150 lines
6.7 KiB
Python
150 lines
6.7 KiB
Python
import torch
|
|
import torch.nn as nn
|
|
|
|
from model import MODEL
|
|
from model.base.module import Conv2dBlock, ResidualBlock, LinearBlock
|
|
|
|
|
|
def _get_down_sampling_sequence(in_channels, base_channels, num_conv, max_down_sampling_multiple=2,
|
|
padding_mode='reflect', activation_type="ReLU", norm_type="NONE"):
|
|
sequence = [Conv2dBlock(
|
|
in_channels, base_channels, kernel_size=7, stride=1, padding=3, padding_mode=padding_mode,
|
|
activation_type=activation_type, norm_type=norm_type
|
|
)]
|
|
multiple_now = 1
|
|
for i in range(1, num_conv + 1):
|
|
multiple_prev = multiple_now
|
|
multiple_now = min(2 ** i, 2 ** max_down_sampling_multiple)
|
|
sequence.append(Conv2dBlock(
|
|
multiple_prev * base_channels, multiple_now * base_channels,
|
|
kernel_size=4, stride=2, padding=1, padding_mode=padding_mode,
|
|
activation_type=activation_type, norm_type=norm_type
|
|
))
|
|
return sequence, multiple_now * base_channels
|
|
|
|
|
|
class StyleEncoder(nn.Module):
|
|
def __init__(self, in_channels, out_dim, num_conv, base_channels=64,
|
|
max_down_sampling_multiple=2, padding_mode='reflect', activation_type="ReLU", norm_type="NONE"):
|
|
super().__init__()
|
|
|
|
sequence, last_channels = _get_down_sampling_sequence(
|
|
in_channels, base_channels, num_conv,
|
|
max_down_sampling_multiple, padding_mode, activation_type, norm_type
|
|
)
|
|
sequence.append(nn.AdaptiveAvgPool2d(1))
|
|
# conv1x1 works as fc when tensor's size is (batch_size, channels, 1, 1), keep same with origin code
|
|
sequence.append(nn.Conv2d(last_channels, out_dim, kernel_size=1, stride=1, padding=0))
|
|
self.sequence = nn.Sequential(*sequence)
|
|
|
|
def forward(self, image):
|
|
return self.sequence(image).view(image.size(0), -1)
|
|
|
|
|
|
class ContentEncoder(nn.Module):
|
|
def __init__(self, in_channels, num_down_sampling, num_residual_blocks, base_channels=64,
|
|
max_down_sampling_multiple=2,
|
|
padding_mode='reflect', activation_type="ReLU", norm_type="IN"):
|
|
super().__init__()
|
|
|
|
sequence, last_channels = _get_down_sampling_sequence(
|
|
in_channels, base_channels, num_down_sampling,
|
|
max_down_sampling_multiple, padding_mode, activation_type, norm_type
|
|
)
|
|
|
|
sequence += [ResidualBlock(last_channels, last_channels, padding_mode, activation_type, norm_type) for _ in
|
|
range(num_residual_blocks)]
|
|
self.sequence = nn.Sequential(*sequence)
|
|
|
|
def forward(self, image):
|
|
return self.sequence(image)
|
|
|
|
|
|
class Decoder(nn.Module):
|
|
def __init__(self, in_channels, out_channels, num_up_sampling, num_residual_blocks,
|
|
res_norm_type="AdaIN", norm_type="LN", activation_type="ReLU", padding_mode='reflect'):
|
|
super().__init__()
|
|
self.residual_blocks = nn.ModuleList([
|
|
ResidualBlock(in_channels, in_channels, padding_mode, activation_type, norm_type=res_norm_type)
|
|
for _ in range(num_residual_blocks)
|
|
])
|
|
|
|
sequence = list()
|
|
channels = in_channels
|
|
for i in range(num_up_sampling):
|
|
sequence.append(nn.Sequential(
|
|
