1234567891011121314151617181920212223242526272829 |
- from torch import nn
- class LRASPP(nn.Module):
- def __init__(self, in_channels, out_channels):
- super().__init__()
- self.aspp1 = nn.Sequential(
- nn.Conv2d(in_channels, out_channels, 1, bias=False),
- nn.BatchNorm2d(out_channels),
- nn.ReLU(True)
- )
- self.aspp2 = nn.Sequential(
- nn.AdaptiveAvgPool2d(1),
- nn.Conv2d(in_channels, out_channels, 1, bias=False),
- nn.Sigmoid()
- )
-
- def forward_single_frame(self, x):
- return self.aspp1(x) * self.aspp2(x)
-
- def forward_time_series(self, x):
- B, T = x.shape[:2]
- x = self.forward_single_frame(x.flatten(0, 1)).unflatten(0, (B, T))
- return x
-
- def forward(self, x):
- if x.ndim == 5:
- return self.forward_time_series(x)
- else:
- return self.forward_single_frame(x)
|