CoolFace
Apppublic

riciii7/FastAPI-Batik-GAN

sourceHugging Faceupdated 1y agoView on Hugging Face
0likes
stylegan_model.py368 linesDownload Raw Back to root
1from torch import nn2import torch3from torch.nn import functional as F4from typing import Optional5import math6 7class WSLinear(nn.Module):8    '''9    Weighted scale linear for equalized learning rate.10 11    Args:12        in_features (int): The number of input features.13        out_features (int): The number of output features.14    '''15 16    def __init__(self, in_features: int, out_features: int) -> None:17        super(WSLinear, self).__init__()18        self.in_features = in_features19        self.out_features = out_features20 21        self.linear = nn.Linear(self.in_features, self.out_features)22        self.scale = (2 / self.in_features) ** 0.523        self.bias = self.linear.bias24        self.linear.bias = None25 26        self._init_weights()27 28    def _init_weights(self) -> None:29        nn.init.normal_(self.linear.weight)30        nn.init.zeros_(self.bias)31 32    def forward(self, x: torch.Tensor) -> torch.Tensor:33        return self.linear(x * self.scale) + self.bias34    35class WSConv2d(nn.Module):36    """37    Weight-scaled Conv2d layer for equalized learning rate.38 39    Args:40        in_channels (int): Number of input channels.41        out_channels (int): Number of output channels.42        kernel_size (int, optional): Size of the convolving kernel. Default: 3.43        stride (int, optional): Stride of the convolution. Default: 1.44        padding (int, optional): Padding added to all sides of the input. Default: 1.45        gain (float, optional): Gain factor for weight initialization. Default: 2.46    """47    def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, padding=1, gain=2):48        super().__init__()49        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride, padding)50        self.scale = (gain / (in_channels * kernel_size ** 2)) ** 0.551        self.bias = self.conv.bias52        self.conv.bias = None  # Remove bias to apply it after scaling53 54        # Initialize weights55        nn.init.normal_(self.conv.weight)56        nn.init.zeros_(self.bias)57 58    def forward(self, x):59        return self.conv(x * self.scale) + self.bias.view(1, self.bias.shape[0], 1, 1)60 61class Mapping(nn.Module):62    '''63    Mapping network.64 65    Args:66        features (int): Number of features in the input and output.67        num_layers (int): Number of layers in the feed forward network.68        num_styles (int): Number of styles to generate.69    '''70 71    def __init__(72        self,73        features: int,74        num_styles: int,75        num_layers: int = 8,76    ) -> None:77        super(Mapping, self).__init__()78        self.features = features79        self.num_layers = num_layers80        self.num_styles = num_styles81 82        layers = []83        for _ in range(self.num_layers):84            layers.append(WSLinear(self.features, self.features))85            layers.append(nn.LeakyReLU(0.2))86 87        self.fc = nn.Sequential(*layers)88 89    def forward(self, x: torch.Tensor) -> torch.Tensor:90        '''91        Args:92            x (torch.Tensor): Input tensor of shape (b, l).93 94        Returns:95            torch.Tensor: Output tensor with the same shape as input.96        '''97 98        x = self.fc(x) # (b, l)99        return x100 101class AdaIN(nn.Module):102    '''103    Adaptive Instance Normalization (AdaIN)104    AdaIN(x_i, y) = y_s,i * (x_i - mean(x_i)) / std(x_i) + y_b,i105 106    Args:107        eps (float, optional): Small value to avoid division by zero. Default value is 0.00001.108    '''109 110    def __init__(self, eps: float= 1e-5) -> None:111        super(AdaIN, self).__init__()112        self.eps = eps113 114    def forward(115        self,116        x: torch.Tensor,117        scale: torch.Tensor,118        shift: torch.Tensor119    ) -> torch.Tensor:120        '''121        Args:122            x (torch.Tensor): Input tensor of shape (b, c, h, w).123            scale (torch.Tensor): Scale tensor of shape (b, c).124            shift (torch.Tensor): Shift tensor of shape (b, c).125 126        Returns:127            torch.Tensor: Output tensor of shape (b, c, h, w).128        '''129 130        b, c, *_ = x.shape131 132        mean = x.mean(dim=(2, 3), keepdim=True) # (b, c, 1, 1)133        std = x.std(dim=(2, 3), keepdim=True) # (b, c, 1, 1)134        x_norm = (x - mean) / (std ** 2 + self.eps) ** .5135 136        scale = scale.view(b, c, 1, 1) # (b, c, 1, 1)137        shift = scale.view(b, c, 1, 1) # (b, c, 1, 1)138        outputs = scale * x_norm + shift # (b, c, h, w)139 140        return outputs141 142class SynthesisLayer(nn.Module):143    '''144    Synthesis network layer which consist of:145    - Conv2d.146    - AdaIN.147    - Affine transformation.148    - Noise injection.149 150    Args:151        in_channels (int): The number of input channels.152        out_channels (int): The number of output channels.153        latent_features (int): The number of latent features.154        use_conv (bool, optional): Whether to use convolution or not. Default value is True.155    '''156 157    def __init__(158        self,159        in_channels: int,160        out_channels: int,161        latent_features: int,162        use_conv: bool = True163    ) -> None:164        super(SynthesisLayer, self).__init__()165        self.in_channels = in_channels166        self.out_channels = out_channels167        self.latent_features = latent_features168        self.use_conv = use_conv169 170        self.conv = nn.Sequential(171            WSConv2d(self.in_channels, self.out_channels, kernel_size=3, padding=1),172            nn.LeakyReLU(0.2)173        ) if self.use_conv else nn.Identity()174        self.norm = AdaIN()175        self.scale_transform = WSLinear(self.latent_features, self.out_channels)176        self.shift_transform = WSLinear(self.latent_features, self.out_channels)177        self.noise_factor = nn.Parameter(torch.zeros(1, self.out_channels, 1, 1))178 179        self._init_weights()180 181    def _init_weights(self) -> None:182        for m in self.modules():183            if isinstance(m, (nn.Conv2d, nn.Linear)):184                nn.init.normal_(m.weight)185                if m.bias is not None:186                    nn.init.zeros_(m.bias)187        nn.init.ones_(self.scale_transform.bias)188 189    def forward(190        self,191        x: torch.Tensor,192        w: torch.Tensor,193        noise: Optional[torch.Tensor] = None194    ) -> torch.Tensor:195        '''196        Args:197            x (torch.Tensor): Input tensor of shape (b, c, h, w).198            w (torch.Tensor): Latent space vector of shape (b, l).199            noise (torch.Tensor, optional): Noise tensor of shape (b, 1, h, w). Default value is None.200 201        Returns:202            torch.Tensor: Output tensor of shape (b, c, h, w).203        '''204 205        b, _, h, w_ = x.shape206        x = self.conv(x) # (b, o_c, h, w)207        if noise is None:208            noise = torch.randn(b, 1, h, w_, device=x.device) # (b, 1, h, w)209        x += self.noise_factor * noise # (b, o_c, h, w)210        y_s = self.scale_transform(w) # (b, o_c)211        y_b = self.shift_transform(w) # (b, o_c)212        x = self.norm(x, y_s, y_b) # (b, i_c, h, w)213 214        return x215    216 217class SynthesisBlock(nn.Module):218    '''219    Synthesis network block which consist of:220    - Optional upsampling.221    - 2 Synthesis Layers.222 223    Args:224        in_channels (int): The number of input channels.225        out_channels (int): The number of output channels.226        latent_features (int): The number of latent features.227        use_conv (bool, optional): Whether to use convolution or not. Default value is True.228        upsample (bool, optional): Whether to use upsampling or not. Default value is True.229    '''230 231    def __init__(232        self,233        in_channels: int,234        out_channels: int,235        latent_features: int,236        *,237        use_conv: bool = True,238        upsample: bool = True239     ) -> None:240        super(SynthesisBlock, self).