| 1 | from typing import Optional |
| 2 | |
| 3 | import torch |
| 4 | |
| 5 | |
| 6 | class ResBlock(torch.nn.Module): |
| 7 | """ |
| 8 | Residual block with two convolutional layers, group normalization, and SiLU activation. |
| 9 | Args: |
| 10 | channels (int): Number of input and output channels. |
| 11 | mid_channels (Optional[int]): Number of channels in the intermediate convolution layer. Defaults to `channels` |
| 12 | if not specified. |
| 13 | dims (int): Dimensionality of the convolution (2 for Conv2d, 3 for Conv3d). Defaults to 3. |
| 14 | """ |
| 15 | |
| 16 | def __init__(self, channels: int, mid_channels: Optional[int] = None, dims: int = 3): |
| 17 | super().__init__() |
| 18 | if mid_channels is None: |
| 19 | mid_channels = channels |
| 20 | |
| 21 | conv = torch.nn.Conv2d if dims == 2 else torch.nn.Conv3d |
| 22 | |
| 23 | self.conv1 = conv(channels, mid_channels, kernel_size=3, padding=1) |
| 24 | self.norm1 = torch.nn.GroupNorm(32, mid_channels) |
| 25 | self.conv2 = conv(mid_channels, channels, kernel_size=3, padding=1) |
| 26 | self.norm2 = torch.nn.GroupNorm(32, channels) |
| 27 | self.activation = torch.nn.SiLU() |
| 28 | |
| 29 | def forward(self, x: torch.Tensor) -> torch.Tensor: |
| 30 | residual = x |
| 31 | x = self.conv1(x) |
| 32 | x = self.norm1(x) |
| 33 | x = self.activation(x) |
| 34 | x = self.conv2(x) |
| 35 | x = self.norm2(x) |
| 36 | x = self.activation(x + residual) |
| 37 | return x |
| 38 |