networks/discriminator.py

import math
import torch
from torch.nn import functional as F
from torch import nn


def fused_leaky_relu(input, bias, negative_slope=0.2, scale=2 ** 0.5):
    return F.leaky_relu(input + bias, negative_slope) * scale


class FusedLeakyReLU(nn.Module):
    def __init__(self, channel, negative_slope=0.2, scale=2 ** 0.5):
        super().__init__()
        self.bias = nn.Parameter(torch.zeros(1, channel, 1, 1))
        self.negative_slope = negative_slope
        self.scale = scale

    def forward(self, input):
        # print("FusedLeakyReLU: ", input.abs().mean())
        out = fused_leaky_relu(input, self.bias, self.negative_slope, self.scale)
        # print("FusedLeakyReLU: ", out.abs().mean())
        return out


def upfirdn2d_native(input, kernel, up_x, up_y, down_x, down_y, pad_x0, pad_x1, pad_y0, pad_y1):
    _, minor, in_h, in_w = input.shape
    kernel_h, kernel_w = kernel.shape

    out = input.view(-1, minor, in_h, 1, in_w, 1)
    out = F.pad(out, [0, up_x - 1, 0, 0, 0, up_y - 1, 0, 0])
    out = out.view(-1, minor, in_h * up_y, in_w * up_x)

    out = F.pad(out, [max(pad_x0, 0), max(pad_x1, 0), max(pad_y0, 0), max(pad_y1, 0)])
    out = out[:, :, max(-pad_y0, 0): out.shape[2] - max(-pad_y1, 0), max(-pad_x0, 0): out.shape[3] - max(-pad_x1, 0), ]

    # out = out.permute(0, 3, 1, 2)
    out = out.reshape([-1, 1, in_h * up_y + pad_y0 + pad_y1, in_w * up_x + pad_x0 + pad_x1])
    w = torch.flip(kernel, [0, 1]).view(1, 1, kernel_h, kernel_w)
    out = F.conv2d(out, w)
    out = out.reshape(-1, minor, in_h * up_y + pad_y0 + pad_y1 - kernel_h + 1,
                      in_w * up_x + pad_x0 + pad_x1 - kernel_w + 1, )
    # out = out.permute(0, 2, 3, 1)

    return out[:, :, ::down_y, ::down_x]


def upfirdn2d(input, kernel, up=1, down=1, pad=(0, 0)):
    return upfirdn2d_native(input, kernel, up, up, down, down, pad[0], pad[1], pad[0], pad[1])


def make_kernel(k):
    k = torch.tensor(k, dtype=torch.float32)

    if k.ndim == 1:
        k = k[None, :] * k[:, None]

    k /= k.sum()

    return k


class Blur(nn.Module):
    def __init__(self, kernel, pad, upsample_factor=1):
        super().__init__()

        kernel = make_kernel(kernel)

        if upsample_factor > 1:
            kernel = kernel * (upsample_factor ** 2)

        self.register_buffer('kernel', kernel)

        self.pad = pad

    def forward(self, input):
        return upfirdn2d(input, self.kernel, pad=self.pad)


class ScaledLeakyReLU(nn.Module):
    def __init__(self, negative_slope=0.2):
        super().__init__()

        self.negative_slope = negative_slope

    def forward(self, input):
        return F.leaky_relu(input, negative_slope=self.negative_slope)


class EqualConv2d(nn.Module):
    def __init__(self, in_channel, out_channel, kernel_size, stride=1, padding=0, bias=True):
        super().__init__()

        self.weight = nn.Parameter(torch.randn(out_channel, in_channel, kernel_size, kernel_size))
        self.scale = 1 / math.sqrt(in_channel * kernel_size ** 2)

        self.stride = stride
        self.padding = padding

        if bias:
            self.bias = nn.Parameter(torch.zeros(out_channel))
        else:
            self.bias = None

    def forward(self, input):

        return F.conv2d(input, self.weight * self.scale, bias=self.bias, stride=self.stride,
                        padding=self.padding, )

    def __repr__(self):
        return (
            f'{self.__class__.__name__}({self.weight.shape[1]}, {self.weight.shape[0]},'
            f' {self.weight.shape[2]}, stride={self.stride}, padding={self.padding})'
        )