__init__()241        self.in_channels = in_channels242        self.out_channels = out_channels243        self.latent_features = latent_features244        self.use_conv = use_conv245        self.upsample = upsample246 247        self.upsample = nn.Upsample(scale_factor=2, mode='bilinear') if self.upsample else nn.Identity()248        self.layers = nn.ModuleList([249            SynthesisLayer(self.in_channels, self.in_channels, self.latent_features, use_conv=self.use_conv),250            SynthesisLayer(self.in_channels, self.out_channels, self.latent_features)251        ])252 253    def forward(self, x: torch.Tensor, w: torch.Tensor) -> torch.Tensor:254        '''255        Args:256            x (torch.Tensor): Input tensor of shape (b, c, h, w).257            w (torch.Tensor): Latent vector of shape (b, l).258 259        Returns:260            torch.Tensor: Output tensor of shape (b, c, h, w) if not upsample else (b, c, 2h, 2w).261        '''262 263        x = self.upsample(x) # (b, c, h, w) if not upsample else (b, c, 2h, 2w)264 265        for layer in self.layers:266            x = layer(x, w) # (b, c, h, w) if not upsample else (b, c, 2h, 2w)267 268        return x269 270class Synthesis(nn.Module):271    '''272    Synthesis network which consist of:273    - Constant tensor.274    - Synthesis blocks.275    - ToRGB convolutions.276 277    Args:278        resolution (int): The resolution of the image.279        const_channels (int): The number of channels in the constant tensor. Default value is 512.280    '''281 282    def __init__(self, resolution: int, const_channels: int = 512) -> None:283        super(Synthesis, self).__init__()284        self.const_channels = const_channels285        self.resolution = resolution286 287        self.resolution_levels = int(math.log2(resolution) - 1)288 289        self.constant = nn.Parameter(torch.ones(1, self.const_channels, 4, 4)) # (c, 4, 4)290 291        in_channels = self.const_channels292        blocks = [ SynthesisBlock(in_channels, in_channels, self.const_channels, use_conv=False, upsample=False) ]293        to_rgb = [ WSConv2d(in_channels, 3, kernel_size=1, padding=0) ]294 295        for _ in range(self.resolution_levels - 1):296            blocks.append(SynthesisBlock(in_channels, in_channels // 2, self.const_channels))297            to_rgb.append(WSConv2d(in_channels // 2, 3, kernel_size=1, padding=0))298            in_channels //= 2299 300        self.blocks = nn.ModuleList(blocks)301        self.to_rgb = nn.ModuleList(to_rgb)302 303    def forward(self, w: torch.Tensor, alpha: float, steps: int) -> torch.Tensor:304        '''305        Args:306            w (torch.Tensor): Latent space vector of shape (b, l).307            alpha (float): Fade in alpha value.308            steps (int): The number of steps starting from 0.309 310        Returns:311            torch.Tensor: Output tensor of shape (b, 3, h, w).312        '''313 314        b = w.size(0)315        x = self.constant.expand(b, -1, -1, -1).clone() # (b, c, h, w)316 317        if steps == 0:318            x = self.blocks[0](x, w) # (b, c, h, w)319            x = self.to_rgb[0](x) # (b, c, h, w)320            return x321 322        for i in range(steps):323            x = self.blocks[i](x, w) # (b, c, h/2, w/2)324 325        old_rgb = self.to_rgb[steps - 1](x) # (b, 3, h/2, w/2)326 327        x = self.blocks[steps](x, w) # (b, 3, h, w)328        new_rgb = self.to_rgb[steps](x) # (b, 3, h, w)329        old_rgb = F.interpolate(old_rgb, scale_factor=2, mode='bilinear', align_corners=False) # (b, 3, h, w)330 331        x = (1 - alpha) * old_rgb + alpha * new_rgb # (b, 3, h, w)332 333        return x334 335class StyleGAN(nn.Module):336    '''337    StyleGAN implementation.338 339    Args:340        num_features (int): The number of features in the latent space vector.341        resolution (int): The resolution of the image.342        num_blocks (int, optional): The number of blocks in the synthesis network. Default value is 10.343    '''344 345    def __init__(self, num_features: int, resolution: int, num_blocks: int = 10):346        super(StyleGAN, self).__init__()347        self.num_features = num_features348        self.resolution = resolution349        self.num_blocks = num_blocks350 351        self.mapping = Mapping(self.num_features, self.num_blocks)352        self.synthesis = Synthesis(self.resolution, self.num_features)353 354    def forward(self, x: torch.Tensor, alpha: float, steps: int) -> torch.Tensor:355        '''356        Args:357            x (torch.Tensor): Random input tensor of shape (b, l).358            alpha (float): Fade in alpha value.359            steps (int): The number of steps starting from 0.360 361        Returns:362            torch.Tensor: Output tensor of shape (b, c, h, w).363        '''364 365        w = self.mapping(x) # (b, l)366        outputs = self.synthesis(w, alpha, steps) # (b, c, h, w)367 368        return outputs