+                # repeat latent code for all the layers
+                latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            else:  # used for encoder with different latent code for each layer
+                latent = styles[0]
+        elif len(styles) == 2:  # mixing noises
+            if inject_index is None:
+                inject_index = random.randint(1, self.num_latent - 1)
+            latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            latent2 = styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
+            latent = torch.cat([latent1, latent2], 1)
+
+        # main generation
+        out = self.constant_input(latent.shape[0])
+        out = self.style_conv1(out, latent[:, 0], noise=noise[0])
+        skip = self.to_rgb1(out, latent[:, 1])
+
+        i = 1
+        for conv1, conv2, noise1, noise2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], noise[1::2],
+                                                        noise[2::2], self.to_rgbs):
+            out = conv1(out, latent[:, i], noise=noise1)
+
+            # the conditions may have fewer levels
+            if i < len(conditions):
+                # SFT part to combine the conditions
+                if self.sft_half:  # only apply SFT to half of the channels
+                    out_same, out_sft = torch.split(out, int(out.size(1) // 2), dim=1)
+                    out_sft = out_sft * conditions[i - 1] + conditions[i]
+                    out = torch.cat([out_same, out_sft], dim=1)
+                else:  # apply SFT to all the channels
+                    out = out * conditions[i - 1] + conditions[i]
+
+            out = conv2(out, latent[:, i + 1], noise=noise2)
+            skip = to_rgb(out, latent[:, i + 2], skip)  # feature back to the rgb space
+            i += 2
+
+        image = skip
+
+        if return_latents:
+            return image, latent
+        else:
+            return image, None
+
+
+class ResBlock(nn.Module):
+    """Residual block with bilinear upsampling/downsampling.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+        mode (str): Upsampling/downsampling mode. Options: down | up. Default: down.
+    """
+
+    def __init__(self, in_channels, out_channels, mode='down'):
+        super(ResBlock, self).__init__()
+
+        self.conv1 = nn.Conv2d(in_channels, in_channels, 3, 1, 1)
+        self.conv2 = nn.Conv2d(in_channels, out_channels, 3, 1, 1)
+        self.skip = nn.Conv2d(in_channels, out_channels, 1, bias=False)
+        if mode == 'down':
+            self.scale_factor = 0.5
+        elif mode == 'up':
+            self.scale_factor = 2
+
+    def forward(self, x):
+        out = F.leaky_relu_(self.conv1(x), negative_slope=0.2)
+        # upsample/downsample
+        out = F.interpolate(out, scale_factor=self.scale_factor, mode='bilinear', align_corners=False)
+        out = F.leaky_relu_(self.conv2(out), negative_slope=0.2)
+        # skip
+        x = F.interpolate(x, scale_factor=self.scale_factor, mode='bilinear', align_corners=False)
+        skip = self.skip(x)
+        out = out + skip
+        return out
+
+
+@ARCH_REGISTRY.register()
+class GFPGANv1Clean(nn.Module):
+    """The GFPGAN architecture: Unet + StyleGAN2 decoder with SFT.
+
+    It is the clean version without custom compiled CUDA extensions used in StyleGAN2.
+
+    Ref: GFP-GAN: Towards Real-World Blind Face Restoration with Generative Facial Prior.
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        num_style_feat (int): Channel number of style features. Default: 512.
+        channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
+        decoder_load_path (str): The path to the pre-trained decoder model (usually, the StyleGAN2). Default: None.
+        fix_decoder (bool): Whether to fix the decoder. Default: True.
+
+        num_mlp (int): Layer number of MLP style layers. Default: 8.
+        input_is_latent (bool): Whether input is latent style. Default: False.
+        different_w (bool): Whether to use different latent w for different layers. Default: False.
+        narrow (float): The narrow ratio for channels. Default: 1.
+        sft_half (bool): Whether to apply SFT on half of the input channels. Default: False.
+    """
+
+    def __init__(
+            self,
+            out_size,
+            num_style_feat=512,
+            channel_multiplier=1,
+            decoder_load_path=None,
+            fix_decoder=True,
+            # for stylegan decoder
+            num_mlp=8,
+            input_is_latent=False,
+            different_w=False,
+            narrow=1,
+            sft_half=False):
+
+        super(GFPGANv1Clean, self).__init__()
+        self.input_is_latent = input_is_latent
+        self.different_w = different_w
+        self.num_style_feat = num_style_feat
+
+        unet_narrow = narrow * 0.5  # by default, use a half of input channels
+        channels = {
+            '4': int(512 * unet_narrow),
+            '8': int(512 * unet_narrow),
+            '16': int(512 * unet_narrow),
+            '32': int(512 * unet_narrow),
+            '64': int(256 * channel_multiplier * unet_narrow),
+            '128': int(128 * channel_multiplier * unet_narrow),
+            '256': int(64 * channel_multiplier * unet_narrow),
+            '512': int(32 * channel_multiplier * unet_narrow),
+            '1024': int(16 * channel_multiplier * unet_narrow)
+        }
+
+        self.log_size = int(math.log(out_size, 2))
+        first_out_size = 2**(int(math.log(out_size, 2)))
+
+        self.conv_body_first = nn.Conv2d(3, channels[f'{first_out_size}'], 1)
+
+        # downsample
+        in_channels = channels[f'{first_out_size}']
+        self.conv_body_down = nn.ModuleList()
+        for i in range(self.log_size, 2, -1):
+            out_channels = channels[f'{2**(i - 1)}']
+            self.conv_body_down.append(ResBlock(in_channels, out_channels, mode='down'))
+            in_channels = out_channels
+
+        self.final_conv = nn.Conv2d(in_channels, channels['4'], 3, 1, 1)
+
+        # upsample
+        in_channels = channels['4']
+        self.conv_body_up = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            out_channels = channels[f'{2**i}']
+            self.conv_body_up.append(ResBlock(in_channels, out_channels, mode='up'))
+            in_channels = out_channels
+
+        # to RGB
+        self.toRGB = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            self.toRGB.append(nn.Conv2d(channels[f'{2**i}'], 3, 1))
+
+        if different_w:
+            linear_out_channel = (int(math.log(out_size, 2)) * 2 - 2) * num_style_feat
+        else:
+            linear_out_channel = num_style_feat
+
+        self.final_linear = nn.Linear(channels['4'] * 4 * 4, linear_out_channel)
+
+        # the decoder: stylegan2 generator with SFT modulations
+        self.stylegan_decoder = StyleGAN2GeneratorCSFT(
+            out_size=out_size,
+            num_style_feat=num_style_feat,
+            num_mlp=num_mlp,
+            channel_multiplier=channel_multiplier,
+            narrow=narrow,
+            sft_half=sft_half)
+
+        # load pre-trained stylegan2 model if necessary
+        if decoder_load_path:
+            self.stylegan_decoder.load_state_dict(
+                torch.load(decoder_load_path, map_location=lambda storage, loc: storage)['params_ema'])
+        # fix decoder without updating params
+        if fix_decoder:
+            for _, param in self.stylegan_decoder.named_parameters():
+                param.requires_grad = False
+
+        # for SFT modulations (scale and shift)
+        self.condition_scale = nn.ModuleList()
+        self.condition_shift = nn.ModuleList()
+        for i in range(3, self.log_size + 1):
+            out_channels = channels[f'{2**i}']
+            if sft_half:
+                sft_out_channels = out_channels
+            else:
+                sft_out_channels = out_channels * 2
+            self.condition_scale.append(
+                nn.Sequential(
+                    nn.Conv2d(out_channels, out_channels, 3, 1, 1), nn.LeakyReLU(0.2, True),
+                    nn.Conv2d(out_channels, sft_out_channels, 3, 1, 1)))
+            self.condition_shift.append(
+                nn.Sequential(
+                    nn.Conv2d(out_channels, out_channels, 3, 1, 1), nn.LeakyReLU(0.2, True),
+                    nn.Conv2d(out_channels, sft_out_channels, 3, 1, 1)))
+
+    def forward(self, x, return_latents=False, return_rgb=True, randomize_noise=True, **kwargs):
+        """Forward function for GFPGANv1Clean.
+
+        Args:
+            x (Tensor): Input images.
+            return_latents (bool): Whether to return style latents. Default: False.
+            return_rgb (bool): Whether return intermediate rgb images. Default: True.
+            randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
+        """
+        conditions = []
+        unet_skips = []
+        out_rgbs = []
+
+        # encoder
+        feat = F.leaky_relu_(self.conv_body_first(x), negative_slope=0.2)
+        for i in range(self.log_size - 2):
+            feat = self.conv_body_down[i](feat)
+            unet_skips.insert(0, feat)
+        feat = F.leaky_relu_(self.final_conv(feat), negative_slope=0.2)
+
+        # style code
+        style_code = self.final_linear(feat.view(feat.size(0), -1))
+        if self.different_w:
+            style_code = style_code.view(style_code.size(0), -1, self.num_style_feat)
+
+        # decode
+        for i in range(self.log_size - 2):
+            # add unet skip
+            feat = feat + unet_skips[i]
+            # ResUpLayer
+            feat = self.conv_body_up[i](feat)
+            # generate scale and shift for SFT layers
+            scale = self.condition_scale[i](feat)
+            conditions.append(scale.clone())
+            shift = self.condition_shift[i](feat)
+            conditions.append(shift.clone())
+            # generate rgb images
+            if return_rgb:
+                out_rgbs.append(self.toRGB[i](feat))
+
+        # decoder
+        image, _ = self.stylegan_decoder([style_code],
+                                         conditions,
+                                         return_latents=return_latents,
+                                         input_is_latent=self.input_is_latent,
+                                         randomize_noise=randomize_noise)
+
+        return image, out_rgbs

