Element 0 of tensors does not require grad and does not have a grad_fn about monodepth2 HOT 2 CLOSED

Hi @sunnyHelen

The code with torch.no_grad() is used in validation only. The main reason it is there is to not store network activations in GPU memory when doing validation, as we are not training from that set of images so activations are not needed. It also may be helpful to catch bugs, such as the one you have posted. It seems that you are trying to backpropagate on the validation set. Without seeing the changes you introduced I'm not sure I can help.

Thanks!

from monodepth2.

Yes – if you remove with torch.no_grad(): here it shouldn't have any effect, as long as you are not trying to update network weights. You can read more here.

Thanks

from monodepth2.