git init

basicsr/archs/spynet_arch.py

+import math
+import torch
+from torch import nn as nn
+from torch.nn import functional as F
+
+from basicsr.utils.registry import ARCH_REGISTRY
+from .arch_util import flow_warp
+
+
+class BasicModule(nn.Module):
+    """Basic Module for SpyNet.
+    """
+
+    def __init__(self):
+        super(BasicModule, self).__init__()
+
+        self.basic_module = nn.Sequential(
+            nn.Conv2d(in_channels=8, out_channels=32, kernel_size=7, stride=1, padding=3), nn.ReLU(inplace=False),
+            nn.Conv2d(in_channels=32, out_channels=64, kernel_size=7, stride=1, padding=3), nn.ReLU(inplace=False),
+            nn.Conv2d(in_channels=64, out_channels=32, kernel_size=7, stride=1, padding=3), nn.ReLU(inplace=False),
+            nn.Conv2d(in_channels=32, out_channels=16, kernel_size=7, stride=1, padding=3), nn.ReLU(inplace=False),
+            nn.Conv2d(in_channels=16, out_channels=2, kernel_size=7, stride=1, padding=3))
+
+    def forward(self, tensor_input):
+        return self.basic_module(tensor_input)
+
+
+@ARCH_REGISTRY.register()
+class SpyNet(nn.Module):
+    """SpyNet architecture.
+
+    Args:
+        load_path (str): path for pretrained SpyNet. Default: None.
+    """
+
+    def __init__(self, load_path=None):
+        super(SpyNet, self).__init__()
+        self.basic_module = nn.ModuleList([BasicModule() for _ in range(6)])
+        if load_path:
+            self.load_state_dict(torch.load(load_path, map_location=lambda storage, loc: storage)['params'])
+
+        self.register_buffer('mean', torch.Tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1))
+        self.register_buffer('std', torch.Tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1))
+
+    def preprocess(self, tensor_input):
+        tensor_output = (tensor_input - self.mean) / self.std
+        return tensor_output
+
+    def process(self, ref, supp):
+        flow = []
+
+        ref = [self.preprocess(ref)]
+        supp = [self.preprocess(supp)]
+
+        for level in range(5):
+            ref.insert(0, F.avg_pool2d(input=ref[0], kernel_size=2, stride=2, count_include_pad=False))
+            supp.insert(0, F.avg_pool2d(input=supp[0], kernel_size=2, stride=2, count_include_pad=False))
+
+        flow = ref[0].new_zeros(
+            [ref[0].size(0), 2,
+             int(math.floor(ref[0].size(2) / 2.0)),
+             int(math.floor(ref[0].size(3) / 2.0))])
+
+        for level in range(len(ref)):
+            upsampled_flow = F.interpolate(input=flow, scale_factor=2, mode='bilinear', align_corners=True) * 2.0
+
+            if upsampled_flow.size(2) != ref[level].size(2):
+                upsampled_flow = F.pad(input=upsampled_flow, pad=[0, 0, 0, 1], mode='replicate')
+            if upsampled_flow.size(3) != ref[level].size(3):
+                upsampled_flow = F.pad(input=upsampled_flow, pad=[0, 1, 0, 0], mode='replicate')
+
+            flow = self.basic_module[level](torch.cat([
+                ref[level],
+                flow_warp(
+                    supp[level], upsampled_flow.permute(0, 2, 3, 1), interp_mode='bilinear', padding_mode='border'),
+                upsampled_flow
+            ], 1)) + upsampled_flow
+
+        return flow
+
+    def forward(self, ref, supp):
+        assert ref.size() == supp.size()
+
+        h, w = ref.size(2), ref.size(3)
+        w_floor = math.floor(math.ceil(w / 32.0) * 32.0)
+        h_floor = math.floor(math.ceil(h / 32.0) * 32.0)
+
+        ref = F.interpolate(input=ref, size=(h_floor, w_floor), mode='bilinear', align_corners=False)
+        supp = F.interpolate(input=supp, size=(h_floor, w_floor), mode='bilinear', align_corners=False)
+
+        flow = F.interpolate(input=self.process(ref, supp), size=(h, w), mode='bilinear', align_corners=False)
+
+        flow[:, 0, :, :] *= float(w) / float(w_floor)
+        flow[:, 1, :, :] *= float(h) / float(h_floor)
+
+        return flow

basicsr/archs/srresnet_arch.py

+from torch import nn as nn
+from torch.nn import functional as F
+
+from basicsr.utils.registry import ARCH_REGISTRY
+from .arch_util import ResidualBlockNoBN, default_init_weights, make_layer
+
+
+@ARCH_REGISTRY.register()
+class MSRResNet(nn.Module):
+    """Modified SRResNet.
+
+    A compacted version modified from SRResNet in
+    "Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network"
+    It uses residual blocks without BN, similar to EDSR.
+    Currently, it supports x2, x3 and x4 upsampling scale factor.
+
+    Args:
+        num_in_ch (int): Channel number of inputs. Default: 3.
+        num_out_ch (int): Channel number of outputs. Default: 3.
+        num_feat (int): Channel number of intermediate features. Default: 64.
+        num_block (int): Block number in the body network. Default: 16.
+        upscale (int): Upsampling factor. Support x2, x3 and x4. Default: 4.
+    """
+
+    def __init__(self, num_in_ch=3, num_out_ch=3, num_feat=64, num_block=16, upscale=4):
+        super(MSRResNet, self).__init__()
+        self.upscale = upscale
+
+        self.conv_first = nn.Conv2d(num_in_ch, num_feat, 3, 1, 1)
+        self.body = make_layer(ResidualBlockNoBN, num_block, num_feat=num_feat)
+
+        # upsampling
+        if self.upscale in [2, 3]:
+            self.upconv1 = nn.Conv2d(num_feat, num_feat * self.upscale * self.upscale, 3, 1, 1)
+            self.pixel_shuffle = nn.PixelShuffle(self.upscale)
+        elif self.upscale == 4:
+            self.upconv1 = nn.Conv2d(num_feat, num_feat * 4, 3, 1, 1)
+            self.upconv2 = nn.Conv2d(num_feat, num_feat * 4, 3, 1, 1)
+            self.pixel_shuffle = nn.PixelShuffle(2)
+
+        self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+        self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+
+        # activation function
+        self.lrelu = nn.LeakyReLU(negative_slope=0.1, inplace=True)
+
+        # initialization
+        default_init_weights([self.conv_first, self.upconv1, self.conv_hr, self.conv_last], 0.1)
+        if self.upscale == 4:
+            default_init_weights(self.upconv2, 0.1)
+
+    def forward(self, x):
+        feat = self.lrelu(self.conv_first(x))
+        out = self.body(feat)
+
+        if self.upscale == 4:
+            out = self.lrelu(self.pixel_shuffle(self.upconv1(out)))
+            out = self.lrelu(self.pixel_shuffle(self.upconv2(out)))
+        elif self.upscale in [2, 3]:
+            out = self.lrelu(self.pixel_shuffle(self.upconv1(out)))
+
+        out = self.conv_last(self.lrelu(self.conv_hr(out)))
+        base = F.interpolate(x, scale_factor=self.upscale, mode='bilinear', align_corners=False)
+        out += base
+        return out

