Skip to content
This repository was archived by the owner on Jul 1, 2023. It is now read-only.

Reverting inference behavior for BatchNorm #426

Merged
merged 1 commit into from
Aug 8, 2019
Merged

Reverting inference behavior for BatchNorm #426

merged 1 commit into from
Aug 8, 2019

Conversation

BradLarson
Copy link
Contributor

This reverts inference behavior for BatchNorm to what it was before PR #384. The reshaping introduced there was triggering shape mismatches in the MiniGo tests and preventing the demo from running due to runtime errors.

@BradLarson BradLarson merged commit 268055e into tensorflow:master Aug 8, 2019
@dan-zheng
Copy link
Member

I wonder if it's possible to check in a "MiniGo shape mismatch reproducer" test to ensure the bug in #384 won't happen again?

(I don't mean to block progress at all)

@BradLarson
Copy link
Contributor Author

That's a good idea. We need any kind of inference tests for BatchNorm and LayerNorm, because I believe both are only tested under training conditions. That would also help verify numerical correctness, like has been done on the training side.

@eaplatanios
Copy link
Contributor

I don't really understand why these reshapes were there. I'm preparing another PR that removes them from a couple other places where they also shouldn't be.

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

4 participants