gfpgan/archs/restoreformer_arch.py

@@ -0,0 +1,658 @@
+"""Modified from https://github.com/wzhouxiff/RestoreFormer
+"""
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class VectorQuantizer(nn.Module):
+    """
+    see https://github.com/MishaLaskin/vqvae/blob/d761a999e2267766400dc646d82d3ac3657771d4/models/quantizer.py
+    ____________________________________________
+    Discretization bottleneck part of the VQ-VAE.
+    Inputs:
+    - n_e : number of embeddings
+    - e_dim : dimension of embedding
+    - beta : commitment cost used in loss term, beta * ||z_e(x)-sg[e]||^2
+    _____________________________________________
+    """
+
+    def __init__(self, n_e, e_dim, beta):
+        super(VectorQuantizer, self).__init__()
+        self.n_e = n_e
+        self.e_dim = e_dim
+        self.beta = beta
+
+        self.embedding = nn.Embedding(self.n_e, self.e_dim)
+        self.embedding.weight.data.uniform_(-1.0 / self.n_e, 1.0 / self.n_e)
+
+    def forward(self, z):
+        """
+        Inputs the output of the encoder network z and maps it to a discrete
+        one-hot vector that is the index of the closest embedding vector e_j
+        z (continuous) -> z_q (discrete)
+        z.shape = (batch, channel, height, width)
+        quantization pipeline:
+            1. get encoder input (B,C,H,W)
+            2. flatten input to (B*H*W,C)
+        """
+        # reshape z -> (batch, height, width, channel) and flatten
+        z = z.permute(0, 2, 3, 1).contiguous()
+        z_flattened = z.view(-1, self.e_dim)
+        # distances from z to embeddings e_j (z - e)^2 = z^2 + e^2 - 2 e * z
+
+        d = torch.sum(z_flattened ** 2, dim=1, keepdim=True) + \
+            torch.sum(self.embedding.weight**2, dim=1) - 2 * \
+            torch.matmul(z_flattened, self.embedding.weight.t())
+
+        # could possible replace this here
+        # #\start...
+        # find closest encodings
+
+        min_value, min_encoding_indices = torch.min(d, dim=1)
+
+        min_encoding_indices = min_encoding_indices.unsqueeze(1)
+
+        min_encodings = torch.zeros(min_encoding_indices.shape[0], self.n_e).to(z)
+        min_encodings.scatter_(1, min_encoding_indices, 1)
+
+        # dtype min encodings: torch.float32
+        # min_encodings shape: torch.Size([2048, 512])
+        # min_encoding_indices.shape: torch.Size([2048, 1])
+
+        # get quantized latent vectors
+        z_q = torch.matmul(min_encodings, self.embedding.weight).view(z.shape)
+        # .........\end
+
+        # with:
+        # .........\start
+        # min_encoding_indices = torch.argmin(d, dim=1)
+        # z_q = self.embedding(min_encoding_indices)
+        # ......\end......... (TODO)
+
+        # compute loss for embedding
+        loss = torch.mean((z_q.detach() - z)**2) + self.beta * torch.mean((z_q - z.detach())**2)
+
+        # preserve gradients
+        z_q = z + (z_q - z).detach()
+
+        # perplexity
+
+        e_mean = torch.mean(min_encodings, dim=0)
+        perplexity = torch.exp(-torch.sum(e_mean * torch.log(e_mean + 1e-10)))
+
+        # reshape back to match original input shape
+        z_q = z_q.permute(0, 3, 1, 2).contiguous()
+
+        return z_q, loss, (perplexity, min_encodings, min_encoding_indices, d)
+
+    def get_codebook_entry(self, indices, shape):
+        # shape specifying (batch, height, width, channel)
+        # TODO: check for more easy handling with nn.Embedding
+        min_encodings = torch.zeros(indices.shape[0], self.n_e).to(indices)
+        min_encodings.scatter_(1, indices[:, None], 1)
+
+        # get quantized latent vectors
+        z_q = torch.matmul(min_encodings.float(), self.embedding.weight)
+
+        if shape is not None:
+            z_q = z_q.view(shape)
+
+            # reshape back to match original input shape
+            z_q = z_q.permute(0, 3, 1, 2).contiguous()
+
+        return z_q
+
+
+# pytorch_diffusion + derived encoder decoder
+def nonlinearity(x):
+    # swish
+    return x * torch.sigmoid(x)
+
+
+def Normalize(in_channels):
+    return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True)
+
+
+class Upsample(nn.Module):
+
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode='nearest')
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=2, padding=0)
+
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0, 1, 0, 1)
+            x = torch.nn.functional.pad(x, pad, mode='constant', value=0)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+
+class ResnetBlock(nn.Module):
+
+    def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False, dropout, temb_channels=512):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels)
+        self.conv1 = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+        if temb_channels > 0:
+            self.temb_proj = torch.nn.Linear(temb_channels, out_channels)
+        self.norm2 = Normalize(out_channels)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = torch.nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1)
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = torch.nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=1, padding=1)
+            else:
+                self.nin_shortcut = torch.nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None]
+
+        h = self.norm2(h)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x + h
+
+
+class MultiHeadAttnBlock(nn.Module):
+
+    def __init__(self, in_channels, head_size=1):
+        super().__init__()
+        self.in_channels = in_channels
+        self.head_size = head_size
+        self.att_size = in_channels // head_size
+        assert (in_channels % head_size == 0), 'The size of head should be divided by the number of channels.'
+
+        self.norm1 = Normalize(in_channels)
+        self.norm2 = Normalize(in_channels)
+
+        self.q = torch.nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.k = torch.nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.v = torch.nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.proj_out = torch.nn.Conv2d(in_channels, in_channels, kernel_size=1, stride=1, padding=0)
+        self.num = 0
+
+    def forward(self, x, y=None):
+        h_ = x
+        h_ = self.norm1(h_)
+        if y is None:
+            y = h_
+        else:
+            y = self.norm2(y)
+
+        q = self.q(y)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b, c, h, w = q.shape
+        q = q.reshape(b, self.head_size, self.att_size, h * w)
+        q = q.permute(0, 3, 1, 2)  # b, hw, head, att
+
+        k = k.reshape(b, self.head_size, self.att_size, h * w)
+        k = k.permute(0, 3, 1, 2)
+
+        v = v.reshape(b, self.head_size, self.att_size, h * w)
+        v = v.permute(0, 3, 1, 2)
+
+        q = q.transpose(1, 2)
+        v = v.transpose(1, 2)
+        k = k.transpose(1, 2).transpose(2, 3)
+
+        scale = int(self.att_size)**(-0.5)
+        q.mul_(scale)
+        w_ = torch.matmul(q, k)
+        w_ = F.softmax(w_, dim=3)
+
+        w_ = w_.matmul(v)
+
+        w_ = w_.transpose(1, 2).contiguous()  # [b, h*w, head, att]
+        w_ = w_.view(b, h, w, -1)
+        w_ = w_.permute(0, 3, 1, 2)
+
+        w_ = self.proj_out(w_)
+
+        return x + w_
+
+
+class MultiHeadEncoder(nn.Module):
+
+    def __init__(self,
+                 ch,
+                 out_ch,
+                 ch_mult=(1, 2, 4, 8),
+                 num_res_blocks=2,
+                 attn_resolutions=(16, ),
+                 dropout=0.0,
+                 resamp_with_conv=True,
+                 in_channels=3,
+                 resolution=512,
+                 z_channels=256,
+                 double_z=True,
+                 enable_mid=True,
+                 head_size=1,
+                 **ignore_kwargs):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.enable_mid = enable_mid
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(in_channels, self.ch, kernel_size=3, stride=1, padding=1)
+
+        curr_res = resolution
+        in_ch_mult = (1, ) + tuple(ch_mult)
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch * in_ch_mult[i_level]
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(MultiHeadAttnBlock(block_in, head_size))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions - 1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        if self.enable_mid:
+            self.mid = nn.Module()
+            self.mid.block_1 = ResnetBlock(
+                in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout)
+            self.mid.attn_1 = MultiHeadAttnBlock(block_in, head_size)
+            self.mid.block_2 = ResnetBlock(
+                in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout)
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(
+            block_in, 2 * z_channels if double_z else z_channels, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, x):
+        hs = {}
+        # timestep embedding
+        temb = None
+
+        # downsampling
+        h = self.conv_in(x)
+        hs['in'] = h
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](h, temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+
+            if i_level != self.num_resolutions - 1:
+                # hs.append(h)
+                hs['block_' + str(i_level)] = h
+                h = self.down[i_level].downsample(h)
+
+        # middle
+        # h = hs[-1]
+        if self.enable_mid:
+            h = self.mid.block_1(h, temb)
+            hs['block_' + str(i_level) + '_atten'] = h
+            h = self.mid.attn_1(h)
+            h = self.mid.block_2(h, temb)
+            hs['mid_atten'] = h
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        # hs.append(h)
+        hs['out'] = h
+
+        return hs
+
+
+class MultiHeadDecoder(nn.Module):
+
+    def __init__(self,
+                 ch,
+                 out_ch,
+                 ch_mult=(1, 2, 4, 8),
+                 num_res_blocks=2,
+                 attn_resolutions=(16, ),
+                 dropout=0.0,
+                 resamp_with_conv=True,
+                 in_channels=3,
+                 resolution=512,
+                 z_channels=256,
+                 give_pre_end=False,
+                 enable_mid=True,
+                 head_size=1,
+                 **ignorekwargs):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+        self.enable_mid = enable_mid
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        block_in = ch * ch_mult[self.num_resolutions - 1]
+        curr_res = resolution // 2**(self.num_resolutions - 1)
+        self.z_shape = (1, z_channels, curr_res, curr_res)
+        print('Working with z of shape {} = {} dimensions.'.format(self.z_shape, np.prod(self.z_shape)))
+
+        # z to block_in
+        self.conv_in = torch.nn.Conv2d(z_channels, block_in, kernel_size=3, stride=1, padding=1)
+
+        # middle
+        if self.enable_mid:
+            self.mid = nn.Module()
+            self.mid.block_1 = ResnetBlock(
+                in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout)
+            self.mid.attn_1 = MultiHeadAttnBlock(block_in, head_size)
+            self.mid.block_2 = ResnetBlock(
+                in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(MultiHeadAttnBlock(block_in, head_size))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in, out_ch, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, z):
+        # assert z.shape[1:] == self.z_shape[1:]
+        self.last_z_shape = z.shape
+
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        if self.enable_mid:
+            h = self.mid.block_1(h, temb)
+            h = self.mid.attn_1(h)
+            h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h, temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class MultiHeadDecoderTransformer(nn.Module):
+
+    def __init__(self,
+                 ch,
+                 out_ch,
+                 ch_mult=(1, 2, 4, 8),
+                 num_res_blocks=2,
+                 attn_resolutions=(16, ),
+                 dropout=0.0,
+                 resamp_with_conv=True,
+                 in_channels=3,
+                 resolution=512,
+                 z_channels=256,
+                 give_pre_end=False,
+                 enable_mid=True,
+                 head_size=1,
+                 **ignorekwargs):
+        super().__init__()
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+        self.enable_mid = enable_mid
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        block_in = ch * ch_mult[self.num_resolutions - 1]
+        curr_res = resolution // 2**(self.num_resolutions - 1)
+        self.z_shape = (1, z_channels, curr_res, curr_res)
+        print('Working with z of shape {} = {} dimensions.'.format(self.z_shape, np.prod(self.z_shape)))
+
+        # z to block_in
+        self.conv_in = torch.nn.Conv2d(z_channels, block_in, kernel_size=3, stride=1, padding=1)
+
+        # middle
+        if self.enable_mid:
+            self.mid = nn.Module()
+            self.mid.block_1 = ResnetBlock(
+                in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout)
+            self.mid.attn_1 = MultiHeadAttnBlock(block_in, head_size)
+            self.mid.block_2 = ResnetBlock(
+                in_channels=block_in, out_channels=block_in, temb_channels=self.temb_ch, dropout=dropout)
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in, out_channels=block_out, temb_channels=self.temb_ch, dropout=dropout))
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(MultiHeadAttnBlock(block_in, head_size))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(block_in, out_ch, kernel_size=3, stride=1, padding=1)
+
+    def forward(self, z, hs):
+        # assert z.shape[1:] == self.z_shape[1:]
+        # self.last_z_shape = z.shape
+
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        if self.enable_mid:
+            h = self.mid.block_1(h, temb)
+            h = self.mid.attn_1(h, hs['mid_atten'])
+            h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h, temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h, hs['block_' + str(i_level) + '_atten'])
+                    # hfeature = h.clone()
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class RestoreFormer(nn.Module):
+
+    def __init__(self,
+                 n_embed=1024,
+                 embed_dim=256,
+                 ch=64,
+                 out_ch=3,
+                 ch_mult=(1, 2, 2, 4, 4, 8),
+                 num_res_blocks=2,
+                 attn_resolutions=(16, ),
+                 dropout=0.0,
+                 in_channels=3,
+                 resolution=512,
+                 z_channels=256,
+                 double_z=False,
+                 enable_mid=True,
+                 fix_decoder=False,
+                 fix_codebook=True,
+                 fix_encoder=False,
+                 head_size=8):
+        super(RestoreFormer, self).__init__()
+
+        self.encoder = MultiHeadEncoder(
+            ch=ch,
+            out_ch=out_ch,
+            ch_mult=ch_mult,
+            num_res_blocks=num_res_blocks,
+            attn_resolutions=attn_resolutions,
+            dropout=dropout,
+            in_channels=in_channels,
+            resolution=resolution,
+            z_channels=z_channels,
+            double_z=double_z,
+            enable_mid=enable_mid,
+            head_size=head_size)
+        self.decoder = MultiHeadDecoderTransformer(
+            ch=ch,
+            out_ch=out_ch,
+            ch_mult=ch_mult,
+            num_res_blocks=num_res_blocks,
+            attn_resolutions=attn_resolutions,
+            dropout=dropout,
+            in_channels=in_channels,
+            resolution=resolution,
+            z_channels=z_channels,
+            enable_mid=enable_mid,
+            head_size=head_size)
+
+        self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25)
+
+        self.quant_conv = torch.nn.Conv2d(z_channels, embed_dim, 1)
+        self.post_quant_conv = torch.nn.Conv2d(embed_dim, z_channels, 1)
+
+        if fix_decoder:
+            for _, param in self.decoder.named_parameters():
+                param.requires_grad = False
+            for _, param in self.post_quant_conv.named_parameters():
+                param.requires_grad = False
+            for _, param in self.quantize.named_parameters():
+                param.requires_grad = False
+        elif fix_codebook:
+            for _, param in self.quantize.named_parameters():
+                param.requires_grad = False
+
+        if fix_encoder:
+            for _, param in self.encoder.named_parameters():
+                param.requires_grad = False
+
+    def encode(self, x):
+
+        hs = self.encoder(x)
+        h = self.quant_conv(hs['out'])
+        quant, emb_loss, info = self.quantize(h)
+        return quant, emb_loss, info, hs
+
+    def decode(self, quant, hs):
+        quant = self.post_quant_conv(quant)
+        dec = self.decoder(quant, hs)
+
+        return dec
+
+    def forward(self, input, **kwargs):
+        quant, diff, info, hs = self.encode(input)
+        dec = self.decode(quant, hs)
+
+        return dec, None

