Too many upsampling layers in decoder when reducing depth of FPN #177
Description
I tried a FPN-Resnet with 3 layers and I get a prediction 2 times too big when computing loss.
I believe you should adapt the number of SegmentationBlock, something like this
self.seg_blocks = nn.ModuleList([
SegmentationBlock(pyramid_channels, segmentation_channels, n_upsamples=n_upsamples)
for n_upsamples in reversed(range(encoder_depth-1))
])