what's meaning of checkpoint name suffixed with ra1/ra2/a1h1? #1499
Answered
by
rwightman
lippman1125
asked this question in
Q&A
-
I found so many pretrained weights suffixed with ra1/a1h1 ? what do those abbreviations stand for? |
Beta Was this translation helpful? Give feedback.
Answered by
rwightman
Oct 17, 2022
Replies: 1 comment
-
|
Beta Was this translation helpful? Give feedback.
0 replies
Answer selected by
lippman1125
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
ra
= rand-augment, so ra/ra2/ra3 are different hparam schemes involving rand-augment, generally adding more goodies / more aug + reg as the number increases. ra2/ra3 would roughly correspond to the 'B' hparams in https://arxiv.org/abs/2110.00476, so they're RMSProp baseda1
corresponds to the LAMB based A1 hparams in RSB paper.h
means more augmentation + reg that the hparams in the paper, esp higher dropout/stochastic depth.