-
Notifications
You must be signed in to change notification settings - Fork 108
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
KFAC Open sourcing #24
Comments
Yes, we hope to release a research-level preview of KFAC soon! |
Great to hear, thanks! Is there any ETA for this? |
KFAC is now integrated (and the default optimiser) in the JAX branch. |
Hey @jsspencer , thanks a lot for open-sourcing KFAC implementation. Great work! However, when I run training for Mg with 8 V-100 GPUs (batch size 512), I got an error as follows
Any clue? BTW, I was using jax 0.2.9 and jaxlib 0.1.59, not sure if related. |
BTW, I can successfully train the net using KFAC on smaller atoms like O and F, but not for Na nor Mg. |
Hard to know. My suspicion is that the batch size is so small that the estimates required for the curvature in KFAC are noisy. KFAC requires solving the linear equations Ax=b, which is done via a Cholesky decomposition and assumes A is symmetric and positive-definite. The latter requirement might not be met for noisy estimates. |
@jsspencer Got it, will give it a try, thanks! |
Hi,
I was wondering whether you are planning on releasing the KFAC optimizer used in both papers as well?
I know that the TensorFlow version is available on GitHub. Is the JAX version also going open-source?
Thank you!
The text was updated successfully, but these errors were encountered: