You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I'm trying to reproduce the ViT-B32 results with 512 V100 GPUs and batch size ~65k on LIAON400M. However, I'm facing this issue where my loss would spike after each epoch and my zero-shot imagenet accuracy would be <1%. @rom1504 have you had this issue before?
The text was updated successfully, but these errors were encountered:
AFAIK we've never seen loss spikes synched with epochs like that, oscillations ala #822 that seem to line up with epochs aren't uncommon ... but dont' appear to have a significant impact and are nowhere near as significant as this based all the runs various people associated with this project have done.
I'm trying to reproduce the ViT-B32 results with 512 V100 GPUs and batch size ~65k on LIAON400M. However, I'm facing this issue where my loss would spike after each epoch and my zero-shot imagenet accuracy would be <1%. @rom1504 have you had this issue before?
The text was updated successfully, but these errors were encountered: