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
Currently even if best.ckpt or other checkpoint exists in the project folder, with --load_best=true the checkpoint will only be loaded if last.ckpt also exists since the first check is for last.ckpt. I think it should be updated so with --load_best=true it can load any checkpoint. If this is updated, then we can update the docs/ to avoid needing to rename best.ckpt to last.ckpt when downloading checkpoints.
It makes sense for --autoresume=true to continue even if there's no existing checkpoint, but I think --load_best=true should only continue if there's checkpoint instead of continuing with random initialization. Otherwise it is not obvious what the problem is when you are running model test or a prediction pipeline that it is really using random weights.
The text was updated successfully, but these errors were encountered:
Two issues:
best.ckpt
or other checkpoint exists in the project folder, with--load_best=true
the checkpoint will only be loaded iflast.ckpt
also exists since the first check is forlast.ckpt
. I think it should be updated so with--load_best=true
it can load any checkpoint. If this is updated, then we can update thedocs/
to avoid needing to renamebest.ckpt
tolast.ckpt
when downloading checkpoints.--autoresume=true
to continue even if there's no existing checkpoint, but I think--load_best=true
should only continue if there's checkpoint instead of continuing with random initialization. Otherwise it is not obvious what the problem is when you are runningmodel test
or a prediction pipeline that it is really using random weights.The text was updated successfully, but these errors were encountered: