Open chenhaocs opened 1 year ago
Hi,perhaps you can try larger batch size or try single target mode. Segformerb5 sometimes has similar situation. However,when training comes to end,we can still get ~45% mIoU or more.
Do you have some ideas about this kind of degradation problem ?
on single target mode,this degradation issue is eased. However,more training cost is introduced.
During my training process, the performance (i.e., mIOU) of the Segformer_b5 increases (from 0 to 8000 iterations) and then decreases sharply (from 8000 to 40000), where the maximum nIOU could be 68.7 and finally the mIOU became 15.26.
Do you have the same observation ?