basicsr/archs/srvgg_arch.py

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

basicsr/archs/stylegan2_arch.py

+import math
+import random
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from basicsr.ops.fused_act import FusedLeakyReLU, fused_leaky_relu
+from basicsr.ops.upfirdn2d import upfirdn2d
+from basicsr.utils.registry import ARCH_REGISTRY
+
+
+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)
+
+
+def make_resample_kernel(k):
+    """Make resampling kernel for UpFirDn.
+
+    Args:
+        k (list[int]): A list indicating the 1D resample kernel magnitude.
+
+    Returns:
+        Tensor: 2D resampled kernel.
+    """
+    k = torch.tensor(k, dtype=torch.float32)
+    if k.ndim == 1:
+        k = k[None, :] * k[:, None]  # to 2D kernel, outer product
+    # normalize
+    k /= k.sum()
+    return k
+
+
+class UpFirDnUpsample(nn.Module):
+    """Upsample, FIR filter, and downsample (upsampole version).
+
+    References:
+    1. https://docs.scipy.org/doc/scipy/reference/generated/scipy.signal.upfirdn.html  # noqa: E501
+    2. http://www.ece.northwestern.edu/local-apps/matlabhelp/toolbox/signal/upfirdn.html  # noqa: E501
+
+    Args:
+        resample_kernel (list[int]): A list indicating the 1D resample kernel
+            magnitude.
+        factor (int): Upsampling scale factor. Default: 2.
+    """
+
+    def __init__(self, resample_kernel, factor=2):
+        super(UpFirDnUpsample, self).__init__()
+        self.kernel = make_resample_kernel(resample_kernel) * (factor**2)
+        self.factor = factor
+
+        pad = self.kernel.shape[0] - factor
+        self.pad = ((pad + 1) // 2 + factor - 1, pad // 2)
+
+    def forward(self, x):
+        out = upfirdn2d(x, self.kernel.type_as(x), up=self.factor, down=1, pad=self.pad)
+        return out
+
+    def __repr__(self):
+        return (f'{self.__class__.__name__}(factor={self.factor})')
+
+
+class UpFirDnDownsample(nn.Module):
+    """Upsample, FIR filter, and downsample (downsampole version).
+
+    Args:
+        resample_kernel (list[int]): A list indicating the 1D resample kernel
+            magnitude.
+        factor (int): Downsampling scale factor. Default: 2.
+    """
+
+    def __init__(self, resample_kernel, factor=2):
+        super(UpFirDnDownsample, self).__init__()
+        self.kernel = make_resample_kernel(resample_kernel)
+        self.factor = factor
+
+        pad = self.kernel.shape[0] - factor
+        self.pad = ((pad + 1) // 2, pad // 2)
+
+    def forward(self, x):
+        out = upfirdn2d(x, self.kernel.type_as(x), up=1, down=self.factor, pad=self.pad)
+        return out
+
+    def __repr__(self):
+        return (f'{self.__class__.__name__}(factor={self.factor})')
+
+
+class UpFirDnSmooth(nn.Module):
+    """Upsample, FIR filter, and downsample (smooth version).
+
+    Args:
+        resample_kernel (list[int]): A list indicating the 1D resample kernel
+            magnitude.
+        upsample_factor (int): Upsampling scale factor. Default: 1.
+        downsample_factor (int): Downsampling scale factor. Default: 1.
+        kernel_size (int): Kernel size: Default: 1.
+    """
+
+    def __init__(self, resample_kernel, upsample_factor=1, downsample_factor=1, kernel_size=1):
+        super(UpFirDnSmooth, self).__init__()
+        self.upsample_factor = upsample_factor
+        self.downsample_factor = downsample_factor
+        self.kernel = make_resample_kernel(resample_kernel)
+        if upsample_factor > 1:
+            self.kernel = self.kernel * (upsample_factor**2)
+
+        if upsample_factor > 1:
+            pad = (self.kernel.shape[0] - upsample_factor) - (kernel_size - 1)
+            self.pad = ((pad + 1) // 2 + upsample_factor - 1, pad // 2 + 1)
+        elif downsample_factor > 1:
+            pad = (self.kernel.shape[0] - downsample_factor) + (kernel_size - 1)
+            self.pad = ((pad + 1) // 2, pad // 2)
+        else:
+            raise NotImplementedError
+
+    def forward(self, x):
+        out = upfirdn2d(x, self.kernel.type_as(x), up=1, down=1, pad=self.pad)
+        return out
+
+    def __repr__(self):
+        return (f'{self.__class__.__name__}(upsample_factor={self.upsample_factor}'
+                f', downsample_factor={self.downsample_factor})')
+
+
+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.
+        resample_kernel (list[int]): A list indicating the 1D resample kernel
+            magnitude. Default: (1, 3, 3, 1).
+        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,
+                 resample_kernel=(1, 3, 3, 1),
+                 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
+
+        if self.sample_mode == 'upsample':
+            self.smooth = UpFirDnSmooth(
+                resample_kernel, upsample_factor=2, downsample_factor=1, kernel_size=kernel_size)
+        elif self.sample_mode == 'downsample':
+            self.smooth = UpFirDnSmooth(
+                resample_kernel, upsample_factor=1, downsample_factor=2, kernel_size=kernel_size)
+        elif self.sample_mode is None:
+            pass
+        else:
+            raise ValueError(f'Wrong sample mode {self.sample_mode}, '
+                             "supported ones are ['upsample', 'downsample', None].")
+
+        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 = x.view(1, b * c, h, w)
+            weight = weight.view(b, self.out_channels, c, self.kernel_size, self.kernel_size)
+            weight = weight.transpose(1, 2).reshape(b * c, self.out_channels, self.kernel_size, self.kernel_size)
+            out = F.conv_transpose2d(x, weight, padding=0, stride=2, groups=b)
+            out = out.view(b, self.out_channels, *out.shape[2:4])
+            out = self.smooth(out)
+        elif self.sample_mode == 'downsample':
+            x = self.smooth(x)
+            x = x.view(1, b * c, *x.shape[2:4])
+            out = F.conv2d(x, weight, padding=0, stride=2, groups=b)
+            out = out.view(b, self.out_channels, *out.shape[2:4])
+        else:
+            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.
+        resample_kernel (list[int]): A list indicating the 1D resample kernel
