Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

DL4J: Lack of input validation for batch norm? #6403

Closed
AlexDBlack opened this issue Sep 10, 2018 · 1 comment

Comments

@AlexDBlack
Copy link
Contributor

commented Sep 10, 2018

As reported in gitter by @DrChainsaw

DrChainsaw @DrChainsaw Sep 09 01:35
@AlexDBlack Running on CPU immediately showed I didn't handle Batchnormalization correctly. Setting nIn on that layer does not do anything. After fixing this I no longer crash at the same point with CUDA so it seems like the CUDA backend does not catch this immediately. Exception when using CPU: https://gist.github.com/DrChainsaw/cc5a9d4dfd5cfc418cebe9e140bf469a
Btw, the fix was to set nOut to what I wanted nIn to be
I'm not sure I can put this in a "crisp" issue. It seems that some precondition which is checked with native backend is not checked with CUDA backend and this causes (or might cause) a delayed error when executing. I could feedforward a few times just fine with the CUDA backend with the "corrupt" batchnorm and nothing seemed off

@AlexDBlack AlexDBlack self-assigned this Sep 10, 2018

AlexDBlack added a commit that referenced this issue Sep 17, 2018
AlexDBlack added a commit that referenced this issue Sep 18, 2018
Various fixes (#6450)
* #6401 transposei fix

* #6378 MmulTranspose fix + cleanup

* #6442 validate array order

* Cleanup and test fixes

* #6403 batch norm validation

* #6389 Fix TransferLearning nOutReplace issues; add nInReplace method

* Trigger CI
@lock

This comment has been minimized.

Copy link

commented Oct 18, 2018

This thread has been automatically locked since there has not been any recent activity after it was closed. Please open a new issue for related bugs.

@lock lock bot locked and limited conversation to collaborators Oct 18, 2018

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Projects
None yet
1 participant
You can’t perform that action at this time.