Does people train object detection methods with early stopping and what were their settings?

Question

I'm working on some stuff related to object detection methods (YOLOv3, Faster-RCNN, RetinaNet, ... ) and I need to train on VOC2007 and VOC2012 (using pretrained models of course). However when I read the relevant papers I do not see people describe if they trained using early stopping or just fixed number of iterations. And if they used early stopping, how many steps were set before stopping ? Because when I tried 100 steps before stopping, it got really poor results . Please help me, thank you very much.

Cormac

Can you provide your code as a stackoverflow.com/help/minimal-reproducible-example and references to the papers you are mentioning? It's too many options for what could go wrong. It's useful to see how loss changes from the iteration count. Try more steps and output the loss on each of the sets - train/test/validation. – Scrimmage

Answer 1

I found an implementation of the PASCAL VOC2012 dataset trained for semantic segmentation that uses the following early stopping parameters:

earlyStopping = EarlyStopping(
    monitor='val_loss', patience=30, verbose=2, mode='auto')

Ribbonwood

That's great. To be honest I tried early stopping with patience=100 and it had poor results. But when I tried running 50000 iterations the results were okay. – Cormac