The decoder feature is always upsampled (aligned in spatial dimensions) before being passed to the attention gate.
"it is correct but i couldnt see on the diagram of attention unet's article . (it seems like there is no up conv on previous layer's signal at decoder just before attention)
so the code is like:
d4 = self.Up4(d5)
s3 = self.Att4(gate=d4, skip_connection=e3)
d4 = torch.cat((s3, d4), dim=1)
d4 = self.UpConv4(d4)
...
is it true
?
The decoder feature is always upsampled (aligned in spatial dimensions) before being passed to the attention gate.
"it is correct but i couldnt see on the diagram of attention unet's article . (it seems like there is no up conv on previous layer's signal at decoder just before attention)
so the code is like:
d4 = self.Up4(d5)
s3 = self.Att4(gate=d4, skip_connection=e3)
d4 = torch.cat((s3, d4), dim=1)
d4 = self.UpConv4(d4)
...
is it true
?