gfpgan/archs/stylegan2_bilinear_arch.py

@@ -0,0 +1,613 @@
+import math
+import random
+import torch
+from basicsr.ops.fused_act import FusedLeakyReLU, fused_leaky_relu
+from basicsr.utils.registry import ARCH_REGISTRY
+from torch import nn
+from torch.nn import functional as F
+
+
+class NormStyleCode(nn.Module):
+
+    def forward(self, x):
+        """Normalize the style codes.
+
+        Args:
+            x (Tensor): Style codes with shape (b, c).
+
+        Returns:
+            Tensor: Normalized tensor.
+        """
+        return x * torch.rsqrt(torch.mean(x**2, dim=1, keepdim=True) + 1e-8)
+
+
+class EqualLinear(nn.Module):
+    """Equalized Linear as StyleGAN2.
+
+    Args:
+        in_channels (int): Size of each sample.
+        out_channels (int): Size of each output sample.
+        bias (bool): If set to ``False``, the layer will not learn an additive
+            bias. Default: ``True``.
+        bias_init_val (float): Bias initialized value. Default: 0.
+        lr_mul (float): Learning rate multiplier. Default: 1.
+        activation (None | str): The activation after ``linear`` operation.
+            Supported: 'fused_lrelu', None. Default: None.
+    """
+
+    def __init__(self, in_channels, out_channels, bias=True, bias_init_val=0, lr_mul=1, activation=None):
+        super(EqualLinear, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.lr_mul = lr_mul
+        self.activation = activation
+        if self.activation not in ['fused_lrelu', None]:
+            raise ValueError(f'Wrong activation value in EqualLinear: {activation}'
+                             "Supported ones are: ['fused_lrelu', None].")
+        self.scale = (1 / math.sqrt(in_channels)) * lr_mul
+
+        self.weight = nn.Parameter(torch.randn(out_channels, in_channels).div_(lr_mul))
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val))
+        else:
+            self.register_parameter('bias', None)
+
+    def forward(self, x):
+        if self.bias is None:
+            bias = None
+        else:
+            bias = self.bias * self.lr_mul
+        if self.activation == 'fused_lrelu':
+            out = F.linear(x, self.weight * self.scale)
+            out = fused_leaky_relu(out, bias)
+        else:
+            out = F.linear(x, self.weight * self.scale, bias=bias)
+        return out
+
+    def __repr__(self):
+        return (f'{self.__class__.__name__}(in_channels={self.in_channels}, '
+                f'out_channels={self.out_channels}, bias={self.bias is not None})')
+
+
+class ModulatedConv2d(nn.Module):
+    """Modulated Conv2d used in StyleGAN2.
+
+    There is no bias in ModulatedConv2d.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+        kernel_size (int): Size of the convolving kernel.
+        num_style_feat (int): Channel number of style features.
+        demodulate (bool): Whether to demodulate in the conv layer.
+            Default: True.
+        sample_mode (str | None): Indicating 'upsample', 'downsample' or None.
+            Default: None.
+        eps (float): A value added to the denominator for numerical stability.
+            Default: 1e-8.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 num_style_feat,
+                 demodulate=True,
+                 sample_mode=None,
+                 eps=1e-8,
+                 interpolation_mode='bilinear'):
+        super(ModulatedConv2d, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.demodulate = demodulate
+        self.sample_mode = sample_mode
+        self.eps = eps
+        self.interpolation_mode = interpolation_mode
+        if self.interpolation_mode == 'nearest':
+            self.align_corners = None
+        else:
+            self.align_corners = False
+
+        self.scale = 1 / math.sqrt(in_channels * kernel_size**2)
+        # modulation inside each modulated conv
+        self.modulation = EqualLinear(
+            num_style_feat, in_channels, bias=True, bias_init_val=1, lr_mul=1, activation=None)
+
+        self.weight = nn.Parameter(torch.randn(1, out_channels, in_channels, kernel_size, kernel_size))
+        self.padding = kernel_size // 2
+
+    def forward(self, x, style):
+        """Forward function.
+
+        Args:
+            x (Tensor): Tensor with shape (b, c, h, w).
+            style (Tensor): Tensor with shape (b, num_style_feat).
+
+        Returns:
+            Tensor: Modulated tensor after convolution.
+        """
+        b, c, h, w = x.shape  # c = c_in
+        # weight modulation
+        style = self.modulation(style).view(b, 1, c, 1, 1)
+        # self.weight: (1, c_out, c_in, k, k); style: (b, 1, c, 1, 1)
+        weight = self.scale * self.weight * style  # (b, c_out, c_in, k, k)
+
+        if self.demodulate:
+            demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + self.eps)
+            weight = weight * demod.view(b, self.out_channels, 1, 1, 1)
+
+        weight = weight.view(b * self.out_channels, c, self.kernel_size, self.kernel_size)
+
+        if self.sample_mode == 'upsample':
+            x = F.interpolate(x, scale_factor=2, mode=self.interpolation_mode, align_corners=self.align_corners)
+        elif self.sample_mode == 'downsample':
+            x = F.interpolate(x, scale_factor=0.5, mode=self.interpolation_mode, align_corners=self.align_corners)
+
+        b, c, h, w = x.shape
+        x = x.view(1, b * c, h, w)
+        # weight: (b*c_out, c_in, k, k), groups=b
+        out = F.conv2d(x, weight, padding=self.padding, groups=b)
+        out = out.view(b, self.out_channels, *out.shape[2:4])
+
+        return out
+
+    def __repr__(self):
+        return (f'{self.__class__.__name__}(in_channels={self.in_channels}, '
+                f'out_channels={self.out_channels}, '
+                f'kernel_size={self.kernel_size}, '
+                f'demodulate={self.demodulate}, sample_mode={self.sample_mode})')
+
+
+class StyleConv(nn.Module):
+    """Style conv.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+        kernel_size (int): Size of the convolving kernel.
+        num_style_feat (int): Channel number of style features.
+        demodulate (bool): Whether demodulate in the conv layer. Default: True.
+        sample_mode (str | None): Indicating 'upsample', 'downsample' or None.
+            Default: None.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 num_style_feat,
+                 demodulate=True,
+                 sample_mode=None,
+                 interpolation_mode='bilinear'):
+        super(StyleConv, self).__init__()
+        self.modulated_conv = ModulatedConv2d(
+            in_channels,
+            out_channels,
+            kernel_size,
+            num_style_feat,
+            demodulate=demodulate,
+            sample_mode=sample_mode,
+            interpolation_mode=interpolation_mode)
+        self.weight = nn.Parameter(torch.zeros(1))  # for noise injection
+        self.activate = FusedLeakyReLU(out_channels)
+
+    def forward(self, x, style, noise=None):
+        # modulate
+        out = self.modulated_conv(x, style)
+        # noise injection
+        if noise is None:
+            b, _, h, w = out.shape
+            noise = out.new_empty(b, 1, h, w).normal_()
+        out = out + self.weight * noise
+        # activation (with bias)
+        out = self.activate(out)
+        return out
+
+
+class ToRGB(nn.Module):
+    """To RGB from features.
+
+    Args:
+        in_channels (int): Channel number of input.
+        num_style_feat (int): Channel number of style features.
+        upsample (bool): Whether to upsample. Default: True.
+    """
+
+    def __init__(self, in_channels, num_style_feat, upsample=True, interpolation_mode='bilinear'):
+        super(ToRGB, self).__init__()
+        self.upsample = upsample
+        self.interpolation_mode = interpolation_mode
+        if self.interpolation_mode == 'nearest':
+            self.align_corners = None
+        else:
+            self.align_corners = False
+        self.modulated_conv = ModulatedConv2d(
+            in_channels,
+            3,
+            kernel_size=1,
+            num_style_feat=num_style_feat,
+            demodulate=False,
+            sample_mode=None,
+            interpolation_mode=interpolation_mode)
+        self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
+
+    def forward(self, x, style, skip=None):
+        """Forward function.
+
+        Args:
+            x (Tensor): Feature tensor with shape (b, c, h, w).
+            style (Tensor): Tensor with shape (b, num_style_feat).
+            skip (Tensor): Base/skip tensor. Default: None.
+
+        Returns:
+            Tensor: RGB images.
+        """
+        out = self.modulated_conv(x, style)
+        out = out + self.bias
+        if skip is not None:
+            if self.upsample:
+                skip = F.interpolate(
+                    skip, scale_factor=2, mode=self.interpolation_mode, align_corners=self.align_corners)
+            out = out + skip
+        return out
+
+
+class ConstantInput(nn.Module):
+    """Constant input.
+
+    Args:
+        num_channel (int): Channel number of constant input.
+        size (int): Spatial size of constant input.
+    """
+
+    def __init__(self, num_channel, size):
+        super(ConstantInput, self).__init__()
+        self.weight = nn.Parameter(torch.randn(1, num_channel, size, size))
+
+    def forward(self, batch):
+        out = self.weight.repeat(batch, 1, 1, 1)
+        return out
+
+
+@ARCH_REGISTRY.register()
+class StyleGAN2GeneratorBilinear(nn.Module):
+    """StyleGAN2 Generator.
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        num_style_feat (int): Channel number of style features. Default: 512.
+        num_mlp (int): Layer number of MLP style layers. Default: 8.
+        channel_multiplier (int): Channel multiplier for large networks of
+            StyleGAN2. Default: 2.
+        lr_mlp (float): Learning rate multiplier for mlp layers. Default: 0.01.
+        narrow (float): Narrow ratio for channels. Default: 1.0.
+    """
+
+    def __init__(self,
+                 out_size,
+                 num_style_feat=512,
+                 num_mlp=8,
+                 channel_multiplier=2,
+                 lr_mlp=0.01,
+                 narrow=1,
+                 interpolation_mode='bilinear'):
+        super(StyleGAN2GeneratorBilinear, self).__init__()
+        # Style MLP layers
+        self.num_style_feat = num_style_feat
+        style_mlp_layers = [NormStyleCode()]
+        for i in range(num_mlp):
+            style_mlp_layers.append(
+                EqualLinear(
+                    num_style_feat, num_style_feat, bias=True, bias_init_val=0, lr_mul=lr_mlp,
+                    activation='fused_lrelu'))
+        self.style_mlp = nn.Sequential(*style_mlp_layers)
+
+        channels = {
+            '4': int(512 * narrow),
+            '8': int(512 * narrow),
+            '16': int(512 * narrow),
+            '32': int(512 * narrow),
+            '64': int(256 * channel_multiplier * narrow),
+            '128': int(128 * channel_multiplier * narrow),
+            '256': int(64 * channel_multiplier * narrow),
+            '512': int(32 * channel_multiplier * narrow),
+            '1024': int(16 * channel_multiplier * narrow)
+        }
+        self.channels = channels
+
+        self.constant_input = ConstantInput(channels['4'], size=4)
+        self.style_conv1 = StyleConv(
+            channels['4'],
+            channels['4'],
+            kernel_size=3,
+            num_style_feat=num_style_feat,
+            demodulate=True,
+            sample_mode=None,
+            interpolation_mode=interpolation_mode)
+        self.to_rgb1 = ToRGB(channels['4'], num_style_feat, upsample=False, interpolation_mode=interpolation_mode)
+
+        self.log_size = int(math.log(out_size, 2))
+        self.num_layers = (self.log_size - 2) * 2 + 1
+        self.num_latent = self.log_size * 2 - 2
+
+        self.style_convs = nn.ModuleList()
+        self.to_rgbs = nn.ModuleList()
+        self.noises = nn.Module()
+
+        in_channels = channels['4']
+        # noise
+        for layer_idx in range(self.num_layers):
+            resolution = 2**((layer_idx + 5) // 2)
+            shape = [1, 1, resolution, resolution]
+            self.noises.register_buffer(f'noise{layer_idx}', torch.randn(*shape))
+        # style convs and to_rgbs
+        for i in range(3, self.log_size + 1):
+            out_channels = channels[f'{2**i}']
+            self.style_convs.append(
+                StyleConv(
+                    in_channels,
+                    out_channels,
+                    kernel_size=3,
+                    num_style_feat=num_style_feat,
+                    demodulate=True,
+                    sample_mode='upsample',
+                    interpolation_mode=interpolation_mode))
+            self.style_convs.append(
+                StyleConv(
+                    out_channels,
+                    out_channels,
+                    kernel_size=3,
+                    num_style_feat=num_style_feat,
+                    demodulate=True,
+                    sample_mode=None,
+                    interpolation_mode=interpolation_mode))
+            self.to_rgbs.append(
+                ToRGB(out_channels, num_style_feat, upsample=True, interpolation_mode=interpolation_mode))
+            in_channels = out_channels
+
+    def make_noise(self):
+        """Make noise for noise injection."""
+        device = self.constant_input.weight.device
+        noises = [torch.randn(1, 1, 4, 4, device=device)]
+
+        for i in range(3, self.log_size + 1):
+            for _ in range(2):
+                noises.append(torch.randn(1, 1, 2**i, 2**i, device=device))
+
+        return noises
+
+    def get_latent(self, x):
+        return self.style_mlp(x)
+
+    def mean_latent(self, num_latent):
+        latent_in = torch.randn(num_latent, self.num_style_feat, device=self.constant_input.weight.device)
+        latent = self.style_mlp(latent_in).mean(0, keepdim=True)
+        return latent
+
+    def forward(self,
+                styles,
+                input_is_latent=False,
+                noise=None,
+                randomize_noise=True,
+                truncation=1,
+                truncation_latent=None,
+                inject_index=None,
+                return_latents=False):
+        """Forward function for StyleGAN2Generator.
+
+        Args:
+            styles (list[Tensor]): Sample codes of styles.
+            input_is_latent (bool): Whether input is latent style.
+                Default: False.
+            noise (Tensor | None): Input noise or None. Default: None.
+            randomize_noise (bool): Randomize noise, used when 'noise' is
+                False. Default: True.
+            truncation (float): TODO. Default: 1.
+            truncation_latent (Tensor | None): TODO. Default: None.
+            inject_index (int | None): The injection index for mixing noise.
+                Default: None.
+            return_latents (bool): Whether to return style latents.
+                Default: False.
+        """
+        # style codes -> latents with Style MLP layer
+        if not input_is_latent:
+            styles = [self.style_mlp(s) for s in styles]
+        # noises
+        if noise is None:
+            if randomize_noise:
+                noise = [None] * self.num_layers  # for each style conv layer
+            else:  # use the stored noise
+                noise = [getattr(self.noises, f'noise{i}') for i in range(self.num_layers)]
+        # style truncation
+        if truncation < 1:
+            style_truncation = []
+            for style in styles:
+                style_truncation.append(truncation_latent + truncation * (style - truncation_latent))
+            styles = style_truncation
+        # get style latent with injection
+        if len(styles) == 1:
+            inject_index = self.num_latent
+
+            if styles[0].ndim < 3:
+                # repeat latent code for all the layers
+                latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            else:  # used for encoder with different latent code for each layer
+                latent = styles[0]
+        elif len(styles) == 2:  # mixing noises
+            if inject_index is None:
+                inject_index = random.randint(1, self.num_latent - 1)
+            latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            latent2 = styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
+            latent = torch.cat([latent1, latent2], 1)
+
+        # main generation
+        out = self.constant_input(latent.shape[0])
+        out = self.style_conv1(out, latent[:, 0], noise=noise[0])
+        skip = self.to_rgb1(out, latent[:, 1])
+
+        i = 1
+        for conv1, conv2, noise1, noise2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], noise[1::2],
+                                                        noise[2::2], self.to_rgbs):
+            out = conv1(out, latent[:, i], noise=noise1)
+            out = conv2(out, latent[:, i + 1], noise=noise2)
+            skip = to_rgb(out, latent[:, i + 2], skip)
+            i += 2
+
+        image = skip
+
+        if return_latents:
+            return image, latent
+        else:
+            return image, None
+
+
+class ScaledLeakyReLU(nn.Module):
+    """Scaled LeakyReLU.
+
+    Args:
+        negative_slope (float): Negative slope. Default: 0.2.
+    """
+
+    def __init__(self, negative_slope=0.2):
+        super(ScaledLeakyReLU, self).__init__()
+        self.negative_slope = negative_slope
+
+    def forward(self, x):
+        out = F.leaky_relu(x, negative_slope=self.negative_slope)
+        return out * math.sqrt(2)
+
+
+class EqualConv2d(nn.Module):
+    """Equalized Linear as StyleGAN2.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+        kernel_size (int): Size of the convolving kernel.
+        stride (int): Stride of the convolution. Default: 1
+        padding (int): Zero-padding added to both sides of the input.
+            Default: 0.
+        bias (bool): If ``True``, adds a learnable bias to the output.
+            Default: ``True``.
+        bias_init_val (float): Bias initialized value. Default: 0.
+    """
+
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, bias=True, bias_init_val=0):
+        super(EqualConv2d, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.stride = stride
+        self.padding = padding
+        self.scale = 1 / math.sqrt(in_channels * kernel_size**2)
+
+        self.weight = nn.Parameter(torch.randn(out_channels, in_channels, kernel_size, kernel_size))
+        if bias:
+            self.bias = nn.Parameter(torch.zeros(out_channels).fill_(bias_init_val))
+        else:
+            self.register_parameter('bias', None)
+
+    def forward(self, x):
+        out = F.conv2d(
+            x,
+            self.weight * self.scale,
+            bias=self.bias,
+            stride=self.stride,
+            padding=self.padding,
+        )
+
+        return out
+
+    def __repr__(self):
+        return (f'{self.__class__.__name__}(in_channels={self.in_channels}, '
+                f'out_channels={self.out_channels}, '
+                f'kernel_size={self.kernel_size},'
+                f' stride={self.stride}, padding={self.padding}, '
+                f'bias={self.bias is not None})')
+
+
+class ConvLayer(nn.Sequential):
+    """Conv Layer used in StyleGAN2 Discriminator.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+        kernel_size (int): Kernel size.
+        downsample (bool): Whether downsample by a factor of 2.
+            Default: False.
+        bias (bool): Whether with bias. Default: True.
+        activate (bool): Whether use activateion. Default: True.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 downsample=False,
+                 bias=True,
+                 activate=True,
+                 interpolation_mode='bilinear'):
+        layers = []
+        self.interpolation_mode = interpolation_mode
+        # downsample
+        if downsample:
+            if self.interpolation_mode == 'nearest':
+                self.align_corners = None
+            else:
+                self.align_corners = False
+
+            layers.append(
+                torch.nn.Upsample(scale_factor=0.5, mode=interpolation_mode, align_corners=self.align_corners))
+        stride = 1
+        self.padding = kernel_size // 2
+        # conv
+        layers.append(
+            EqualConv2d(
+                in_channels, out_channels, kernel_size, stride=stride, padding=self.padding, bias=bias
+                and not activate))
+        # activation
+        if activate:
+            if bias:
+                layers.append(FusedLeakyReLU(out_channels))
+            else:
+                layers.append(ScaledLeakyReLU(0.2))
+
+        super(ConvLayer, self).__init__(*layers)
+
+
+class ResBlock(nn.Module):
+    """Residual block used in StyleGAN2 Discriminator.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+    """
+
+    def __init__(self, in_channels, out_channels, interpolation_mode='bilinear'):
+        super(ResBlock, self).__init__()
+
+        self.conv1 = ConvLayer(in_channels, in_channels, 3, bias=True, activate=True)
+        self.conv2 = ConvLayer(
+            in_channels,
+            out_channels,
+            3,
+            downsample=True,
+            interpolation_mode=interpolation_mode,
+            bias=True,
+            activate=True)
+        self.skip = ConvLayer(
+            in_channels,
+            out_channels,
+            1,
+            downsample=True,
+            interpolation_mode=interpolation_mode,
+            bias=False,
+            activate=False)
+
+    def forward(self, x):
+        out = self.conv1(x)
+        out = self.conv2(out)
+        skip = self.skip(x)
+        out = (out + skip) / math.sqrt(2)
+        return out

gfpgan/archs/stylegan2_clean_arch.py

@@ -0,0 +1,368 @@
+import math
+import random
+import torch
+from basicsr.archs.arch_util import default_init_weights
+from basicsr.utils.registry import ARCH_REGISTRY
+from torch import nn
+from torch.nn import functional as F
+
+
+class NormStyleCode(nn.Module):
+
+    def forward(self, x):
+        """Normalize the style codes.
+
+        Args:
+            x (Tensor): Style codes with shape (b, c).
+
+        Returns:
+            Tensor: Normalized tensor.
+        """
+        return x * torch.rsqrt(torch.mean(x**2, dim=1, keepdim=True) + 1e-8)
+
+
+class ModulatedConv2d(nn.Module):
+    """Modulated Conv2d used in StyleGAN2.
+
+    There is no bias in ModulatedConv2d.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+        kernel_size (int): Size of the convolving kernel.
+        num_style_feat (int): Channel number of style features.
+        demodulate (bool): Whether to demodulate in the conv layer. Default: True.
+        sample_mode (str | None): Indicating 'upsample', 'downsample' or None. Default: None.
+        eps (float): A value added to the denominator for numerical stability. Default: 1e-8.
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 num_style_feat,
+                 demodulate=True,
+                 sample_mode=None,
+                 eps=1e-8):
+        super(ModulatedConv2d, self).__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.kernel_size = kernel_size
+        self.demodulate = demodulate
+        self.sample_mode = sample_mode
+        self.eps = eps
+
+        # modulation inside each modulated conv
+        self.modulation = nn.Linear(num_style_feat, in_channels, bias=True)
+        # initialization
+        default_init_weights(self.modulation, scale=1, bias_fill=1, a=0, mode='fan_in', nonlinearity='linear')
+
+        self.weight = nn.Parameter(
+            torch.randn(1, out_channels, in_channels, kernel_size, kernel_size) /
+            math.sqrt(in_channels * kernel_size**2))
+        self.padding = kernel_size // 2
+
+    def forward(self, x, style):
+        """Forward function.
+
+        Args:
+            x (Tensor): Tensor with shape (b, c, h, w).
+            style (Tensor): Tensor with shape (b, num_style_feat).
+
+        Returns:
+            Tensor: Modulated tensor after convolution.
+        """
+        b, c, h, w = x.shape  # c = c_in
+        # weight modulation
+        style = self.modulation(style).view(b, 1, c, 1, 1)
+        # self.weight: (1, c_out, c_in, k, k); style: (b, 1, c, 1, 1)
+        weight = self.weight * style  # (b, c_out, c_in, k, k)
+
+        if self.demodulate:
+            demod = torch.rsqrt(weight.pow(2).sum([2, 3, 4]) + self.eps)
+            weight = weight * demod.view(b, self.out_channels, 1, 1, 1)
+
+        weight = weight.view(b * self.out_channels, c, self.kernel_size, self.kernel_size)
+
+        # upsample or downsample if necessary
+        if self.sample_mode == 'upsample':
+            x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=False)
+        elif self.sample_mode == 'downsample':
+            x = F.interpolate(x, scale_factor=0.5, mode='bilinear', align_corners=False)
+
+        b, c, h, w = x.shape
+        x = x.view(1, b * c, h, w)
+        # weight: (b*c_out, c_in, k, k), groups=b
+        out = F.conv2d(x, weight, padding=self.padding, groups=b)
+        out = out.view(b, self.out_channels, *out.shape[2:4])
+
+        return out
+
+    def __repr__(self):
+        return (f'{self.__class__.__name__}(in_channels={self.in_channels}, out_channels={self.out_channels}, '
+                f'kernel_size={self.kernel_size}, demodulate={self.demodulate}, sample_mode={self.sample_mode})')
+
+
+class StyleConv(nn.Module):
+    """Style conv used in StyleGAN2.
+
+    Args:
+        in_channels (int): Channel number of the input.
+        out_channels (int): Channel number of the output.
+        kernel_size (int): Size of the convolving kernel.
+        num_style_feat (int): Channel number of style features.
+        demodulate (bool): Whether demodulate in the conv layer. Default: True.
+        sample_mode (str | None): Indicating 'upsample', 'downsample' or None. Default: None.
+    """
+
+    def __init__(self, in_channels, out_channels, kernel_size, num_style_feat, demodulate=True, sample_mode=None):
+        super(StyleConv, self).__init__()
+        self.modulated_conv = ModulatedConv2d(
+            in_channels, out_channels, kernel_size, num_style_feat, demodulate=demodulate, sample_mode=sample_mode)
+        self.weight = nn.Parameter(torch.zeros(1))  # for noise injection
+        self.bias = nn.Parameter(torch.zeros(1, out_channels, 1, 1))
+        self.activate = nn.LeakyReLU(negative_slope=0.2, inplace=True)
+
+    def forward(self, x, style, noise=None):
+        # modulate
+        out = self.modulated_conv(x, style) * 2**0.5  # for conversion
+        # noise injection
+        if noise is None:
+            b, _, h, w = out.shape
+            noise = out.new_empty(b, 1, h, w).normal_()
+        out = out + self.weight * noise
+        # add bias
+        out = out + self.bias
+        # activation
+        out = self.activate(out)
+        return out
+
+
+class ToRGB(nn.Module):
+    """To RGB (image space) from features.
+
+    Args:
+        in_channels (int): Channel number of input.
+        num_style_feat (int): Channel number of style features.
+        upsample (bool): Whether to upsample. Default: True.
+    """
+
+    def __init__(self, in_channels, num_style_feat, upsample=True):
+        super(ToRGB, self).__init__()
+        self.upsample = upsample
+        self.modulated_conv = ModulatedConv2d(
+            in_channels, 3, kernel_size=1, num_style_feat=num_style_feat, demodulate=False, sample_mode=None)
+        self.bias = nn.Parameter(torch.zeros(1, 3, 1, 1))
+
+    def forward(self, x, style, skip=None):
+        """Forward function.
+
+        Args:
+            x (Tensor): Feature tensor with shape (b, c, h, w).
+            style (Tensor): Tensor with shape (b, num_style_feat).
+            skip (Tensor): Base/skip tensor. Default: None.
+
+        Returns:
+            Tensor: RGB images.
+        """
+        out = self.modulated_conv(x, style)
+        out = out + self.bias
+        if skip is not None:
+            if self.upsample:
+                skip = F.interpolate(skip, scale_factor=2, mode='bilinear', align_corners=False)
+            out = out + skip
+        return out
+
+
+class ConstantInput(nn.Module):
+    """Constant input.
+
+    Args:
+        num_channel (int): Channel number of constant input.
+        size (int): Spatial size of constant input.
+    """
+
+    def __init__(self, num_channel, size):
+        super(ConstantInput, self).__init__()
+        self.weight = nn.Parameter(torch.randn(1, num_channel, size, size))
+
+    def forward(self, batch):
+        out = self.weight.repeat(batch, 1, 1, 1)
+        return out
+
+
+@ARCH_REGISTRY.register()
+class StyleGAN2GeneratorClean(nn.Module):
+    """Clean version of StyleGAN2 Generator.
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        num_style_feat (int): Channel number of style features. Default: 512.
+        num_mlp (int): Layer number of MLP style layers. Default: 8.
+        channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
+        narrow (float): Narrow ratio for channels. Default: 1.0.
+    """
+
+    def __init__(self, out_size, num_style_feat=512, num_mlp=8, channel_multiplier=2, narrow=1):
+        super(StyleGAN2GeneratorClean, self).__init__()
+        # Style MLP layers
+        self.num_style_feat = num_style_feat
+        style_mlp_layers = [NormStyleCode()]
+        for i in range(num_mlp):
+            style_mlp_layers.extend(
+                [nn.Linear(num_style_feat, num_style_feat, bias=True),
+                 nn.LeakyReLU(negative_slope=0.2, inplace=True)])
+        self.style_mlp = nn.Sequential(*style_mlp_layers)
+        # initialization
+        default_init_weights(self.style_mlp, scale=1, bias_fill=0, a=0.2, mode='fan_in', nonlinearity='leaky_relu')
+
+        # channel list
+        channels = {
+            '4': int(512 * narrow),
+            '8': int(512 * narrow),
+            '16': int(512 * narrow),
+            '32': int(512 * narrow),
+            '64': int(256 * channel_multiplier * narrow),
+            '128': int(128 * channel_multiplier * narrow),
+            '256': int(64 * channel_multiplier * narrow),
+            '512': int(32 * channel_multiplier * narrow),
+            '1024': int(16 * channel_multiplier * narrow)
+        }
+        self.channels = channels
+
+        self.constant_input = ConstantInput(channels['4'], size=4)
+        self.style_conv1 = StyleConv(
+            channels['4'],
+            channels['4'],
+            kernel_size=3,
+            num_style_feat=num_style_feat,
+            demodulate=True,
+            sample_mode=None)
+        self.to_rgb1 = ToRGB(channels['4'], num_style_feat, upsample=False)
+
+        self.log_size = int(math.log(out_size, 2))
+        self.num_layers = (self.log_size - 2) * 2 + 1
+        self.num_latent = self.log_size * 2 - 2
+
+        self.style_convs = nn.ModuleList()
+        self.to_rgbs = nn.ModuleList()
+        self.noises = nn.Module()
+
+        in_channels = channels['4']
+        # noise
+        for layer_idx in range(self.num_layers):
+            resolution = 2**((layer_idx + 5) // 2)
+            shape = [1, 1, resolution, resolution]
+            self.noises.register_buffer(f'noise{layer_idx}', torch.randn(*shape))
+        # style convs and to_rgbs
+        for i in range(3, self.log_size + 1):
+            out_channels = channels[f'{2**i}']
+            self.style_convs.append(
+                StyleConv(
+                    in_channels,
+                    out_channels,
+                    kernel_size=3,
+                    num_style_feat=num_style_feat,
+                    demodulate=True,
+                    sample_mode='upsample'))
+            self.style_convs.append(
+                StyleConv(
+                    out_channels,
+                    out_channels,
+                    kernel_size=3,
+                    num_style_feat=num_style_feat,
+                    demodulate=True,
+                    sample_mode=None))
+            self.to_rgbs.append(ToRGB(out_channels, num_style_feat, upsample=True))
+            in_channels = out_channels
+
+    def make_noise(self):
+        """Make noise for noise injection."""
+        device = self.constant_input.weight.device
+        noises = [torch.randn(1, 1, 4, 4, device=device)]
+
+        for i in range(3, self.log_size + 1):
+            for _ in range(2):
+                noises.append(torch.randn(1, 1, 2**i, 2**i, device=device))
+
+        return noises
+
+    def get_latent(self, x):
+        return self.style_mlp(x)
+
+    def mean_latent(self, num_latent):
+        latent_in = torch.randn(num_latent, self.num_style_feat, device=self.constant_input.weight.device)
+        latent = self.style_mlp(latent_in).mean(0, keepdim=True)
+        return latent
+
+    def forward(self,
+                styles,
+                input_is_latent=False,
+                noise=None,
+                randomize_noise=True,
+                truncation=1,
+                truncation_latent=None,
+                inject_index=None,
+                return_latents=False):
+        """Forward function for StyleGAN2GeneratorClean.
+
+        Args:
+            styles (list[Tensor]): Sample codes of styles.
+            input_is_latent (bool): Whether input is latent style. Default: False.
+            noise (Tensor | None): Input noise or None. Default: None.
+            randomize_noise (bool): Randomize noise, used when 'noise' is False. Default: True.
+            truncation (float): The truncation ratio. Default: 1.
+            truncation_latent (Tensor | None): The truncation latent tensor. Default: None.
+            inject_index (int | None): The injection index for mixing noise. Default: None.
+            return_latents (bool): Whether to return style latents. Default: False.
+        """
+        # style codes -> latents with Style MLP layer
+        if not input_is_latent:
+            styles = [self.style_mlp(s) for s in styles]
+        # noises
+        if noise is None:
+            if randomize_noise:
+                noise = [None] * self.num_layers  # for each style conv layer
