A question about UNet Crop and nn.functional.interpolate

vision
1

I’m trying to code the UNet paper, however I’ve come across the issue of implementing the crop when trying to code the decoder.

One person did this:

class Decoder(nn.Module):
    def __init__(self, chs=(1024, 512, 256, 128, 64)):
        super().__init__()
        self.chs         = chs
        self.upconvs    = nn.ModuleList([nn.ConvTranspose2d(chs[i], chs[i+1], 2, 2) for i in range(len(chs)-1)])
        self.dec_blocks = nn.ModuleList([Block(chs[i], chs[i+1]) for i in range(len(chs)-1)]) 
        
    def forward(self, x, encoder_features):
        for i in range(len(self.chs)-1):
            x        = self.upconvs[i](x)
            enc_ftrs = self.crop(encoder_features[i], x)
            x        = torch.cat([x, enc_ftrs], dim=1)
            x        = self.dec_blocks[i](x)
        return x
    
    def crop(self, enc_ftrs, x):
        _, _, H, W = x.shape
        enc_ftrs   = torchvision.transforms.CenterCrop([H, W])(enc_ftrs)
        return enc_ftrs

If I used nn.functional.interpolate() instead, would I get the same effect as the crop function in the above code statement?

2

I don’t know what you mean by “effect”, but an interpolation is not producing the same output as a cropping. Both manipulate the spatial size and you can use both to create a smaller output but the values won’t necessarily be equal.

1 Like