Potential bug in gradient accumulation? #9
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
For the gradient accumulation case, fix Discriminator regularization to use a different batch of real images every round.
|
@tkarras will the accumulation bug be addressed? |
|
While I believe that might be a bug, I tested the difference between with and without loading new data and they were quite small so probably that doesn't affect things significantly. It also doesn't affect the performance at all for the case of using multi-gpu because than there's anyway only one round. |
patrickxia
pushed a commit
to patrickxia/stylegan2
that referenced
this pull request
Jan 10, 2021
fixed encoder
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
For the gradient accumulation case, currently the regularization for the discriminator performs multiple rounds on the same batch of images, instead of averaging the regularization loss over multiple different batches (as is done for the standard discriminator loss).
Is this a bug or is it arranged this way intentionally?
Thanks!