+            else:  # use the stored noise
+                noise = [getattr(self.noises, f'noise{i}') for i in range(self.num_layers)]
+        # style truncation
+        if truncation < 1:
+            style_truncation = []
+            for style in styles:
+                style_truncation.append(truncation_latent + truncation * (style - truncation_latent))
+            styles = style_truncation
+        # get style latents with injection
+        if len(styles) == 1:
+            inject_index = self.num_latent
+
+            if styles[0].ndim < 3:
+                # repeat latent code for all the layers
+                latent = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            else:  # used for encoder with different latent code for each layer
+                latent = styles[0]
+        elif len(styles) == 2:  # mixing noises
+            if inject_index is None:
+                inject_index = random.randint(1, self.num_latent - 1)
+            latent1 = styles[0].unsqueeze(1).repeat(1, inject_index, 1)
+            latent2 = styles[1].unsqueeze(1).repeat(1, self.num_latent - inject_index, 1)
+            latent = torch.cat([latent1, latent2], 1)
+
+        # main generation
+        out = self.constant_input(latent.shape[0])
+        out = self.style_conv1(out, latent[:, 0], noise=noise[0])
+        skip = self.to_rgb1(out, latent[:, 1])
+
+        i = 1
+        for conv1, conv2, noise1, noise2, to_rgb in zip(self.style_convs[::2], self.style_convs[1::2], noise[1::2],
+                                                        noise[2::2], self.to_rgbs):
+            out = conv1(out, latent[:, i], noise=noise1)
+            out = conv2(out, latent[:, i + 1], noise=noise2)
+            skip = to_rgb(out, latent[:, i + 2], skip)  # feature back to the rgb space
+            i += 2
+
+        image = skip
+
+        if return_latents:
+            return image, latent
+        else:
+            return image, None

gfpgan/data/__init__.py

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

gfpgan/data/ffhq_degradation_dataset.py

@@ -0,0 +1,230 @@
+import cv2
+import math
+import numpy as np
+import os.path as osp
+import torch
+import torch.utils.data as data
+from basicsr.data import degradations as degradations
+from basicsr.data.data_util import paths_from_folder
+from basicsr.data.transforms import augment
+from basicsr.utils import FileClient, get_root_logger, imfrombytes, img2tensor
+from basicsr.utils.registry import DATASET_REGISTRY
+from torchvision.transforms.functional import (adjust_brightness, adjust_contrast, adjust_hue, adjust_saturation,
+                                               normalize)
+
+
+@DATASET_REGISTRY.register()
+class FFHQDegradationDataset(data.Dataset):
+    """FFHQ dataset for GFPGAN.
+
+    It reads high resolution images, and then generate low-quality (LQ) images on-the-fly.
+
+    Args:
+        opt (dict): Config for train datasets. It contains the following keys:
+            dataroot_gt (str): Data root path for gt.
+            io_backend (dict): IO backend type and other kwarg.
+            mean (list | tuple): Image mean.
+            std (list | tuple): Image std.
+            use_hflip (bool): Whether to horizontally flip.
+            Please see more options in the codes.
+    """
+
+    def __init__(self, opt):
+        super(FFHQDegradationDataset, 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.mean = opt['mean']
+        self.std = opt['std']
+        self.out_size = opt['out_size']
+
+        self.crop_components = opt.get('crop_components', False)  # facial components
+        self.eye_enlarge_ratio = opt.get('eye_enlarge_ratio', 1)  # whether enlarge eye regions
+
+        if self.crop_components:
+            # load component list from a pre-process pth files
+            self.components_list = torch.load(opt.get('component_path'))
+
+        # file client (lmdb io backend)
+        if self.io_backend_opt['type'] == 'lmdb':
+            self.io_backend_opt['db_paths'] = self.gt_folder
+            if not self.gt_folder.endswith('.lmdb'):
+                raise ValueError(f"'dataroot_gt' should end with '.lmdb', but received {self.gt_folder}")
+            with open(osp.join(self.gt_folder, 'meta_info.txt')) as fin:
+                self.paths = [line.split('.')[0] for line in fin]
+        else:
+            # disk backend: scan file list from a folder
+            self.paths = paths_from_folder(self.gt_folder)
+
+        # degradation configurations
+        self.blur_kernel_size = opt['blur_kernel_size']
+        self.kernel_list = opt['kernel_list']
+        self.kernel_prob = opt['kernel_prob']
+        self.blur_sigma = opt['blur_sigma']
+        self.downsample_range = opt['downsample_range']
+        self.noise_range = opt['noise_range']
+        self.jpeg_range = opt['jpeg_range']
+
+        # color jitter
+        self.color_jitter_prob = opt.get('color_jitter_prob')
+        self.color_jitter_pt_prob = opt.get('color_jitter_pt_prob')
+        self.color_jitter_shift = opt.get('color_jitter_shift', 20)
+        # to gray
+        self.gray_prob = opt.get('gray_prob')
+
+        logger = get_root_logger()
+        logger.info(f'Blur: blur_kernel_size {self.blur_kernel_size}, sigma: [{", ".join(map(str, self.blur_sigma))}]')
+        logger.info(f'Downsample: downsample_range [{", ".join(map(str, self.downsample_range))}]')
+        logger.info(f'Noise: [{", ".join(map(str, self.noise_range))}]')
+        logger.info(f'JPEG compression: [{", ".join(map(str, self.jpeg_range))}]')
+
+        if self.color_jitter_prob is not None:
+            logger.info(f'Use random color jitter. Prob: {self.color_jitter_prob}, shift: {self.color_jitter_shift}')
+        if self.gray_prob is not None:
+            logger.info(f'Use random gray. Prob: {self.gray_prob}')
+        self.color_jitter_shift /= 255.
+
+    @staticmethod
+    def color_jitter(img, shift):
+        """jitter color: randomly jitter the RGB values, in numpy formats"""
+        jitter_val = np.random.uniform(-shift, shift, 3).astype(np.float32)
+        img = img + jitter_val
+        img = np.clip(img, 0, 1)
+        return img
+
+    @staticmethod
+    def color_jitter_pt(img, brightness, contrast, saturation, hue):
+        """jitter color: randomly jitter the brightness, contrast, saturation, and hue, in torch Tensor formats"""
+        fn_idx = torch.randperm(4)
+        for fn_id in fn_idx:
+            if fn_id == 0 and brightness is not None:
+                brightness_factor = torch.tensor(1.0).uniform_(brightness[0], brightness[1]).item()
+                img = adjust_brightness(img, brightness_factor)
+
+            if fn_id == 1 and contrast is not None:
+                contrast_factor = torch.tensor(1.0).uniform_(contrast[0], contrast[1]).item()
+                img = adjust_contrast(img, contrast_factor)
+
+            if fn_id == 2 and saturation is not None:
+                saturation_factor = torch.tensor(1.0).uniform_(saturation[0], saturation[1]).item()
+                img = adjust_saturation(img, saturation_factor)
+
+            if fn_id == 3 and hue is not None:
+                hue_factor = torch.tensor(1.0).uniform_(hue[0], hue[1]).item()
+                img = adjust_hue(img, hue_factor)
+        return img
+
+    def get_component_coordinates(self, index, status):
+        """Get facial component (left_eye, right_eye, mouth) coordinates from a pre-loaded pth file"""
+        components_bbox = self.components_list[f'{index:08d}']
+        if status[0]:  # hflip
+            # exchange right and left eye
+            tmp = components_bbox['left_eye']
+            components_bbox['left_eye'] = components_bbox['right_eye']
+            components_bbox['right_eye'] = tmp
+            # modify the width coordinate
+            components_bbox['left_eye'][0] = self.out_size - components_bbox['left_eye'][0]
+            components_bbox['right_eye'][0] = self.out_size - components_bbox['right_eye'][0]
+            components_bbox['mouth'][0] = self.out_size - components_bbox['mouth'][0]
+
+        # get coordinates
+        locations = []
+        for part in ['left_eye', 'right_eye', 'mouth']:
+            mean = components_bbox[part][0:2]
+            half_len = components_bbox[part][2]
+            if 'eye' in part:
+                half_len *= self.eye_enlarge_ratio
+            loc = np.hstack((mean - half_len + 1, mean + half_len))
+            loc = torch.from_numpy(loc).float()
+            locations.append(loc)
+        return locations
+
+    def __getitem__(self, index):
+        if self.file_client is None:
+            self.file_client = FileClient(self.io_backend_opt.pop('type'), **self.io_backend_opt)
+
+        # load gt image
+        # Shape: (h, w, c); channel order: BGR; image range: [0, 1], float32.
+        gt_path = self.paths[index]
+        img_bytes = self.file_client.get(gt_path)
+        img_gt = imfrombytes(img_bytes, float32=True)
+
+        # random horizontal flip
+        img_gt, status = augment(img_gt, hflip=self.opt['use_hflip'], rotation=False, return_status=True)
+        h, w, _ = img_gt.shape
+
+        # get facial component coordinates
+        if self.crop_components:
+            locations = self.get_component_coordinates(index, status)
+            loc_left_eye, loc_right_eye, loc_mouth = locations
+
+        # ------------------------ generate lq image ------------------------ #
+        # blur
+        kernel = degradations.random_mixed_kernels(
+            self.kernel_list,
+            self.kernel_prob,
+            self.blur_kernel_size,
+            self.blur_sigma,
+            self.blur_sigma, [-math.pi, math.pi],
+            noise_range=None)
+        img_lq = cv2.filter2D(img_gt, -1, kernel)
+        # downsample
+        scale = np.random.uniform(self.downsample_range[0], self.downsample_range[1])
+        img_lq = cv2.resize(img_lq, (int(w // scale), int(h // scale)), interpolation=cv2.INTER_LINEAR)
+        # noise
+        if self.noise_range is not None:
+            img_lq = degradations.random_add_gaussian_noise(img_lq, self.noise_range)
+        # jpeg compression
+        if self.jpeg_range is not None:
+            img_lq = degradations.random_add_jpg_compression(img_lq, self.jpeg_range)
+
+        # resize to original size
+        img_lq = cv2.resize(img_lq, (w, h), interpolation=cv2.INTER_LINEAR)
+
+        # random color jitter (only for lq)
+        if self.color_jitter_prob is not None and (np.random.uniform() < self.color_jitter_prob):
+            img_lq = self.color_jitter(img_lq, self.color_jitter_shift)
+        # random to gray (only for lq)
+        if self.gray_prob and np.random.uniform() < self.gray_prob:
+            img_lq = cv2.cvtColor(img_lq, cv2.COLOR_BGR2GRAY)
+            img_lq = np.tile(img_lq[:, :, None], [1, 1, 3])
+            if self.opt.get('gt_gray'):  # whether convert GT to gray images
+                img_gt = cv2.cvtColor(img_gt, cv2.COLOR_BGR2GRAY)
+                img_gt = np.tile(img_gt[:, :, None], [1, 1, 3])  # repeat the color channels
+
+        # BGR to RGB, HWC to CHW, numpy to tensor
+        img_gt, img_lq = img2tensor([img_gt, img_lq], bgr2rgb=True, float32=True)
+
+        # random color jitter (pytorch version) (only for lq)
+        if self.color_jitter_pt_prob is not None and (np.random.uniform() < self.color_jitter_pt_prob):
+            brightness = self.opt.get('brightness', (0.5, 1.5))
+            contrast = self.opt.get('contrast', (0.5, 1.5))
+            saturation = self.opt.get('saturation', (0, 1.5))
+            hue = self.opt.get('hue', (-0.1, 0.1))
+            img_lq = self.color_jitter_pt(img_lq, brightness, contrast, saturation, hue)
+
+        # round and clip
+        img_lq = torch.clamp((img_lq * 255.0).round(), 0, 255) / 255.
+
+        # normalize
+        normalize(img_gt, self.mean, self.std, inplace=True)
+        normalize(img_lq, self.mean, self.std, inplace=True)
+
+        if self.crop_components:
+            return_dict = {
+                'lq': img_lq,
+                'gt': img_gt,
+                'gt_path': gt_path,
+                'loc_left_eye': loc_left_eye,
+                'loc_right_eye': loc_right_eye,
+                'loc_mouth': loc_mouth
+            }
+            return return_dict
+        else:
+            return {'lq': img_lq, 'gt': img_gt, 'gt_path': gt_path}
+
+    def __len__(self):
+        return len(self.paths)

