Why the warning：UserWarning: a target size that is different to the input size ？？ about kernelgan HOT 9 OPEN

Understandable. Just want to note that even with the warnings and no changes, it also works perfectly, so it's more of a minor inconvenience than a problem.

from kernelgan.

1. Thanks for the compliments.
2. IDK why you get these messages. Please verify your environment is exactly as specified in the requirements.txt.
3. I didn't try the network on huge images, but on realistic sized ones (few MegaPixels) - it works.. On the other hand, very small images (200x200 pixels) would not work very well. If you are interested in such images, I can elaborate.

from kernelgan.

1. Thanks for the compliments.
2. IDK why you get these messages. Please verify your environment is exactly as specified in the requirements.txt.
3. I didn't try the network on huge images, but on realistic sized ones (few MegaPixels) - it works.. On the other hand, very small images (200x200 pixels) would not work very well. If you are interested in such images, I can elaborate.

Thank you for your prompt reply
My operating environment is indeed different from the requirement. Is this just related to the environment?Does this warning affect the outcome?
The images I currently test are really around 200200, but most of the images I will run next time are above 70007000, so the program will be very easy to rebuild, right?

from kernelgan.

I am not familiar with these warnings but I suspect the environment.
7000x7000 is larger than I experimented on - I hope it works well... It should!

from kernelgan.

I am not familiar with these warnings but I suspect the environment.
7000x7000 is larger than I experimented on - I hope it works well... It should!

ok,Thank you for your reply

from kernelgan.

Hello. I (and probably everyone using a recent pytorch version) also get the warnings. These happen on three of the losses: BoundariesLoss, SumOfWeightsLoss and CentralizedLoss. Two of them are easy to fix, as they only need to add the missing dimension, but BoundariesLoss compares a fully expanded kernel (torch.Size([1, 1, 13, 13])) to the self.zero_label Variable, which only has one dimension of size 'k_size'. If expanded to the correct size, (1 x 1 x k_size x k_size) the warning goes away.

from kernelgan.

What torch version are you using?

from kernelgan.

Pytorch 1.3.0 and 1.6.0, tested on both.

from kernelgan.

I specified in the requirements torch=1.0.0.
Try that version, I do not want to change the code without understanding the degradation

from kernelgan.