Question about downstream

Hello,

 I noticed that the loss calculation is performed on downsampled images within the network. Is it necessary to compute the loss on downsampled images, or could it also be effectively done on the original resolution images?

Looking forward to your response.

Thank you!