+            magnitude. Default: (1, 3, 3, 1).
+    """
+
+    def __init__(self,
+                 in_channels,
+                 out_channels,
+                 kernel_size,
+                 num_style_feat,
+                 demodulate=True,
+                 sample_mode=None,
+                 resample_kernel=(1, 3, 3, 1)):
+        super(StyleConv, self).__init__()
+        self.modulated_conv = ModulatedConv2d(
+            in_channels,
+            out_channels,
+            kernel_size,
+            num_style_feat,
+            demodulate=demodulate,
+            sample_mode=sample_mode,
+            resample_kernel=resample_kernel)
+        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.
+        resample_kernel (list[int]): A list indicating the 1D resample kernel
+            magnitude. Default: (1, 3, 3, 1).
+    """
+
+    def __init__(self, in_channels, num_style_feat, upsample=True, resample_kernel=(1, 3, 3, 1)):
+        super(ToRGB, self).__init__()
+        if upsample:
+            self.upsample = UpFirDnUpsample(resample_kernel, factor=2)
+        else:
+            self.upsample = None
+        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 = self.upsample(skip)
+            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 StyleGAN2Generator(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.
+        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): Narrow ratio for channels. Default: 1.0.
+    """
+
+    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):
+        super(StyleGAN2Generator, 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,
+            resample_kernel=resample_kernel)
+        self.to_rgb1 = ToRGB(channels['4'], num_style_feat, upsample=False, resample_kernel=resample_kernel)
+
+        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',
+                    resample_kernel=resample_kernel,
+                ))
+            self.style_convs.append(
+                StyleConv(
+                    out_channels,
+                    out_channels,
+                    kernel_size=3,
+                    num_style_feat=num_style_feat,
+                    demodulate=True,
+                    sample_mode=None,
+                    resample_kernel=resample_kernel))
+            self.to_rgbs.append(ToRGB(out_channels, num_style_feat, upsample=True, resample_kernel=resample_kernel))
+            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.
+        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).
+        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,
+                 resample_kernel=(1, 3, 3, 1),
+                 bias=True,
+                 activate=True):
+        layers = []
+        # downsample
+        if downsample:
+            layers.append(
+                UpFirDnSmooth(resample_kernel, upsample_factor=1, downsample_factor=2, kernel_size=kernel_size))
+            stride = 2
+            self.padding = 0
+        else:
+            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.
+        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).
+    """
+
+    def __init__(self, in_channels, out_channels, resample_kernel=(1, 3, 3, 1)):
+        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, resample_kernel=resample_kernel, bias=True, activate=True)
+        self.skip = ConvLayer(
+            in_channels, out_channels, 1, downsample=True, resample_kernel=resample_kernel, 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 StyleGAN2Discriminator(nn.Module):
+    """StyleGAN2 Discriminator.
+
+    Args:
+        out_size (int): The spatial size of outputs.
+        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).
+        stddev_group (int): For group stddev statistics. Default: 4.
+        narrow (float): Narrow ratio for channels. Default: 1.0.
+    """
+
+    def __init__(self, out_size, channel_multiplier=2, resample_kernel=(1, 3, 3, 1), stddev_group=4, narrow=1):
+        super(StyleGAN2Discriminator, self).__init__()
+
+        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)
+        }
+
+        log_size = int(math.log(out_size, 2))
+
+        conv_body = [ConvLayer(3, channels[f'{out_size}'], 1, bias=True, activate=True)]
+
+        in_channels = channels[f'{out_size}']
+        for i in range(log_size, 2, -1):
+            out_channels = channels[f'{2**(i - 1)}']
+            conv_body.append(ResBlock(in_channels, out_channels, resample_kernel))
+            in_channels = out_channels
+        self.conv_body = nn.Sequential(*conv_body)
+
+        self.final_conv = ConvLayer(in_channels + 1, channels['4'], 3, bias=True, activate=True)
+        self.final_linear = nn.Sequential(
+            EqualLinear(
+                channels['4'] * 4 * 4, channels['4'], bias=True, bias_init_val=0, lr_mul=1, activation='fused_lrelu'),
+            EqualLinear(channels['4'], 1, bias=True, bias_init_val=0, lr_mul=1, activation=None),
+        )
+        self.stddev_group = stddev_group
+        self.stddev_feat = 1
+
+    def forward(self, x):
+        out = self.conv_body(x)
+
+        b, c, h, w = out.shape
+        # concatenate a group stddev statistics to out
+        group = min(b, self.stddev_group)  # Minibatch must be divisible by (or smaller than) group_size
+        stddev = out.view(group, -1, self.stddev_feat, c // self.stddev_feat, h, w)
+        stddev = torch.sqrt(stddev.var(0, unbiased=False) + 1e-8)
+        stddev = stddev.mean([2, 3, 4], keepdims=True).squeeze(2)
+        stddev = stddev.repeat(group, 1, h, w)
+        out = torch.cat([out, stddev], 1)
+
+        out = self.final_conv(out)
+        out = out.view(b, -1)
+        out = self.final_linear(out)
+
+        return out

basicsr/archs/stylegan2_bilinear_arch.py

+import math
+import random
+import torch
+from torch import nn
+from torch.nn import functional as F
+
+from basicsr.ops.fused_act import FusedLeakyReLU, fused_leaky_relu
+from basicsr.utils.registry import ARCH_REGISTRY
+
+
+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(suffix='basicsr')
+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

