model.py

import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np


class ResidualBlock(nn.Module):
    """Residual Block with instance normalization."""
    def __init__(self, dim_in, dim_out):
        super(ResidualBlock, self).__init__()
        self.main = nn.Sequential(
            nn.Conv2d(dim_in, dim_out, kernel_size=3, stride=1, padding=1, bias=False),
            nn.InstanceNorm2d(dim_out, affine=True, track_running_stats=True),
            nn.ReLU(inplace=True),
            nn.Conv2d(dim_out, dim_out, kernel_size=3, stride=1, padding=1, bias=False),
            nn.InstanceNorm2d(dim_out, affine=True, track_running_stats=True))

    def forward(self, x):
        return x + self.main(x)


class Generator(nn.Module):
    """Generator network."""
    def __init__(self, conv_dim=64, c_dim=8, r_dim=2, repeat_num=6):
        super(Generator, self).__init__()

        layers = []
        layers.append(nn.Conv2d(3+c_dim+r_dim, conv_dim, kernel_size=7, stride=1, padding=3, bias=False))
        layers.append(nn.InstanceNorm2d(conv_dim, affine=True, track_running_stats=True))
        layers.append(nn.ReLU(inplace=True))

        # Down-sampling layers.
        curr_dim = conv_dim
        for i in range(2):
            layers.append(nn.Conv2d(curr_dim, curr_dim*2, kernel_size=4, stride=2, padding=1, bias=False))
            layers.append(nn.InstanceNorm2d(curr_dim*2, affine=True, track_running_stats=True))
            layers.append(nn.ReLU(inplace=True))
            curr_dim = curr_dim * 2

        # Bottleneck layers.
        for i in range(repeat_num):
            layers.append(ResidualBlock(dim_in=curr_dim, dim_out=curr_dim))

        # Up-sampling layers.
        for i in range(2):
            layers.append(nn.ConvTranspose2d(curr_dim, curr_dim//2, kernel_size=4, stride=2, padding=1, bias=False))
            layers.append(nn.InstanceNorm2d(curr_dim//2, affine=True, track_running_stats=True))
            layers.append(nn.ReLU(inplace=True))
            curr_dim = curr_dim // 2

        layers.append(nn.Conv2d(curr_dim, 3, kernel_size=7, stride=1, padding=3, bias=False))
        layers.append(nn.Tanh())
        self.main = nn.Sequential(*layers)

    def forward(self, x, c, r):
        """ Input shape
        :param x: N,C,H,W
        :param c: N,8
        :param r: N,2
        """
        # Replicate spatially and concatenate domain information.
        # Note that this type of label conditioning does not work at all if we use reflection padding in Conv2d.
        # This is because instance normalization ignores the shifting (or bias) effect.
        c = c.view(c.size(0), c.size(1), 1, 1)    # N,8,1,1
        c = c.repeat(1, 1, x.size(2), x.size(3))  # N,8,H,W
        r = r.view(r.size(0), r.size(1), 1, 1)    # N,2,1,1
        r = r.repeat(1, 1, x.size(2), x.size(3))  # N,2,H,W
        x = torch.cat([x, c, r], dim=1)              # N,C+8+2,H,W
        return self.main(x)


class Discriminator(nn.Module):
    """Discriminator network with PatchGAN."""
    def __init__(self, image_size=128, conv_dim=64, c_dim=8, r_dim=2, repeat_num=6):
        super(Discriminator, self).__init__()
        layers = []
        layers.append(nn.Conv2d(3, conv_dim, kernel_size=4, stride=2, padding=1))
        layers.append(nn.LeakyReLU(0.01))

        curr_dim = conv_dim
        for i in range(1, repeat_num):
            layers.append(nn.Conv2d(curr_dim, curr_dim*2, kernel_size=4, stride=2, padding=1))
            layers.append(nn.LeakyReLU(0.01))
            curr_dim = curr_dim * 2

        kernel_size = int(image_size / np.power(2, repeat_num))
        self.main = nn.Sequential(*layers)
        self.conv1 = nn.Conv2d(curr_dim, 1, kernel_size=3, stride=1, padding=1, bias=False) # real or fake
        self.conv2 = nn.Conv2d(curr_dim, c_dim, kernel_size=kernel_size, bias=False) # classification
        self.conv3 = nn.Conv2d(curr_dim, r_dim, kernel_size=kernel_size, bias=False) # regression
        
    def forward(self, x):
        h = self.main(x)
        out_src = self.conv1(h) # N, 2, 2, 2
        out_cls = self.conv2(h) # N, 8, 1, 1
        out_reg = self.conv3(h) # N, 2, 1, 1
        return out_src, out_cls.view(out_cls.size(0), out_cls.size(1)), out_reg.view(out_reg.size(0), out_reg.size(1))