50 epochs usually way too much

[tommiekerssies]

Hi, I notice when doing the aggregated training, that usually after 1 or 2 epochs the best img, pixel AUROC is already reached. The many epochs after that only fluctuate around it and do not further improve. Is this something normal?

[tommiekerssies]

For example this is the training curves I get for tile with k=2:

[chaoqinhuang]

You are correct. In aggregated training, as the aggregated training set is large enough, training several epochs is enough for some categories. In your case, as the detection result has reached >98% AUC after 3 epochs, it is fine to stop the training.