basicsr/archs/swinir_arch.py

+# Modified from https://github.com/JingyunLiang/SwinIR
+# SwinIR: Image Restoration Using Swin Transformer, https://arxiv.org/abs/2108.10257
+# Originally Written by Ze Liu, Modified by Jingyun Liang.
+
+import math
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint as checkpoint
+
+from basicsr.utils.registry import ARCH_REGISTRY
+from .arch_util import to_2tuple, trunc_normal_
+
+
+def drop_path(x, drop_prob: float = 0., training: bool = False):
+    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).
+
+    From: https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/drop.py
+    """
+    if drop_prob == 0. or not training:
+        return x
+    keep_prob = 1 - drop_prob
+    shape = (x.shape[0], ) + (1, ) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
+    random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)
+    random_tensor.floor_()  # binarize
+    output = x.div(keep_prob) * random_tensor
+    return output
+
+
+class DropPath(nn.Module):
+    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
+
+    From: https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/layers/drop.py
+    """
+
+    def __init__(self, drop_prob=None):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+
+    def forward(self, x):
+        return drop_path(x, self.drop_prob, self.training)
+
+
+class Mlp(nn.Module):
+
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+def window_partition(x, window_size):
+    """
+    Args:
+        x: (b, h, w, c)
+        window_size (int): window size
+
+    Returns:
+        windows: (num_windows*b, window_size, window_size, c)
+    """
+    b, h, w, c = x.shape
+    x = x.view(b, h // window_size, window_size, w // window_size, window_size, c)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, c)
+    return windows
+
+
+def window_reverse(windows, window_size, h, w):
+    """
+    Args:
+        windows: (num_windows*b, window_size, window_size, c)
+        window_size (int): Window size
+        h (int): Height of image
+        w (int): Width of image
+
+    Returns:
+        x: (b, h, w, c)
+    """
+    b = int(windows.shape[0] / (h * w / window_size / window_size))
+    x = windows.view(b, h // window_size, w // window_size, window_size, window_size, -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(b, h, w, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    r""" Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+    """
+
+    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        # define a parameter table of relative position bias
+        self.relative_position_bias_table = nn.Parameter(
+            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer('relative_position_index', relative_position_index)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+
+        self.proj_drop = nn.Dropout(proj_drop)
+
+        trunc_normal_(self.relative_position_bias_table, std=.02)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask=None):
+        """
+        Args:
+            x: input features with shape of (num_windows*b, n, c)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        b_, n, c = x.shape
+        qkv = self.qkv(x).reshape(b_, n, 3, self.num_heads, c // self.num_heads).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)
+
+        q = q * self.scale
+        attn = (q @ k.transpose(-2, -1))
+
+        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
+            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nw = mask.shape[0]
+            attn = attn.view(b_ // nw, nw, self.num_heads, n, n) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, n, n)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(b_, n, c)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+    def extra_repr(self) -> str:
+        return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}'
+
+    def flops(self, n):
+        # calculate flops for 1 window with token length of n
+        flops = 0
+        # qkv = self.qkv(x)
+        flops += n * self.dim * 3 * self.dim
+        # attn = (q @ k.transpose(-2, -1))
+        flops += self.num_heads * n * (self.dim // self.num_heads) * n
+        #  x = (attn @ v)
+        flops += self.num_heads * n * n * (self.dim // self.num_heads)
+        # x = self.proj(x)
+        flops += n * self.dim * self.dim
+        return flops
+
+
+class SwinTransformerBlock(nn.Module):
+    r""" Swin Transformer Block.
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self,
+                 dim,
+                 input_resolution,
+                 num_heads,
+                 window_size=7,
+                 shift_size=0,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop=0.,
+                 attn_drop=0.,
+                 drop_path=0.,
+                 act_layer=nn.GELU,
+                 norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.num_heads = num_heads
+        self.window_size = window_size
+        self.shift_size = shift_size
+        self.mlp_ratio = mlp_ratio
+        if min(self.input_resolution) <= self.window_size:
+            # if window size is larger than input resolution, we don't partition windows
+            self.shift_size = 0
+            self.window_size = min(self.input_resolution)
+        assert 0 <= self.shift_size < self.window_size, 'shift_size must in 0-window_size'
+
+        self.norm1 = norm_layer(dim)
+        self.attn = WindowAttention(
+            dim,
+            window_size=to_2tuple(self.window_size),
+            num_heads=num_heads,
+            qkv_bias=qkv_bias,
+            qk_scale=qk_scale,
+            attn_drop=attn_drop,
+            proj_drop=drop)
+
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+        self.norm2 = norm_layer(dim)
+        mlp_hidden_dim = int(dim * mlp_ratio)
+        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)
+
+        if self.shift_size > 0:
+            attn_mask = self.calculate_mask(self.input_resolution)
+        else:
+            attn_mask = None
+
+        self.register_buffer('attn_mask', attn_mask)
+
+    def calculate_mask(self, x_size):
+        # calculate attention mask for SW-MSA
+        h, w = x_size
+        img_mask = torch.zeros((1, h, w, 1))  # 1 h w 1
+        h_slices = (slice(0, -self.window_size), slice(-self.window_size,
+                                                       -self.shift_size), slice(-self.shift_size, None))
+        w_slices = (slice(0, -self.window_size), slice(-self.window_size,
+                                                       -self.shift_size), slice(-self.shift_size, None))
+        cnt = 0
+        for h in h_slices:
+            for w in w_slices:
+                img_mask[:, h, w, :] = cnt
+                cnt += 1
+
+        mask_windows = window_partition(img_mask, self.window_size)  # nw, window_size, window_size, 1
+        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
+        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+
+        return attn_mask
+
+    def forward(self, x, x_size):
