Closed johnlockejrr closed 2 months ago
Sorry for only getting back to you on this now. The validation metric is val_mean_iu
not pixel accuracy which is most likely the reason for it picking the original model (as there wasn't any improvement). But the metrics aren't really representative of segmentation quality so using early stopping for segmentation training is not advised. Rather evaluate the output after a fixed number of epochs (your dataset is rather small so 50-100 is reasonable).
No problem, better later than never :) Thanks for the tip, I'll consider it!
I'm not sure this is an issue or a missunderstanding, anyway, I finetuned a segmentation model on a small data but kraken picks the first output model although following models seem to have higher accuracy.