mmdetection
mmdetection copied to clipboard
open-mmlab
Add Validation Loss
Motivation In the latest version (v3), validation loss has been removed, making it difficult to properly evaluate model performance. Validation loss is a crucial metric for assessing overfitting, hyperparameter tuning, and model selection. Without it, users have no straightforward way to monitor generalization performance. Removing this feature in the newer version is confusing and impacts the evaluation process.
Related Resources If there is an official reason for its removal or alternative solutions, please provide references. Additionally, if there are third-party implementations or workarounds, sharing those would be helpful.
Additional Context It would be greatly appreciated if validation loss could be reintroduced or an alternative method for evaluation could be officially provided. If community contributions are welcomed, I’d be happy to help with a PR.
This has caused me unnecessary pain, too! Other libraries, e.g., MMSegmentation, have removed validation losses as well
Any update on this? I would greatly appreciate if someone could provide a Hook!!
Any update on this? I would greatly appreciate if someone could provide a Hook!!
I tried to implement a hook for val loss with no success, also using workflow=[('train', 1), ('val', 1)] does not seem to work. Did you manage to implement a working hook?