+        h, w = x_size
+        b, _, c = x.shape
+        # assert seq_len == h * w, "input feature has wrong size"
+
+        shortcut = x
+        x = self.norm1(x)
+        x = x.view(b, h, w, c)
+
+        # cyclic shift
+        if self.shift_size > 0:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
+        else:
+            shifted_x = x
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nw*b, window_size, window_size, c
+        x_windows = x_windows.view(-1, self.window_size * self.window_size, c)  # nw*b, window_size*window_size, c
+
+        # W-MSA/SW-MSA (to be compatible for testing on images whose shapes are the multiple of window size
+        if self.input_resolution == x_size:
+            attn_windows = self.attn(x_windows, mask=self.attn_mask)  # nw*b, window_size*window_size, c
+        else:
+            attn_windows = self.attn(x_windows, mask=self.calculate_mask(x_size).to(x.device))
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, c)
+        shifted_x = window_reverse(attn_windows, self.window_size, h, w)  # b h' w' c
+
+        # reverse cyclic shift
+        if self.shift_size > 0:
+            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
+        else:
+            x = shifted_x
+        x = x.view(b, h * w, c)
+
+        # FFN
+        x = shortcut + self.drop_path(x)
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+
+        return x
+
+    def extra_repr(self) -> str:
+        return (f'dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, '
+                f'window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}')
+
+    def flops(self):
+        flops = 0
+        h, w = self.input_resolution
+        # norm1
+        flops += self.dim * h * w
+        # W-MSA/SW-MSA
+        nw = h * w / self.window_size / self.window_size
+        flops += nw * self.attn.flops(self.window_size * self.window_size)
+        # mlp
+        flops += 2 * h * w * self.dim * self.dim * self.mlp_ratio
+        # norm2
+        flops += self.dim * h * w
+        return flops
+
+
+class PatchMerging(nn.Module):
+    r""" Patch Merging Layer.
+
+    Args:
+        input_resolution (tuple[int]): Resolution of input feature.
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(4 * dim)
+
+    def forward(self, x):
+        """
+        x: b, h*w, c
+        """
+        h, w = self.input_resolution
+        b, seq_len, c = x.shape
+        assert seq_len == h * w, 'input feature has wrong size'
+        assert h % 2 == 0 and w % 2 == 0, f'x size ({h}*{w}) are not even.'
+
+        x = x.view(b, h, w, c)
+
+        x0 = x[:, 0::2, 0::2, :]  # b h/2 w/2 c
+        x1 = x[:, 1::2, 0::2, :]  # b h/2 w/2 c
+        x2 = x[:, 0::2, 1::2, :]  # b h/2 w/2 c
+        x3 = x[:, 1::2, 1::2, :]  # b h/2 w/2 c
+        x = torch.cat([x0, x1, x2, x3], -1)  # b h/2 w/2 4*c
+        x = x.view(b, -1, 4 * c)  # b h/2*w/2 4*c
+
+        x = self.norm(x)
+        x = self.reduction(x)
+
+        return x
+
+    def extra_repr(self) -> str:
+        return f'input_resolution={self.input_resolution}, dim={self.dim}'
+
+    def flops(self):
+        h, w = self.input_resolution
+        flops = h * w * self.dim
+        flops += (h // 2) * (w // 2) * 4 * self.dim * 2 * self.dim
+        return flops
+
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+    """
+
+    def __init__(self,
+                 dim,
+                 input_resolution,
+                 depth,
+                 num_heads,
+                 window_size,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop=0.,
+                 attn_drop=0.,
+                 drop_path=0.,
+                 norm_layer=nn.LayerNorm,
+                 downsample=None,
+                 use_checkpoint=False):
+
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.depth = depth
+        self.use_checkpoint = use_checkpoint
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(
+                dim=dim,
+                input_resolution=input_resolution,
+                num_heads=num_heads,
+                window_size=window_size,
+                shift_size=0 if (i % 2 == 0) else window_size // 2,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop,
+                attn_drop=attn_drop,
+                drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                norm_layer=norm_layer) for i in range(depth)
+        ])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = None
+
+    def forward(self, x, x_size):
+        for blk in self.blocks:
+            if self.use_checkpoint:
+                x = checkpoint.checkpoint(blk, x)
+            else:
+                x = blk(x, x_size)
+        if self.downsample is not None:
+            x = self.downsample(x)
+        return x
+
+    def extra_repr(self) -> str:
+        return f'dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}'
+
+    def flops(self):
+        flops = 0
+        for blk in self.blocks:
+            flops += blk.flops()
+        if self.downsample is not None:
+            flops += self.downsample.flops()
+        return flops
+
+
+class RSTB(nn.Module):
+    """Residual Swin Transformer Block (RSTB).
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
+        img_size: Input image size.
+        patch_size: Patch size.
+        resi_connection: The convolutional block before residual connection.
+    """
+
+    def __init__(self,
+                 dim,
+                 input_resolution,
+                 depth,
+                 num_heads,
+                 window_size,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop=0.,
+                 attn_drop=0.,
+                 drop_path=0.,
+                 norm_layer=nn.LayerNorm,
+                 downsample=None,
+                 use_checkpoint=False,
+                 img_size=224,
+                 patch_size=4,
+                 resi_connection='1conv'):
+        super(RSTB, self).__init__()
+
+        self.dim = dim
+        self.input_resolution = input_resolution
+
+        self.residual_group = BasicLayer(
+            dim=dim,
+            input_resolution=input_resolution,
+            depth=depth,
+            num_heads=num_heads,
+            window_size=window_size,
+            mlp_ratio=mlp_ratio,
+            qkv_bias=qkv_bias,
+            qk_scale=qk_scale,
+            drop=drop,
+            attn_drop=attn_drop,
+            drop_path=drop_path,
+            norm_layer=norm_layer,
+            downsample=downsample,
+            use_checkpoint=use_checkpoint)
+
+        if resi_connection == '1conv':
+            self.conv = nn.Conv2d(dim, dim, 3, 1, 1)
+        elif resi_connection == '3conv':
+            # to save parameters and memory
+            self.conv = nn.Sequential(
+                nn.Conv2d(dim, dim // 4, 3, 1, 1), nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                nn.Conv2d(dim // 4, dim // 4, 1, 1, 0), nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                nn.Conv2d(dim // 4, dim, 3, 1, 1))
+
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim, norm_layer=None)
