Replies: 1 comment
-
Informative, yes, but note that it isn't a guide. It's actually a Q & A quiz to test your knowledge. The list of causes for each scenario (loss patterns) are multiple choice answers for the quiz, not definitive troubleshooting steps. You can take the quiz, ignore the wrong answers, and what's left are the legitimate troubleshooting possibilities for each scenario. I wanted to clarify so noone gets confused when they realize the "wrong answers" are making their loss issues worse instead of better. |
Beta Was this translation helpful? Give feedback.
0 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
This is a good quiz on model troubleshooting
As an example, I arrived at scenario 5 and sorted out the problem by switching on the crop jitter and Random flip in the 'Concept' configuration window, increasing the batch size to 6 (Adafactor optimizer on an RTX 4090), and increasing the Accumulation Steps to 6.
P.S. Cosine with restarts worked better for me compared to constant, cosine or REX Learning Rate Schedules.
Beta Was this translation helpful? Give feedback.
All reactions