gfpgan/models/__init__.py

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

gfpgan/models/gfpgan_model.py

@@ -0,0 +1,579 @@
+import math
+import os.path as osp
+import torch
+from basicsr.archs import build_network
+from basicsr.losses import build_loss
+from basicsr.losses.gan_loss import r1_penalty
+from basicsr.metrics import calculate_metric
+from basicsr.models.base_model import BaseModel
+from basicsr.utils import get_root_logger, imwrite, tensor2img
+from basicsr.utils.registry import MODEL_REGISTRY
+from collections import OrderedDict
+from torch.nn import functional as F
+from torchvision.ops import roi_align
+from tqdm import tqdm
+
+
+@MODEL_REGISTRY.register()
+class GFPGANModel(BaseModel):
+    """The GFPGAN model for Towards real-world blind face restoratin with generative facial prior"""
+
+    def __init__(self, opt):
+        super(GFPGANModel, self).__init__(opt)
+        self.idx = 0  # it is used for saving data for check
+
+        # define network
+        self.net_g = build_network(opt['network_g'])
+        self.net_g = self.model_to_device(self.net_g)
+        self.print_network(self.net_g)
+
+        # load pretrained model
+        load_path = self.opt['path'].get('pretrain_network_g', None)
+        if load_path is not None:
+            param_key = self.opt['path'].get('param_key_g', 'params')
+            self.load_network(self.net_g, load_path, self.opt['path'].get('strict_load_g', True), param_key)
+
+        self.log_size = int(math.log(self.opt['network_g']['out_size'], 2))
+
+        if self.is_train:
+            self.init_training_settings()
+
+    def init_training_settings(self):
+        train_opt = self.opt['train']
+
+        # ----------- define net_d ----------- #
+        self.net_d = build_network(self.opt['network_d'])
+        self.net_d = self.model_to_device(self.net_d)
+        self.print_network(self.net_d)
+        # load pretrained model
+        load_path = self.opt['path'].get('pretrain_network_d', None)
+        if load_path is not None:
+            self.load_network(self.net_d, load_path, self.opt['path'].get('strict_load_d', True))
+
+        # ----------- define net_g with Exponential Moving Average (EMA) ----------- #
+        # net_g_ema only used for testing on one GPU and saving. There is no need to wrap with DistributedDataParallel
+        self.net_g_ema = build_network(self.opt['network_g']).to(self.device)
+        # load pretrained model
+        load_path = self.opt['path'].get('pretrain_network_g', None)
+        if load_path is not None:
+            self.load_network(self.net_g_ema, load_path, self.opt['path'].get('strict_load_g', True), 'params_ema')
+        else:
+            self.model_ema(0)  # copy net_g weight
+
+        self.net_g.train()
+        self.net_d.train()
+        self.net_g_ema.eval()
+
+        # ----------- facial component networks ----------- #
+        if ('network_d_left_eye' in self.opt and 'network_d_right_eye' in self.opt and 'network_d_mouth' in self.opt):
+            self.use_facial_disc = True
+        else:
+            self.use_facial_disc = False
+
+        if self.use_facial_disc:
+            # left eye
+            self.net_d_left_eye = build_network(self.opt['network_d_left_eye'])
+            self.net_d_left_eye = self.model_to_device(self.net_d_left_eye)
+            self.print_network(self.net_d_left_eye)
+            load_path = self.opt['path'].get('pretrain_network_d_left_eye')
+            if load_path is not None:
+                self.load_network(self.net_d_left_eye, load_path, True, 'params')
+            # right eye
+            self.net_d_right_eye = build_network(self.opt['network_d_right_eye'])
+            self.net_d_right_eye = self.model_to_device(self.net_d_right_eye)
+            self.print_network(self.net_d_right_eye)
+            load_path = self.opt['path'].get('pretrain_network_d_right_eye')
+            if load_path is not None:
+                self.load_network(self.net_d_right_eye, load_path, True, 'params')
+            # mouth
+            self.net_d_mouth = build_network(self.opt['network_d_mouth'])
+            self.net_d_mouth = self.model_to_device(self.net_d_mouth)
+            self.print_network(self.net_d_mouth)
+            load_path = self.opt['path'].get('pretrain_network_d_mouth')
+            if load_path is not None:
+                self.load_network(self.net_d_mouth, load_path, True, 'params')
+
+            self.net_d_left_eye.train()
+            self.net_d_right_eye.train()
+            self.net_d_mouth.train()
+
+            # ----------- define facial component gan loss ----------- #
+            self.cri_component = build_loss(train_opt['gan_component_opt']).to(self.device)
+
+        # ----------- define losses ----------- #
+        # pixel loss
+        if train_opt.get('pixel_opt'):
+            self.cri_pix = build_loss(train_opt['pixel_opt']).to(self.device)
+        else:
+            self.cri_pix = None
+
+        # perceptual loss
+        if train_opt.get('perceptual_opt'):
+            self.cri_perceptual = build_loss(train_opt['perceptual_opt']).to(self.device)
+        else:
+            self.cri_perceptual = None
+
+        # L1 loss is used in pyramid loss, component style loss and identity loss
+        self.cri_l1 = build_loss(train_opt['L1_opt']).to(self.device)
+
+        # gan loss (wgan)
+        self.cri_gan = build_loss(train_opt['gan_opt']).to(self.device)
+
+        # ----------- define identity loss ----------- #
+        if 'network_identity' in self.opt:
+            self.use_identity = True
+        else:
+            self.use_identity = False
+
+        if self.use_identity:
+            # define identity network
+            self.network_identity = build_network(self.opt['network_identity'])
+            self.network_identity = self.model_to_device(self.network_identity)
+            self.print_network(self.network_identity)
+            load_path = self.opt['path'].get('pretrain_network_identity')
+            if load_path is not None:
+                self.load_network(self.network_identity, load_path, True, None)
+            self.network_identity.eval()
+            for param in self.network_identity.parameters():
+                param.requires_grad = False
+
+        # regularization weights
+        self.r1_reg_weight = train_opt['r1_reg_weight']  # for discriminator
+        self.net_d_iters = train_opt.get('net_d_iters', 1)
+        self.net_d_init_iters = train_opt.get('net_d_init_iters', 0)
+        self.net_d_reg_every = train_opt['net_d_reg_every']
+
+        # set up optimizers and schedulers
+        self.setup_optimizers()
+        self.setup_schedulers()
+
+    def setup_optimizers(self):
+        train_opt = self.opt['train']
+
+        # ----------- optimizer g ----------- #
+        net_g_reg_ratio = 1
+        normal_params = []
+        for _, param in self.net_g.named_parameters():
+            normal_params.append(param)
+        optim_params_g = [{  # add normal params first
+            'params': normal_params,
+            'lr': train_opt['optim_g']['lr']
+        }]
+        optim_type = train_opt['optim_g'].pop('type')
+        lr = train_opt['optim_g']['lr'] * net_g_reg_ratio
+        betas = (0**net_g_reg_ratio, 0.99**net_g_reg_ratio)
+        self.optimizer_g = self.get_optimizer(optim_type, optim_params_g, lr, betas=betas)
+        self.optimizers.append(self.optimizer_g)
+
+        # ----------- optimizer d ----------- #
+        net_d_reg_ratio = self.net_d_reg_every / (self.net_d_reg_every + 1)
+        normal_params = []
+        for _, param in self.net_d.named_parameters():
+            normal_params.append(param)
+        optim_params_d = [{  # add normal params first
+            'params': normal_params,
+            'lr': train_opt['optim_d']['lr']
+        }]
+        optim_type = train_opt['optim_d'].pop('type')
+        lr = train_opt['optim_d']['lr'] * net_d_reg_ratio
+        betas = (0**net_d_reg_ratio, 0.99**net_d_reg_ratio)
+        self.optimizer_d = self.get_optimizer(optim_type, optim_params_d, lr, betas=betas)
+        self.optimizers.append(self.optimizer_d)
+
+        # ----------- optimizers for facial component networks ----------- #
+        if self.use_facial_disc:
+            # setup optimizers for facial component discriminators
+            optim_type = train_opt['optim_component'].pop('type')
+            lr = train_opt['optim_component']['lr']
+            # left eye
+            self.optimizer_d_left_eye = self.get_optimizer(
+                optim_type, self.net_d_left_eye.parameters(), lr, betas=(0.9, 0.99))
+            self.optimizers.append(self.optimizer_d_left_eye)
+            # right eye
+            self.optimizer_d_right_eye = self.get_optimizer(
+                optim_type, self.net_d_right_eye.parameters(), lr, betas=(0.9, 0.99))
+            self.optimizers.append(self.optimizer_d_right_eye)
+            # mouth
+            self.optimizer_d_mouth = self.get_optimizer(
+                optim_type, self.net_d_mouth.parameters(), lr, betas=(0.9, 0.99))
+            self.optimizers.append(self.optimizer_d_mouth)
+
+    def feed_data(self, data):
+        self.lq = data['lq'].to(self.device)
+        if 'gt' in data:
+            self.gt = data['gt'].to(self.device)
+
+        if 'loc_left_eye' in data:
+            # get facial component locations, shape (batch, 4)
+            self.loc_left_eyes = data['loc_left_eye']
+            self.loc_right_eyes = data['loc_right_eye']
+            self.loc_mouths = data['loc_mouth']
+
+        # uncomment to check data
+        # import torchvision
+        # if self.opt['rank'] == 0:
+        #     import os
+        #     os.makedirs('tmp/gt', exist_ok=True)
+        #     os.makedirs('tmp/lq', exist_ok=True)
+        #     print(self.idx)
+        #     torchvision.utils.save_image(
+        #         self.gt, f'tmp/gt/gt_{self.idx}.png', nrow=4, padding=2, normalize=True, range=(-1, 1))
+        #     torchvision.utils.save_image(
+        #         self.lq, f'tmp/lq/lq{self.idx}.png', nrow=4, padding=2, normalize=True, range=(-1, 1))
+        #     self.idx = self.idx + 1
+
+    def construct_img_pyramid(self):
+        """Construct image pyramid for intermediate restoration loss"""
+        pyramid_gt = [self.gt]
+        down_img = self.gt
+        for _ in range(0, self.log_size - 3):
+            down_img = F.interpolate(down_img, scale_factor=0.5, mode='bilinear', align_corners=False)
+            pyramid_gt.insert(0, down_img)
+        return pyramid_gt
+
+    def get_roi_regions(self, eye_out_size=80, mouth_out_size=120):
+        face_ratio = int(self.opt['network_g']['out_size'] / 512)
+        eye_out_size *= face_ratio
+        mouth_out_size *= face_ratio
+
+        rois_eyes = []
+        rois_mouths = []
+        for b in range(self.loc_left_eyes.size(0)):  # loop for batch size
+            # left eye and right eye
+            img_inds = self.loc_left_eyes.new_full((2, 1), b)
+            bbox = torch.stack([self.loc_left_eyes[b, :], self.loc_right_eyes[b, :]], dim=0)  # shape: (2, 4)
+            rois = torch.cat([img_inds, bbox], dim=-1)  # shape: (2, 5)
+            rois_eyes.append(rois)
+            # mouse
+            img_inds = self.loc_left_eyes.new_full((1, 1), b)
+            rois = torch.cat([img_inds, self.loc_mouths[b:b + 1, :]], dim=-1)  # shape: (1, 5)
+            rois_mouths.append(rois)
+
+        rois_eyes = torch.cat(rois_eyes, 0).to(self.device)
+        rois_mouths = torch.cat(rois_mouths, 0).to(self.device)
+
+        # real images
+        all_eyes = roi_align(self.gt, boxes=rois_eyes, output_size=eye_out_size) * face_ratio
+        self.left_eyes_gt = all_eyes[0::2, :, :, :]
+        self.right_eyes_gt = all_eyes[1::2, :, :, :]
+        self.mouths_gt = roi_align(self.gt, boxes=rois_mouths, output_size=mouth_out_size) * face_ratio
+        # output
+        all_eyes = roi_align(self.output, boxes=rois_eyes, output_size=eye_out_size) * face_ratio
+        self.left_eyes = all_eyes[0::2, :, :, :]
+        self.right_eyes = all_eyes[1::2, :, :, :]
+        self.mouths = roi_align(self.output, boxes=rois_mouths, output_size=mouth_out_size) * face_ratio
+
+    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
+
+    def gray_resize_for_identity(self, out, size=128):
+        out_gray = (0.2989 * out[:, 0, :, :] + 0.5870 * out[:, 1, :, :] + 0.1140 * out[:, 2, :, :])
+        out_gray = out_gray.unsqueeze(1)
+        out_gray = F.interpolate(out_gray, (size, size), mode='bilinear', align_corners=False)
+        return out_gray
+
+    def optimize_parameters(self, current_iter):
+        # optimize net_g
+        for p in self.net_d.parameters():
+            p.requires_grad = False
+        self.optimizer_g.zero_grad()
+
+        # do not update facial component net_d
+        if self.use_facial_disc:
+            for p in self.net_d_left_eye.parameters():
+                p.requires_grad = False
+            for p in self.net_d_right_eye.parameters():
+                p.requires_grad = False
+            for p in self.net_d_mouth.parameters():
+                p.requires_grad = False
+
+        # image pyramid loss weight
+        pyramid_loss_weight = self.opt['train'].get('pyramid_loss_weight', 0)