+
+        self.patch_unembed = PatchUnEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=0, embed_dim=dim, norm_layer=None)
+
+    def forward(self, x, x_size):
+        return self.patch_embed(self.conv(self.patch_unembed(self.residual_group(x, x_size), x_size))) + x
+
+    def flops(self):
+        flops = 0
+        flops += self.residual_group.flops()
+        h, w = self.input_resolution
+        flops += h * w * self.dim * self.dim * 9
+        flops += self.patch_embed.flops()
+        flops += self.patch_unembed.flops()
+
+        return flops
+
+
+class PatchEmbed(nn.Module):
+    r""" Image to Patch Embedding
+
+    Args:
+        img_size (int): Image size.  Default: 224.
+        patch_size (int): Patch token size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patches_resolution = patches_resolution
+        self.num_patches = patches_resolution[0] * patches_resolution[1]
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+        if norm_layer is not None:
+            self.norm = norm_layer(embed_dim)
+        else:
+            self.norm = None
+
+    def forward(self, x):
+        x = x.flatten(2).transpose(1, 2)  # b Ph*Pw c
+        if self.norm is not None:
+            x = self.norm(x)
+        return x
+
+    def flops(self):
+        flops = 0
+        h, w = self.img_size
+        if self.norm is not None:
+            flops += h * w * self.embed_dim
+        return flops
+
+
+class PatchUnEmbed(nn.Module):
+    r""" Image to Patch Unembedding
+
+    Args:
+        img_size (int): Image size.  Default: 224.
+        patch_size (int): Patch token size. Default: 4.
+        in_chans (int): Number of input image channels. Default: 3.
+        embed_dim (int): Number of linear projection output channels. Default: 96.
+        norm_layer (nn.Module, optional): Normalization layer. Default: None
+    """
+
+    def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
+        super().__init__()
+        img_size = to_2tuple(img_size)
+        patch_size = to_2tuple(patch_size)
+        patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
+        self.img_size = img_size
+        self.patch_size = patch_size
+        self.patches_resolution = patches_resolution
+        self.num_patches = patches_resolution[0] * patches_resolution[1]
+
+        self.in_chans = in_chans
+        self.embed_dim = embed_dim
+
+    def forward(self, x, x_size):
+        x = x.transpose(1, 2).view(x.shape[0], self.embed_dim, x_size[0], x_size[1])  # b Ph*Pw c
+        return x
+
+    def flops(self):
+        flops = 0
+        return flops
+
+
+class Upsample(nn.Sequential):
+    """Upsample module.
+
+    Args:
+        scale (int): Scale factor. Supported scales: 2^n and 3.
+        num_feat (int): Channel number of intermediate features.
+    """
+
+    def __init__(self, scale, num_feat):
+        m = []
+        if (scale & (scale - 1)) == 0:  # scale = 2^n
+            for _ in range(int(math.log(scale, 2))):
+                m.append(nn.Conv2d(num_feat, 4 * num_feat, 3, 1, 1))
+                m.append(nn.PixelShuffle(2))
+        elif scale == 3:
+            m.append(nn.Conv2d(num_feat, 9 * num_feat, 3, 1, 1))
+            m.append(nn.PixelShuffle(3))
+        else:
+            raise ValueError(f'scale {scale} is not supported. Supported scales: 2^n and 3.')
+        super(Upsample, self).__init__(*m)
+
+
+class UpsampleOneStep(nn.Sequential):
+    """UpsampleOneStep module (the difference with Upsample is that it always only has 1conv + 1pixelshuffle)
+       Used in lightweight SR to save parameters.
+
+    Args:
+        scale (int): Scale factor. Supported scales: 2^n and 3.
+        num_feat (int): Channel number of intermediate features.
+
+    """
+
+    def __init__(self, scale, num_feat, num_out_ch, input_resolution=None):
+        self.num_feat = num_feat
+        self.input_resolution = input_resolution
+        m = []
+        m.append(nn.Conv2d(num_feat, (scale**2) * num_out_ch, 3, 1, 1))
+        m.append(nn.PixelShuffle(scale))
+        super(UpsampleOneStep, self).__init__(*m)
+
+    def flops(self):
+        h, w = self.input_resolution
+        flops = h * w * self.num_feat * 3 * 9
+        return flops
+
+
+@ARCH_REGISTRY.register()
+class SwinIR(nn.Module):
+    r""" SwinIR
+        A PyTorch impl of : `SwinIR: Image Restoration Using Swin Transformer`, based on Swin Transformer.
+
+    Args:
+        img_size (int | tuple(int)): Input image size. Default 64
+        patch_size (int | tuple(int)): Patch size. Default: 1
+        in_chans (int): Number of input image channels. Default: 3
+        embed_dim (int): Patch embedding dimension. Default: 96
+        depths (tuple(int)): Depth of each Swin Transformer layer.
+        num_heads (tuple(int)): Number of attention heads in different layers.
+        window_size (int): Window size. Default: 7
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None
+        drop_rate (float): Dropout rate. Default: 0
+        attn_drop_rate (float): Attention dropout rate. Default: 0
+        drop_path_rate (float): Stochastic depth rate. Default: 0.1
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
+        upscale: Upscale factor. 2/3/4/8 for image SR, 1 for denoising and compress artifact reduction
+        img_range: Image range. 1. or 255.
+        upsampler: The reconstruction reconstruction module. 'pixelshuffle'/'pixelshuffledirect'/'nearest+conv'/None
+        resi_connection: The convolutional block before residual connection. '1conv'/'3conv'
+    """
+
+    def __init__(self,
+                 img_size=64,
+                 patch_size=1,
+                 in_chans=3,
+                 embed_dim=96,
+                 depths=(6, 6, 6, 6),
+                 num_heads=(6, 6, 6, 6),
+                 window_size=7,
+                 mlp_ratio=4.,
+                 qkv_bias=True,
+                 qk_scale=None,
+                 drop_rate=0.,
+                 attn_drop_rate=0.,
+                 drop_path_rate=0.1,
+                 norm_layer=nn.LayerNorm,
+                 ape=False,
+                 patch_norm=True,
+                 use_checkpoint=False,
+                 upscale=2,
+                 img_range=1.,
+                 upsampler='',
+                 resi_connection='1conv',
+                 **kwargs):
+        super(SwinIR, self).__init__()
+        num_in_ch = in_chans
+        num_out_ch = in_chans
+        num_feat = 64
+        self.img_range = img_range
+        if in_chans == 3:
+            rgb_mean = (0.4488, 0.4371, 0.4040)
+            self.mean = torch.Tensor(rgb_mean).view(1, 3, 1, 1)
+        else:
+            self.mean = torch.zeros(1, 1, 1, 1)
+        self.upscale = upscale
+        self.upsampler = upsampler
+
+        # ------------------------- 1, shallow feature extraction ------------------------- #
+        self.conv_first = nn.Conv2d(num_in_ch, embed_dim, 3, 1, 1)
+
+        # ------------------------- 2, deep feature extraction ------------------------- #
