-
|
I think that the option of early stopping when your val loss is not improving during some time is very useful, how that happened that in timm's training pipeline there is no such arg parameter?( |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 1 reply
-
|
@detkov well, I don't believe in early stopping, just keeping your best checkpoints as measured by validation ... it's hard to 'calibrate' and get the stopping condition right, and it doesn't work well with a lot of good training schemes for larger datasets with cosine lr decay, heavy augmentation and regularization... |
Beta Was this translation helpful? Give feedback.
-
|
Should we at least have to option to pick it? It would save us a lot of resources if we can early stop. |
Beta Was this translation helpful? Give feedback.
@detkov well, I don't believe in early stopping, just keeping your best checkpoints as measured by validation ... it's hard to 'calibrate' and get the stopping condition right, and it doesn't work well with a lot of good training schemes for larger datasets with cosine lr decay, heavy augmentation and regularization...