alexbrx
109 linhas
4.7 KiB
Python
109 linhas
4.7 KiB
Python
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=5, repeat_num=6, depth_concat=True):
|
|
super(Generator, self).__init__()
|
|
|
|
self.depth_concat = depth_concat
|
|
layers = []
|
|
if self.depth_concat:
|
|
layers.append(nn.Conv2d(3+c_dim, conv_dim, kernel_size=7, stride=1, padding=3, bias=False))
|
|
else:
|
|
layers.append(nn.Conv2d(4, 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):
|
|
# 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.
|
|
if self.depth_concat:
|
|
c = c.view(c.size(0), c.size(1), 1, 1)
|
|
c = c.repeat(1, 1, x.size(2), x.size(3))
|
|
else:
|
|
if c.size(1) == 2: #labels are assumed to have length 2 or 64 and images are 112x112
|
|
# c = c.view(c.size(0), 1, 1, 2)
|
|
# c = c.repeat(1, 1, 112, 56)
|
|
# c = c.view(-1).repeat_interleave(8).view(c.size(0),1,1,-1).repeat(1,1,112,7)
|
|
c = c.view(-1).repeat_interleave(56).view(c.size(0),1,1,-1).repeat(1,1,112,1)
|
|
elif c.size(1) == 64:
|
|
c = c.view(c.size(0), 1, 8, 8)
|
|
c = c.repeat(1, 1, 14, 14)
|
|
|
|
x = torch.cat([x, c], dim=1)
|
|
return self.main(x)
|
|
|
|
|
|
class Discriminator(nn.Module):
|
|
"""Discriminator network with PatchGAN."""
|
|
def __init__(self, image_size=128, conv_dim=64, c_dim=5, repeat_num=6, affectnet_emo_descr=None, d_loss_cls_type=None):
|
|
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)
|
|
self.conv2 = nn.Conv2d(curr_dim, c_dim, kernel_size=kernel_size, bias=False)
|
|
|
|
if affectnet_emo_descr == '64d_cls' and (d_loss_cls_type in ['both', 'pred']):
|
|
self.fc = torch.nn.Linear(c_dim, 7)
|
|
elif affectnet_emo_descr == '64d_reg' and (d_loss_cls_type in ['both', 'pred']):
|
|
self.fc = torch.nn.Linear(c_dim, 2)
|
|
|
|
def forward(self, x):
|
|
h = self.main(x)
|
|
out_src = self.conv1(h)
|
|
out_cls = self.conv2(h)
|
|
return out_src, out_cls.view(out_cls.size(0), out_cls.size(1))
|