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.ape = ape
+        self.patch_norm = patch_norm
+        self.num_features = embed_dim
+        self.mlp_ratio = mlp_ratio
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            img_size=img_size,
+            patch_size=patch_size,
+            in_chans=embed_dim,
+            embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+        num_patches = self.patch_embed.num_patches
+        patches_resolution = self.patch_embed.patches_resolution
+        self.patches_resolution = patches_resolution
+
+        # merge non-overlapping patches into image
+        self.patch_unembed = PatchUnEmbed(
+            img_size=img_size,
+            patch_size=patch_size,
+            in_chans=embed_dim,
+            embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+
+        # absolute position embedding
+        if self.ape:
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        # stochastic depth
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+
+        # build Residual Swin Transformer blocks (RSTB)
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            layer = RSTB(
+                dim=embed_dim,
+                input_resolution=(patches_resolution[0], patches_resolution[1]),
+                depth=depths[i_layer],
+                num_heads=num_heads[i_layer],
+                window_size=window_size,
+                mlp_ratio=self.mlp_ratio,
+                qkv_bias=qkv_bias,
+                qk_scale=qk_scale,
+                drop=drop_rate,
+                attn_drop=attn_drop_rate,
+                drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],  # no impact on SR results
+                norm_layer=norm_layer,
+                downsample=None,
+                use_checkpoint=use_checkpoint,
+                img_size=img_size,
+                patch_size=patch_size,
+                resi_connection=resi_connection)
+            self.layers.append(layer)
+        self.norm = norm_layer(self.num_features)
+
+        # build the last conv layer in deep feature extraction
+        if resi_connection == '1conv':
+            self.conv_after_body = nn.Conv2d(embed_dim, embed_dim, 3, 1, 1)
+        elif resi_connection == '3conv':
+            # to save parameters and memory
+            self.conv_after_body = nn.Sequential(
+                nn.Conv2d(embed_dim, embed_dim // 4, 3, 1, 1), nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                nn.Conv2d(embed_dim // 4, embed_dim // 4, 1, 1, 0), nn.LeakyReLU(negative_slope=0.2, inplace=True),
+                nn.Conv2d(embed_dim // 4, embed_dim, 3, 1, 1))
+
+        # ------------------------- 3, high quality image reconstruction ------------------------- #
+        if self.upsampler == 'pixelshuffle':
+            # for classical SR
+            self.conv_before_upsample = nn.Sequential(
+                nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True))
+            self.upsample = Upsample(upscale, num_feat)
+            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+        elif self.upsampler == 'pixelshuffledirect':
+            # for lightweight SR (to save parameters)
+            self.upsample = UpsampleOneStep(upscale, embed_dim, num_out_ch,
+                                            (patches_resolution[0], patches_resolution[1]))
+        elif self.upsampler == 'nearest+conv':
+            # for real-world SR (less artifacts)
+            assert self.upscale == 4, 'only support x4 now.'
+            self.conv_before_upsample = nn.Sequential(
+                nn.Conv2d(embed_dim, num_feat, 3, 1, 1), nn.LeakyReLU(inplace=True))
+            self.conv_up1 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+            self.conv_up2 = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+            self.conv_hr = nn.Conv2d(num_feat, num_feat, 3, 1, 1)
+            self.conv_last = nn.Conv2d(num_feat, num_out_ch, 3, 1, 1)
+            self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
+        else:
+            # for image denoising and JPEG compression artifact reduction
+            self.conv_last = nn.Conv2d(embed_dim, num_out_ch, 3, 1, 1)
+
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        return {'absolute_pos_embed'}
+
+    @torch.jit.ignore
+    def no_weight_decay_keywords(self):
+        return {'relative_position_bias_table'}
+
+    def forward_features(self, x):
+        x_size = (x.shape[2], x.shape[3])
+        x = self.patch_embed(x)
+        if self.ape:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+
+        for layer in self.layers:
+            x = layer(x, x_size)
+
+        x = self.norm(x)  # b seq_len c
+        x = self.patch_unembed(x, x_size)
+
+        return x
+
+    def forward(self, x):
+        self.mean = self.mean.type_as(x)
+        x = (x - self.mean) * self.img_range
+
+        if self.upsampler == 'pixelshuffle':
+            # for classical SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.conv_before_upsample(x)
+            x = self.conv_last(self.upsample(x))
+        elif self.upsampler == 'pixelshuffledirect':
+            # for lightweight SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.upsample(x)
+        elif self.upsampler == 'nearest+conv':
+            # for real-world SR
+            x = self.conv_first(x)
+            x = self.conv_after_body(self.forward_features(x)) + x
+            x = self.conv_before_upsample(x)
+            x = self.lrelu(self.conv_up1(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
+            x = self.lrelu(self.conv_up2(torch.nn.functional.interpolate(x, scale_factor=2, mode='nearest')))
+            x = self.conv_last(self.lrelu(self.conv_hr(x)))
+        else:
+            # for image denoising and JPEG compression artifact reduction
+            x_first = self.conv_first(x)
+            res = self.conv_after_body(self.forward_features(x_first)) + x_first
+            x = x + self.conv_last(res)
+
+        x = x / self.img_range + self.mean
+
+        return x
+
+    def flops(self):
+        flops = 0
+        h, w = self.patches_resolution
+        flops += h * w * 3 * self.embed_dim * 9
+        flops += self.patch_embed.flops()
+        for layer in self.layers:
+            flops += layer.flops()
+        flops += h * w * 3 * self.embed_dim * self.embed_dim
+        flops += self.upsample.flops()
+        return flops
+
+
+if __name__ == '__main__':
+    upscale = 4
+    window_size = 8
+    height = (1024 // upscale // window_size + 1) * window_size
+    width = (720 // upscale // window_size + 1) * window_size
+    model = SwinIR(
+        upscale=2,
+        img_size=(height, width),
+        window_size=window_size,
+        img_range=1.,
+        depths=[6, 6, 6, 6],
+        embed_dim=60,
+        num_heads=[6, 6, 6, 6],
+        mlp_ratio=2,
+        upsampler='pixelshuffledirect')
+    print(model)
+    print(height, width, model.flops() / 1e9)
+
+    x = torch.randn((1, 3, height, width))
+    x = model(x)
+    print(x.shape)

