Fix adam optimizer #70
Open
Conversation
This file contains hidden or 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
According to the Adam Paper-v8 https://arxiv.org/pdf/1412.6980v8.pdf Algorithm 1 (p. 2). Behaves significantly better when running the trainer demo on MNIST (even better when changing adam learning rate to recommended 0.001) .
|
Thank you so much! Can someone please proof this and push this into release (on npm too please)? |
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Outdated suggestions cannot be applied.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.
Hi Andrej,
I recently ran the trainer demo on MNIST and wondered why the Adam optimizer performs so much more worse than Adadelta.
I think I found a little bug in the Adam implementation.
According to the Adam Paper-v8 https://arxiv.org/pdf/1412.6980v8.pdf Algorithm 1 (p. 2) the bias estimates use division instead of multiplication. The fixed version behaves significantly better when running the trainer demo on MNIST. To get the results as below I also changed the learning rate to 0.001 and the beta2 parameter to 0.999 (from 0.01 and 0.99 respectively) as recommended in the paper.
Before:
After:
