|
from torch import tanh, Tensor |
|
import torch.nn as nn |
|
from omegaconf import DictConfig |
|
from abc import ABC, abstractmethod |
|
|
|
|
|
class BaseGenerator(ABC, nn.Module): |
|
def __init__(self, channels: int = 3): |
|
super().__init__() |
|
self.channels = channels |
|
|
|
@abstractmethod |
|
def forward(self, x: Tensor) -> Tensor: |
|
pass |
|
|
|
|
|
class Generator(BaseGenerator): |
|
def __init__(self, cfg: DictConfig): |
|
super().__init__(cfg.channels) |
|
self.cfg = cfg |
|
self.model = self._construct_model() |
|
|
|
def _construct_model(self): |
|
initial_layer = nn.Sequential( |
|
nn.Conv2d( |
|
self.cfg.channels, |
|
self.cfg.num_features, |
|
kernel_size=7, |
|
stride=1, |
|
padding=3, |
|
padding_mode="reflect", |
|
), |
|
nn.ReLU(inplace=True), |
|
) |
|
|
|
down_blocks = nn.Sequential( |
|
ConvBlock( |
|
self.cfg.num_features, |
|
self.cfg.num_features * 2, |
|
kernel_size=3, |
|
stride=2, |
|
padding=1, |
|
), |
|
ConvBlock( |
|
self.cfg.num_features * 2, |
|
self.cfg.num_features * 4, |
|
kernel_size=3, |
|
stride=2, |
|
padding=1, |
|
), |
|
) |
|
|
|
residual_blocks = nn.Sequential( |
|
*[ |
|
ResidualBlock(self.cfg.num_features * 4) |
|
for _ in range(self.cfg.num_residuals) |
|
] |
|
) |
|
|
|
up_blocks = nn.Sequential( |
|
ConvBlock( |
|
self.cfg.num_features * 4, |
|
self.cfg.num_features * 2, |
|
down=False, |
|
kernel_size=3, |
|
stride=2, |
|
padding=1, |
|
output_padding=1, |
|
), |
|
ConvBlock( |
|
self.cfg.num_features * 2, |
|
self.cfg.num_features, |
|
down=False, |
|
kernel_size=3, |
|
stride=2, |
|
padding=1, |
|
output_padding=1, |
|
), |
|
) |
|
|
|
last_layer = nn.Conv2d( |
|
self.cfg.num_features, |
|
self.cfg.channels, |
|
kernel_size=7, |
|
stride=1, |
|
padding=3, |
|
padding_mode="reflect", |
|
) |
|
|
|
return nn.Sequential( |
|
initial_layer, down_blocks, residual_blocks, up_blocks, last_layer |
|
) |
|
|
|
def forward(self, x: Tensor) -> Tensor: |
|
return tanh(self.model(x)) |
|
|
|
|
|
class ConvBlock(nn.Module): |
|
def __init__( |
|
self, in_channels, out_channels, down=True, use_activation=True, **kwargs |
|
): |
|
super().__init__() |
|
self.conv = nn.Sequential( |
|
nn.Conv2d(in_channels, out_channels, padding_mode="reflect", **kwargs) |
|
if down |
|
else nn.ConvTranspose2d(in_channels, out_channels, **kwargs), |
|
nn.InstanceNorm2d(out_channels), |
|
nn.ReLU(inplace=True) if use_activation else nn.Identity(), |
|
) |
|
|
|
def forward(self, x: Tensor) -> Tensor: |
|
return self.conv(x) |
|
|
|
|
|
class ResidualBlock(nn.Module): |
|
def __init__(self, channels: int): |
|
super().__init__() |
|
self.block = nn.Sequential( |
|
ConvBlock(channels, channels, kernel_size=3, padding=1), |
|
ConvBlock( |
|
channels, channels, use_activation=False, kernel_size=3, padding=1 |
|
), |
|
) |
|
|
|
def forward(self, x: Tensor) -> Tensor: |
|
return x + self.block(x) |
|
|