Skip to content

Conversation

@albanD
Copy link
Collaborator

@albanD albanD commented Jul 6, 2017

Fixes #1972

@apaszke I am starting to worry that the conv double backward test are taking too much time because they test all possible combination of parameters.
If you think this is a problem, maybe I could leave a subset of these tests (most common cases only) in test_nn.py and move the complete tests to a separate file? I am not sure if it's a good solution though.

@albanD
Copy link
Collaborator Author

albanD commented Jul 6, 2017

Ok, so my worry is confirmed as the cont built timeout :'(
Not sure what is the way to go here.

@soumith
Copy link
Contributor

soumith commented Jul 6, 2017

hmmm, they ran for 10 minutes? that's weird. Maybe heavier tests are enabled by default, but can be disabled with an env flag (we can keep them enabled for the GPU builds, but not the travis builds?)

@albanD
Copy link
Collaborator Author

albanD commented Jul 6, 2017

I will do some testing locally and try and reduce the number of tests while covering all use cases.
Otherwise an env variable may be a good solution.

@albanD
Copy link
Collaborator Author

albanD commented Jul 11, 2017

@soumith I reduced unnecessary tests such as testing many possible possible output channel or all even and uneven kernel/input sizes.
Should be good to go now.

@soumith soumith merged commit c888857 into pytorch:master Jul 13, 2017
@soumith
Copy link
Contributor

soumith commented Jul 13, 2017

awesome, thanks Alban!

zou3519 pushed a commit to zou3519/pytorch that referenced this pull request Mar 30, 2018
* wip: cuda targets

* Remove FindCuDNN.cmake as it is no longer needed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Conv double backwards with groups != 1 fails with size mismatch

3 participants