basicsr/archs/tof_arch.py

+import torch
+from torch import nn as nn
+from torch.nn import functional as F
+
+from basicsr.utils.registry import ARCH_REGISTRY
+from .arch_util import flow_warp
+
+
+class BasicModule(nn.Module):
+    """Basic module of SPyNet.
+
+    Note that unlike the architecture in spynet_arch.py, the basic module
+    here contains batch normalization.
+    """
+
+    def __init__(self):
+        super(BasicModule, self).__init__()
+        self.basic_module = nn.Sequential(
+            nn.Conv2d(in_channels=8, out_channels=32, kernel_size=7, stride=1, padding=3, bias=False),
+            nn.BatchNorm2d(32), nn.ReLU(inplace=True),
+            nn.Conv2d(in_channels=32, out_channels=64, kernel_size=7, stride=1, padding=3, bias=False),
+            nn.BatchNorm2d(64), nn.ReLU(inplace=True),
+            nn.Conv2d(in_channels=64, out_channels=32, kernel_size=7, stride=1, padding=3, bias=False),
+            nn.BatchNorm2d(32), nn.ReLU(inplace=True),
+            nn.Conv2d(in_channels=32, out_channels=16, kernel_size=7, stride=1, padding=3, bias=False),
+            nn.BatchNorm2d(16), nn.ReLU(inplace=True),
+            nn.Conv2d(in_channels=16, out_channels=2, kernel_size=7, stride=1, padding=3))
+
+    def forward(self, tensor_input):
+        """
+        Args:
+            tensor_input (Tensor): Input tensor with shape (b, 8, h, w).
+                8 channels contain:
+                [reference image (3), neighbor image (3), initial flow (2)].
+
+        Returns:
+            Tensor: Estimated flow with shape (b, 2, h, w)
+        """
+        return self.basic_module(tensor_input)
+
+
+class SPyNetTOF(nn.Module):
+    """SPyNet architecture for TOF.
+
+    Note that this implementation is specifically for TOFlow. Please use :file:`spynet_arch.py` for general use.
+    They differ in the following aspects:
+
+    1. The basic modules here contain BatchNorm.
+    2. Normalization and denormalization are not done here, as they are done in TOFlow.
+
+    ``Paper: Optical Flow Estimation using a Spatial Pyramid Network``
+
+    Reference: https://github.com/Coldog2333/pytoflow
+
+    Args:
+        load_path (str): Path for pretrained SPyNet. Default: None.
+    """
+
+    def __init__(self, load_path=None):
+        super(SPyNetTOF, self).__init__()
+
+        self.basic_module = nn.ModuleList([BasicModule() for _ in range(4)])
+        if load_path:
+            self.load_state_dict(torch.load(load_path, map_location=lambda storage, loc: storage)['params'])
+
+    def forward(self, ref, supp):
+        """
+        Args:
+            ref (Tensor): Reference image with shape of (b, 3, h, w).
+            supp: The supporting image to be warped: (b, 3, h, w).
+
+        Returns:
+            Tensor: Estimated optical flow: (b, 2, h, w).
+        """
+        num_batches, _, h, w = ref.size()
+        ref = [ref]
+        supp = [supp]
+
+        # generate downsampled frames
+        for _ in range(3):
+            ref.insert(0, F.avg_pool2d(input=ref[0], kernel_size=2, stride=2, count_include_pad=False))
+            supp.insert(0, F.avg_pool2d(input=supp[0], kernel_size=2, stride=2, count_include_pad=False))
+
+        # flow computation
+        flow = ref[0].new_zeros(num_batches, 2, h // 16, w // 16)
+        for i in range(4):
+            flow_up = F.interpolate(input=flow, scale_factor=2, mode='bilinear', align_corners=True) * 2.0
+            flow = flow_up + self.basic_module[i](
+                torch.cat([ref[i], flow_warp(supp[i], flow_up.permute(0, 2, 3, 1)), flow_up], 1))
+        return flow
+
+
+@ARCH_REGISTRY.register()
+class TOFlow(nn.Module):
+    """PyTorch implementation of TOFlow.
+
+    In TOFlow, the LR frames are pre-upsampled and have the same size with the GT frames.
+
+    ``Paper: Video Enhancement with Task-Oriented Flow``
+
+    Reference: https://github.com/anchen1011/toflow
+
+    Reference: https://github.com/Coldog2333/pytoflow
+
+    Args:
+        adapt_official_weights (bool): Whether to adapt the weights translated
+            from the official implementation. Set to false if you want to
+            train from scratch. Default: False
+    """
+
+    def __init__(self, adapt_official_weights=False):
+        super(TOFlow, self).__init__()
+        self.adapt_official_weights = adapt_official_weights
+        self.ref_idx = 0 if adapt_official_weights else 3
+
+        self.register_buffer('mean', torch.Tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1))
+        self.register_buffer('std', torch.Tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1))
+
+        # flow estimation module
+        self.spynet = SPyNetTOF()
+
+        # reconstruction module
+        self.conv_1 = nn.Conv2d(3 * 7, 64, 9, 1, 4)
+        self.conv_2 = nn.Conv2d(64, 64, 9, 1, 4)
+        self.conv_3 = nn.Conv2d(64, 64, 1)
+        self.conv_4 = nn.Conv2d(64, 3, 1)
+
+        # activation function
+        self.relu = nn.ReLU(inplace=True)
+
+    def normalize(self, img):
+        return (img - self.mean) / self.std
+
+    def denormalize(self, img):
+        return img * self.std + self.mean
+
+    def forward(self, lrs):
+        """
+        Args:
+            lrs: Input lr frames: (b, 7, 3, h, w).
+
+        Returns:
+            Tensor: SR frame: (b, 3, h, w).
+        """
+        # In the official implementation, the 0-th frame is the reference frame
+        if self.adapt_official_weights:
+            lrs = lrs[:, [3, 0, 1, 2, 4, 5, 6], :, :, :]
+
+        num_batches, num_lrs, _, h, w = lrs.size()
+
+        lrs = self.normalize(lrs.view(-1, 3, h, w))
+        lrs = lrs.view(num_batches, num_lrs, 3, h, w)
+
+        lr_ref = lrs[:, self.ref_idx, :, :, :]
+        lr_aligned = []
+        for i in range(7):  # 7 frames
+            if i == self.ref_idx:
+                lr_aligned.append(lr_ref)
+            else:
+                lr_supp = lrs[:, i, :, :, :]
+                flow = self.spynet(lr_ref, lr_supp)
+                lr_aligned.append(flow_warp(lr_supp, flow.permute(0, 2, 3, 1)))
+
+        # reconstruction
+        hr = torch.stack(lr_aligned, dim=1)
+        hr = hr.view(num_batches, -1, h, w)
+        hr = self.relu(self.conv_1(hr))
+        hr = self.relu(self.conv_2(hr))
+        hr = self.relu(self.conv_3(hr))
+        hr = self.conv_4(hr) + lr_ref
+
+        return self.denormalize(hr)

basicsr/archs/vgg_arch.py

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

basicsr/data/__init__.py

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