Unstable accuracy and Loss


HI friends
Please help indentifying the mistake here.
I suspect my weights from new layer are not added properly.
Check out the training output/ArcMarging layer ,

CustomHead from where ArcMargin is invoked.

In what sense are the accuracy and loss unstable?
Do you see large changes running the same script?
If so, you could apply some changes to your code based on the Reproducibility docs.

in version 9 i faced the problem.check accuracy trend.
In the latest version i removed fp16 part which was for reducing the accuracy from float 32 to 16. After that i am not facing the issue.

Not sure how this change solved the issue.