+        if pyramid_loss_weight > 0 and current_iter > self.opt['train'].get('remove_pyramid_loss', float('inf')):
+            pyramid_loss_weight = 1e-12  # very small weight to avoid unused param error
+        if pyramid_loss_weight > 0:
+            self.output, out_rgbs = self.net_g(self.lq, return_rgb=True)
+            pyramid_gt = self.construct_img_pyramid()
+        else:
+            self.output, out_rgbs = self.net_g(self.lq, return_rgb=False)
+
+        # get roi-align regions
+        if self.use_facial_disc:
+            self.get_roi_regions(eye_out_size=80, mouth_out_size=120)
+
+        l_g_total = 0
+        loss_dict = OrderedDict()
+        if (current_iter % self.net_d_iters == 0 and current_iter > self.net_d_init_iters):
+            # pixel loss
+            if self.cri_pix:
+                l_g_pix = self.cri_pix(self.output, self.gt)
+                l_g_total += l_g_pix
+                loss_dict['l_g_pix'] = l_g_pix
+
+            # image pyramid loss
+            if pyramid_loss_weight > 0:
+                for i in range(0, self.log_size - 2):
+                    l_pyramid = self.cri_l1(out_rgbs[i], pyramid_gt[i]) * pyramid_loss_weight
+                    l_g_total += l_pyramid
+                    loss_dict[f'l_p_{2**(i+3)}'] = l_pyramid
+
+            # perceptual loss
+            if self.cri_perceptual:
+                l_g_percep, l_g_style = self.cri_perceptual(self.output, self.gt)
+                if l_g_percep is not None:
+                    l_g_total += l_g_percep
+                    loss_dict['l_g_percep'] = l_g_percep
+                if l_g_style is not None:
+                    l_g_total += l_g_style
+                    loss_dict['l_g_style'] = l_g_style
+
+            # gan loss
+            fake_g_pred = self.net_d(self.output)
+            l_g_gan = self.cri_gan(fake_g_pred, True, is_disc=False)
+            l_g_total += l_g_gan
+            loss_dict['l_g_gan'] = l_g_gan
+
+            # facial component loss
+            if self.use_facial_disc:
+                # left eye
+                fake_left_eye, fake_left_eye_feats = self.net_d_left_eye(self.left_eyes, return_feats=True)
+                l_g_gan = self.cri_component(fake_left_eye, True, is_disc=False)
+                l_g_total += l_g_gan
+                loss_dict['l_g_gan_left_eye'] = l_g_gan
+                # right eye
+                fake_right_eye, fake_right_eye_feats = self.net_d_right_eye(self.right_eyes, return_feats=True)
+                l_g_gan = self.cri_component(fake_right_eye, True, is_disc=False)
+                l_g_total += l_g_gan
+                loss_dict['l_g_gan_right_eye'] = l_g_gan
+                # mouth
+                fake_mouth, fake_mouth_feats = self.net_d_mouth(self.mouths, return_feats=True)
+                l_g_gan = self.cri_component(fake_mouth, True, is_disc=False)
+                l_g_total += l_g_gan
+                loss_dict['l_g_gan_mouth'] = l_g_gan
+
+                if self.opt['train'].get('comp_style_weight', 0) > 0:
+                    # get gt feat
+                    _, real_left_eye_feats = self.net_d_left_eye(self.left_eyes_gt, return_feats=True)
+                    _, real_right_eye_feats = self.net_d_right_eye(self.right_eyes_gt, return_feats=True)
+                    _, real_mouth_feats = self.net_d_mouth(self.mouths_gt, return_feats=True)
+
+                    def _comp_style(feat, feat_gt, criterion):
+                        return criterion(self._gram_mat(feat[0]), self._gram_mat(
+                            feat_gt[0].detach())) * 0.5 + criterion(
+                                self._gram_mat(feat[1]), self._gram_mat(feat_gt[1].detach()))
+
+                    # facial component style loss
+                    comp_style_loss = 0
+                    comp_style_loss += _comp_style(fake_left_eye_feats, real_left_eye_feats, self.cri_l1)
+                    comp_style_loss += _comp_style(fake_right_eye_feats, real_right_eye_feats, self.cri_l1)
+                    comp_style_loss += _comp_style(fake_mouth_feats, real_mouth_feats, self.cri_l1)
+                    comp_style_loss = comp_style_loss * self.opt['train']['comp_style_weight']
+                    l_g_total += comp_style_loss
+                    loss_dict['l_g_comp_style_loss'] = comp_style_loss
+
+            # identity loss
+            if self.use_identity:
+                identity_weight = self.opt['train']['identity_weight']
+                # get gray images and resize
+                out_gray = self.gray_resize_for_identity(self.output)
+                gt_gray = self.gray_resize_for_identity(self.gt)
+
+                identity_gt = self.network_identity(gt_gray).detach()
+                identity_out = self.network_identity(out_gray)
+                l_identity = self.cri_l1(identity_out, identity_gt) * identity_weight
+                l_g_total += l_identity
+                loss_dict['l_identity'] = l_identity
+
+            l_g_total.backward()
+            self.optimizer_g.step()
+
+        # EMA
+        self.model_ema(decay=0.5**(32 / (10 * 1000)))
+
+        # ----------- optimize net_d ----------- #
+        for p in self.net_d.parameters():
+            p.requires_grad = True
+        self.optimizer_d.zero_grad()
+        if self.use_facial_disc:
+            for p in self.net_d_left_eye.parameters():
+                p.requires_grad = True
+            for p in self.net_d_right_eye.parameters():
+                p.requires_grad = True
+            for p in self.net_d_mouth.parameters():
+                p.requires_grad = True
+            self.optimizer_d_left_eye.zero_grad()
+            self.optimizer_d_right_eye.zero_grad()
+            self.optimizer_d_mouth.zero_grad()
+
+        fake_d_pred = self.net_d(self.output.detach())
+        real_d_pred = self.net_d(self.gt)
+        l_d = self.cri_gan(real_d_pred, True, is_disc=True) + self.cri_gan(fake_d_pred, False, is_disc=True)
+        loss_dict['l_d'] = l_d
+        # In WGAN, real_score should be positive and fake_score should be negative
+        loss_dict['real_score'] = real_d_pred.detach().mean()
+        loss_dict['fake_score'] = fake_d_pred.detach().mean()
+        l_d.backward()
+
+        # regularization loss
+        if current_iter % self.net_d_reg_every == 0:
+            self.gt.requires_grad = True
+            real_pred = self.net_d(self.gt)
+            l_d_r1 = r1_penalty(real_pred, self.gt)
+            l_d_r1 = (self.r1_reg_weight / 2 * l_d_r1 * self.net_d_reg_every + 0 * real_pred[0])
+            loss_dict['l_d_r1'] = l_d_r1.detach().mean()
+            l_d_r1.backward()
+
+        self.optimizer_d.step()
+
+        # optimize facial component discriminators
+        if self.use_facial_disc:
+            # left eye
+            fake_d_pred, _ = self.net_d_left_eye(self.left_eyes.detach())
+            real_d_pred, _ = self.net_d_left_eye(self.left_eyes_gt)
+            l_d_left_eye = self.cri_component(
+                real_d_pred, True, is_disc=True) + self.cri_gan(
+                    fake_d_pred, False, is_disc=True)
+            loss_dict['l_d_left_eye'] = l_d_left_eye
+            l_d_left_eye.backward()
+            # right eye
+            fake_d_pred, _ = self.net_d_right_eye(self.right_eyes.detach())
+            real_d_pred, _ = self.net_d_right_eye(self.right_eyes_gt)
+            l_d_right_eye = self.cri_component(
+                real_d_pred, True, is_disc=True) + self.cri_gan(
+                    fake_d_pred, False, is_disc=True)
+            loss_dict['l_d_right_eye'] = l_d_right_eye
+            l_d_right_eye.backward()
+            # mouth
+            fake_d_pred, _ = self.net_d_mouth(self.mouths.detach())
+            real_d_pred, _ = self.net_d_mouth(self.mouths_gt)
+            l_d_mouth = self.cri_component(
+                real_d_pred, True, is_disc=True) + self.cri_gan(
+                    fake_d_pred, False, is_disc=True)
+            loss_dict['l_d_mouth'] = l_d_mouth
+            l_d_mouth.backward()
+
+            self.optimizer_d_left_eye.step()
+            self.optimizer_d_right_eye.step()
+            self.optimizer_d_mouth.step()
+
+        self.log_dict = self.reduce_loss_dict(loss_dict)
+
+    def test(self):
+        with torch.no_grad():
+            if hasattr(self, 'net_g_ema'):
+                self.net_g_ema.eval()
+                self.output, _ = self.net_g_ema(self.lq)
+            else:
+                logger = get_root_logger()
+                logger.warning('Do not have self.net_g_ema, use self.net_g.')
+                self.net_g.eval()
+                self.output, _ = self.net_g(self.lq)
+                self.net_g.train()
+
+    def dist_validation(self, dataloader, current_iter, tb_logger, save_img):
+        if self.opt['rank'] == 0:
+            self.nondist_validation(dataloader, current_iter, tb_logger, save_img)
+
+    def nondist_validation(self, dataloader, current_iter, tb_logger, save_img):
+        dataset_name = dataloader.dataset.opt['name']
+        with_metrics = self.opt['val'].get('metrics') is not None
+        use_pbar = self.opt['val'].get('pbar', False)
+
+        if with_metrics:
+            if not hasattr(self, 'metric_results'):  # only execute in the first run
+                self.metric_results = {metric: 0 for metric in self.opt['val']['metrics'].keys()}
+            # initialize the best metric results for each dataset_name (supporting multiple validation datasets)
+            self._initialize_best_metric_results(dataset_name)
+            # zero self.metric_results
+            self.metric_results = {metric: 0 for metric in self.metric_results}
+
+        metric_data = dict()
+        if use_pbar:
+            pbar = tqdm(total=len(dataloader), unit='image')
+
+        for idx, val_data in enumerate(dataloader):
+            img_name = osp.splitext(osp.basename(val_data['lq_path'][0]))[0]
+            self.feed_data(val_data)
+            self.test()
+
+            sr_img = tensor2img(self.output.detach().cpu(), min_max=(-1, 1))
+            metric_data['img'] = sr_img
+            if hasattr(self, 'gt'):
+                gt_img = tensor2img(self.gt.detach().cpu(), min_max=(-1, 1))
+                metric_data['img2'] = gt_img
+                del self.gt
+
+            # tentative for out of GPU memory
+            del self.lq
+            del self.output
+            torch.cuda.empty_cache()
+
+            if save_img:
+                if self.opt['is_train']:
+                    save_img_path = osp.join(self.opt['path']['visualization'], img_name,
+                                             f'{img_name}_{current_iter}.png')
+                else:
+                    if self.opt['val']['suffix']:
+                        save_img_path = osp.join(self.opt['path']['visualization'], dataset_name,
+                                                 f'{img_name}_{self.opt["val"]["suffix"]}.png')
+                    else:
+                        save_img_path = osp.join(self.opt['path']['visualization'], dataset_name,
+                                                 f'{img_name}_{self.opt["name"]}.png')
+                imwrite(sr_img, save_img_path)
+
+            if with_metrics:
+                # calculate metrics
+                for name, opt_ in self.opt['val']['metrics'].items():
+                    self.metric_results[name] += calculate_metric(metric_data, opt_)
+            if use_pbar:
+                pbar.update(1)
+                pbar.set_description(f'Test {img_name}')
+        if use_pbar:
+            pbar.close()
+
+        if with_metrics:
+            for metric in self.metric_results.keys():
+                self.metric_results[metric] /= (idx + 1)
+                # update the best metric result
+                self._update_best_metric_result(dataset_name, metric, self.metric_results[metric], current_iter)
+
+            self._log_validation_metric_values(current_iter, dataset_name, tb_logger)
+
+    def _log_validation_metric_values(self, current_iter, dataset_name, tb_logger):
+        log_str = f'Validation {dataset_name}\n'
+        for metric, value in self.metric_results.items():
+            log_str += f'\t # {metric}: {value:.4f}'
+            if hasattr(self, 'best_metric_results'):
+                log_str += (f'\tBest: {self.best_metric_results[dataset_name][metric]["val"]:.4f} @ '
+                            f'{self.best_metric_results[dataset_name][metric]["iter"]} iter')
+            log_str += '\n'
+
+        logger = get_root_logger()
+        logger.info(log_str)
+        if tb_logger:
+            for metric, value in self.metric_results.items():
+                tb_logger.add_scalar(f'metrics/{dataset_name}/{metric}', value, current_iter)
+
+    def save(self, epoch, current_iter):
+        # save net_g and net_d
+        self.save_network([self.net_g, self.net_g_ema], 'net_g', current_iter, param_key=['params', 'params_ema'])
+        self.save_network(self.net_d, 'net_d', current_iter)
+        # save component discriminators
+        if self.use_facial_disc:
+            self.save_network(self.net_d_left_eye, 'net_d_left_eye', current_iter)
+            self.save_network(self.net_d_right_eye, 'net_d_right_eye', current_iter)
+            self.save_network(self.net_d_mouth, 'net_d_mouth', current_iter)
+        # save training state
+        self.save_training_state(epoch, current_iter)