class EqualLinear(nn.Module):
    def __init__(self, in_dim, out_dim, bias=True, bias_init=0, lr_mul=1, activation=None):
        super().__init__()

        self.weight = nn.Parameter(torch.randn(out_dim, in_dim).div_(lr_mul))

        if bias:
            self.bias = nn.Parameter(torch.zeros(out_dim).fill_(bias_init))
        else:
            self.bias = None

        self.activation = activation

        self.scale = (1 / math.sqrt(in_dim)) * lr_mul
        self.lr_mul = lr_mul

    def forward(self, input):

        if self.activation:
            out = F.linear(input, self.weight * self.scale)
            out = fused_leaky_relu(out, self.bias * self.lr_mul)
        else:
            out = F.linear(input, self.weight * self.scale, bias=self.bias * self.lr_mul)

        return out

    def __repr__(self):
        return (f'{self.__class__.__name__}({self.weight.shape[1]}, {self.weight.shape[0]})')


class ConvLayer(nn.Sequential):
    def __init__(
            self,
            in_channel,
            out_channel,
            kernel_size,
            downsample=False,
            blur_kernel=[1, 3, 3, 1],
            bias=True,
            activate=True,
    ):
        layers = []

        if downsample:
            factor = 2
            p = (len(blur_kernel) - factor) + (kernel_size - 1)
            pad0 = (p + 1) // 2
            pad1 = p // 2

            layers.append(Blur(blur_kernel, pad=(pad0, pad1)))

            stride = 2
            self.padding = 0

        else:
            stride = 1
            self.padding = kernel_size // 2

        layers.append(EqualConv2d(in_channel, out_channel, kernel_size, padding=self.padding, stride=stride,
                                  bias=bias and not activate))

        if activate:
            if bias:
                layers.append(FusedLeakyReLU(out_channel))
            else:
                layers.append(ScaledLeakyReLU(0.2))

        super().__init__(*layers)


class ResBlock(nn.Module):
    def __init__(self, in_channel, out_channel, blur_kernel=[1, 3, 3, 1]):
        super().__init__()

        self.conv1 = ConvLayer(in_channel, in_channel, 3)
        self.conv2 = ConvLayer(in_channel, out_channel, 3, downsample=True)

        self.skip = ConvLayer(in_channel, out_channel, 1, downsample=True, activate=False, bias=False)

    def forward(self, input):
        out = self.conv1(input)
        out = self.conv2(out)

        skip = self.skip(input)
        out = (out + skip) / math.sqrt(2)

        return out


class Discriminator(nn.Module):
    def __init__(self, size, channel_multiplier=1, blur_kernel=[1, 3, 3, 1]):
        super().__init__()

        self.size = size

        channels = {
            4: 512,
            8: 512,
            16: 512,
            32: 512,
            64: 256 * channel_multiplier,
            128: 128 * channel_multiplier,
            256: 64 * channel_multiplier,
            512: 32 * channel_multiplier,
            1024: 16 * channel_multiplier,
        }

        convs = [ConvLayer(3, channels[size], 1)]
        log_size = int(math.log(size, 2))
        in_channel = channels[size]

        for i in range(log_size, 2, -1):
            out_channel = channels[2 ** (i - 1)]
            convs.append(ResBlock(in_channel, out_channel, blur_kernel))
            in_channel = out_channel

        self.convs = nn.Sequential(*convs)

        self.stddev_group = 4
        self.stddev_feat = 1

        self.final_conv = ConvLayer(in_channel + 1, channels[4], 3)
        self.final_linear = nn.Sequential(
            EqualLinear(channels[4] * 4 * 4, channels[4], activation='fused_lrelu'),
            EqualLinear(channels[4], 1),
        )

    def forward(self, input):
        out = self.convs(input)
        batch, channel, height, width = out.shape

        group = min(batch, self.stddev_group)
        stddev = out.view(group, -1, self.stddev_feat, channel // self.stddev_feat, height, width)
        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, height, width)
        out = torch.cat([out, stddev], 1)

        out = self.final_conv(out)

        out = out.view(batch, -1)
        out = self.final_linear(out)

        return out