Why was learning rate warmup steps set to a maximum of 1000? #1427
morphinapg
started this conversation in
General
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
In the old version of the extension, I believe the maximum was 10,000. I like to set my warmup to be 5 epochs, and I often use datasets well over 1000 images, which means 5000+ warmup steps. I've found that I can modify this in scripts/main.py, but still, it seems 1000 is a very small number for a default maximum. I think the old maximum of 10,000 is reasonable but even then there may be some scenarios with very large datasets that could exceed that.
Maybe setting it to epochs instead of steps would be a better approach? I think a default of 1 epoch is probably a good place to start, but with more room above that, maybe a 10 epoch maximum, as I have seen some people who like to use 10% of their steps for warmup, and 100 epochs is a common total.
Beta Was this translation helpful? Give feedback.
All reactions