nn.Upsample(scale_factor=2),
|
|
Conv2dBlock(channels, channels // 2,
|
|
kernel_size=5, stride=1, padding=2, padding_mode=padding_mode,
|
|
activation_type=activation_type, norm_type=norm_type),
|
|
))
|
|
channels = channels // 2
|
|
sequence.append(Conv2dBlock(channels, out_channels,
|
|
kernel_size=7, stride=1, padding=3, padding_mode="reflect",
|
|
activation_type="Tanh", norm_type="NONE"))
|
|
|
|
self.up_sequence = nn.Sequential(*sequence)
|
|
|
|
def forward(self, x, style):
|
|
as_param_style = torch.chunk(style, 2 * len(self.residual_blocks), dim=1)
|
|
# set style for decoder
|
|
for i, blk in enumerate(self.residual_blocks):
|
|
blk.conv1.normalization.set_style(as_param_style[2 * i])
|
|
blk.conv2.normalization.set_style(as_param_style[2 * i + 1])
|
|
x = blk(x)
|
|
return self.up_sequence(x)
|
|
|
|
|
|
class MLPFusion(nn.Module):
|
|
def __init__(self, in_features, out_features, base_features, n_blocks, activation_type="ReLU", norm_type="NONE"):
|
|
super().__init__()
|
|
|
|
sequence = [LinearBlock(in_features, base_features, activation_type=activation_type, norm_type=norm_type)]
|
|
sequence += [
|
|
LinearBlock(base_features, base_features, activation_type=activation_type, norm_type=norm_type)
|
|
for _ in range(n_blocks - 2)
|
|
]
|
|
sequence.append(LinearBlock(base_features, out_features, activation_type=activation_type, norm_type=norm_type))
|
|
self.sequence = nn.Sequential(*sequence)
|
|
|
|
def forward(self, x):
|
|
return self.sequence(x)
|
|
|
|
|
|
@MODEL.register_module("MUNIT-Generator")
|
|
class Generator(nn.Module):
|
|
def __init__(self, in_channels, out_channels, base_channels=64, style_dim=8,
|
|
num_mlp_base_feature=256, num_mlp_blocks=3,
|
|
max_down_sampling_multiple=2, num_content_down_sampling=2, num_style_down_sampling=2,
|
|
encoder_num_residual_blocks=4, decoder_num_residual_blocks=4,
|
|
padding_mode='reflect', activation_type="ReLU"):
|
|
super().__init__()
|
|
self.content_encoder = ContentEncoder(
|
|
in_channels, num_content_down_sampling, encoder_num_residual_blocks,
|
|
base_channels, max_down_sampling_multiple,
|
|
padding_mode, activation_type, norm_type="IN")
|
|
|
|
self.style_encoder = StyleEncoder(in_channels, style_dim, num_style_down_sampling, base_channels,
|
|
max_down_sampling_multiple, padding_mode, activation_type,
|
|
norm_type="NONE")
|
|
|
|
content_channels = base_channels * (2 ** max_down_sampling_multiple)
|
|
|
|
self.fusion = MLPFusion(style_dim, decoder_num_residual_blocks * 2 * content_channels * 2,
|
|
num_mlp_base_feature, num_mlp_blocks, activation_type,
|
|
norm_type="NONE")
|
|
|
|
self.decoder = Decoder(content_channels, out_channels, max_down_sampling_multiple, decoder_num_residual_blocks,
|
|
res_norm_type="AdaIN", norm_type="LN", activation_type=activation_type,
|
|
padding_mode=padding_mode)
|
|
|
|
def encode(self, x):
|
|
return self.content_encoder(x), self.style_encoder(x)
|
|
|
|
def decode(self, content, style):
|
|
self.decoder(content, self.fusion(style))
|
|
|
|
def forward(self, x):
|
|
content, style = self.encode(x)
|
|
return self.decode(content, style)
|