About the labels in train loop

[silence-tang]

Hi, dear friends:
I think in the paper the author wanted to convey an idea that we should upsample the embedded feature maps to the same size as labels in order to calculate the loss. However, in the source code, it seems that it downsamples the labels as the same size as embedded faeture maps to calculate the loss. It confuses me, could you please tell me how it goes like this?