Why disable gradients of on lookup images? about manydepth HOT 7 CLOSED

Thanks for your interest.

The benefit of doing it is as written in the comment there – we disable the gradients to save memory.

In practice you can enable gradient if you like. We found that it made little difference to scores.

from manydepth.

Thanks for reply！I'll have a try

from manydepth.

Thanks for your interest.

The benefit of doing it is as written in the comment there – we disable the gradients to save memory.

In practice you can enable gradient if you like. We found that it made little difference to scores.

I try to enable gradients of on lookup images, and modifiy the code like this:
in line 276, remove

with torch.no_grad():

in line 290~291, remove detach(), like this:

confidence_mask = self.compute_confidence_mask(cost_volume *(1 - missing_mask))

is this correct?
and if I enable gradients of on lookup images, the result will get worse? Should I change the the hyper-parameters such as num_epoch/lr ?

from manydepth.

Yes it looks like those changes should be enough to get gradients to flow to the lookup images.

But I don't know if the result would get better or worse if you do this. You will have to try it out and see!

from manydepth.

Yes it looks like those changes should be enough to get gradients to flow to the lookup images.

But I don't know if the result would get better or worse if you do this. You will have to try it out and see!

Thx! If I have any result, I will post here

from manydepth.

After modified the code as above, I got this result:

it seems like a little worse than origin result.
I'm not sure if it's due to randomness.
Overall, I think the two methods made little difference to score.

from manydepth.

Thanks a lot for reporting back. Glad you could confirm our suspicions.

from manydepth.