Non-equivalence between batch_size
and accumulation_steps
#259
14thibea
started this conversation in
Show and tell
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hi everyone!
I tried to implement a reproducibility test to check that increasing the
accumulation_steps
to the value of thebatch_size
would give the same value... But actually it doesn't.I don't know from which component this discrepancy comes from, maybe as the batch size is different the images are actually not loaded in the same order? This issue could be investigated.
But actually even if setting
accumulation_steps
to mimick the batch_size don't lead exactly to the same results, maybe at least it brings closer the performance: then it would really correspond to the same batch size, but with a different seed.Beta Was this translation helpful? Give feedback.
All reactions