gfpgan/train.py

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

gfpgan/utils.py

@@ -0,0 +1,148 @@
+import cv2
+import os
+import torch
+from basicsr.utils import img2tensor, tensor2img
+from basicsr.utils.download_util import load_file_from_url
+from facexlib.utils.face_restoration_helper import FaceRestoreHelper
+from torchvision.transforms.functional import normalize
+
+from gfpgan.archs.gfpgan_bilinear_arch import GFPGANBilinear
+from gfpgan.archs.gfpganv1_arch import GFPGANv1
+from gfpgan.archs.gfpganv1_clean_arch import GFPGANv1Clean
+
+ROOT_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+
+
+class GFPGANer():
+    """Helper for restoration with GFPGAN.
+
+    It will detect and crop faces, and then resize the faces to 512x512.
+    GFPGAN is used to restored the resized faces.
+    The background is upsampled with the bg_upsampler.
+    Finally, the faces will be pasted back to the upsample background image.
+
+    Args:
+        model_path (str): The path to the GFPGAN model. It can be urls (will first download it automatically).
+        upscale (float): The upscale of the final output. Default: 2.
+        arch (str): The GFPGAN architecture. Option: clean | original. Default: clean.
+        channel_multiplier (int): Channel multiplier for large networks of StyleGAN2. Default: 2.
+        bg_upsampler (nn.Module): The upsampler for the background. Default: None.
+    """
+
+    def __init__(self, model_path, upscale=2, arch='clean', channel_multiplier=2, bg_upsampler=None, device=None):
+        self.upscale = upscale
+        self.bg_upsampler = bg_upsampler
+
+        # initialize model
+        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') if device is None else device
+        # initialize the GFP-GAN
+        if arch == 'clean':
+            self.gfpgan = GFPGANv1Clean(
+                out_size=512,
+                num_style_feat=512,
+                channel_multiplier=channel_multiplier,
+                decoder_load_path=None,
+                fix_decoder=False,
+                num_mlp=8,
+                input_is_latent=True,
+                different_w=True,
+                narrow=1,
+                sft_half=True)
+        elif arch == 'bilinear':
+            self.gfpgan = GFPGANBilinear(
+                out_size=512,
+                num_style_feat=512,
+                channel_multiplier=channel_multiplier,
+                decoder_load_path=None,
+                fix_decoder=False,
+                num_mlp=8,
+                input_is_latent=True,
+                different_w=True,
+                narrow=1,
+                sft_half=True)
+        elif arch == 'original':
+            self.gfpgan = GFPGANv1(
+                out_size=512,
+                num_style_feat=512,
+                channel_multiplier=channel_multiplier,
+                decoder_load_path=None,
+                fix_decoder=True,
+                num_mlp=8,
+                input_is_latent=True,
+                different_w=True,
+                narrow=1,
+                sft_half=True)
+        elif arch == 'RestoreFormer':
+            from gfpgan.archs.restoreformer_arch import RestoreFormer
+            self.gfpgan = RestoreFormer()
+        # initialize face helper
+        self.face_helper = FaceRestoreHelper(
+            upscale,
+            face_size=512,
+            crop_ratio=(1, 1),
+            det_model='retinaface_resnet50',
+            save_ext='png',
+            use_parse=True,
+            device=self.device,
+            model_rootpath='gfpgan/weights')
+
+        if model_path.startswith('https://'):
+            model_path = load_file_from_url(
+                url=model_path, model_dir=os.path.join(ROOT_DIR, 'gfpgan/weights'), progress=True, file_name=None)
+        loadnet = torch.load(model_path)
+        if 'params_ema' in loadnet:
+            keyname = 'params_ema'
+        else:
+            keyname = 'params'
+        self.gfpgan.load_state_dict(loadnet[keyname], strict=True)
+        self.gfpgan.eval()
+        self.gfpgan = self.gfpgan.to(self.device)
+
+    @torch.no_grad()
+    def enhance(self, img, has_aligned=False, only_center_face=False, paste_back=True, weight=0.5):
+        self.face_helper.clean_all()
+
+        if has_aligned:  # the inputs are already aligned
+            img = cv2.resize(img, (512, 512))
+            self.face_helper.cropped_faces = [img]
+        else:
+            self.face_helper.read_image(img)
+            # get face landmarks for each face
+            self.face_helper.get_face_landmarks_5(only_center_face=only_center_face, eye_dist_threshold=5)
+            # eye_dist_threshold=5: skip faces whose eye distance is smaller than 5 pixels
+            # TODO: even with eye_dist_threshold, it will still introduce wrong detections and restorations.
+            # align and warp each face
+            self.face_helper.align_warp_face()
+
+        # face restoration
+        for cropped_face in self.face_helper.cropped_faces:
+            # prepare data
+            cropped_face_t = img2tensor(cropped_face / 255., bgr2rgb=True, float32=True)
+            normalize(cropped_face_t, (0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True)
+            cropped_face_t = cropped_face_t.unsqueeze(0).to(self.device)
+
+            try:
+                output = self.gfpgan(cropped_face_t, return_rgb=False, weight=weight)[0]
+                # convert to image
+                restored_face = tensor2img(output.squeeze(0), rgb2bgr=True, min_max=(-1, 1))
+            except RuntimeError as error:
+                print(f'\tFailed inference for GFPGAN: {error}.')
+                restored_face = cropped_face
+
+            restored_face = restored_face.astype('uint8')
+            self.face_helper.add_restored_face(restored_face)
+
+        if not has_aligned and paste_back:
+            # upsample the background
+            if self.bg_upsampler is not None:
+                # Now only support RealESRGAN for upsampling background
+                bg_img = self.bg_upsampler.enhance(img, outscale=self.upscale)[0]
+            else:
+                bg_img = None
+
+            self.face_helper.get_inverse_affine(None)
+            # paste each restored face to the input image
+            restored_img = self.face_helper.paste_faces_to_input_image(upsample_img=bg_img)
+            return self.face_helper.cropped_faces, self.face_helper.restored_faces, restored_img
+        else:
+            return self.face_helper.cropped_faces, self.face_helper.restored_faces, None

gfpgan/version.py

@@ -0,0 +1,5 @@
+# GENERATED VERSION FILE
+# TIME: Sat May 25 17:25:12 2024
+__version__ = '1.3.8'
+__gitsha__ = '7552a77'
+version_info = (1, 3, 8)

gfpgan/weights/README.md

@@ -0,0 +1,3 @@
+# Weights
+
+Put the downloaded weights to this folder.

gfpgan/weights/detection_Resnet50_Final.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6d1de9c2944f2ccddca5f5e010ea5ae64a39845a86311af6fdf30841b0a5a16d
+size 109497761

gfpgan/weights/parsing_parsenet.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d558d8d0e42c20224f13cf5a29c79eba2d59913419f945545d8cf7b72920de2
+size 85331193