ENet prediction become worse after 100 epoch

xiaoyufenfei / Efficient-Segmentation-Networks

Lightweight models for real-time semantic segmentationon PyTorch (include SQNet, LinkNet, SegNet, UNet, ENet, ERFNet, EDANet, ESPNet, ESPNetv2, LEDNet, ESNet, FSSNet, CGNet, DABNet, Fast-SCNN, ContextNet, FPENet, etc.)

MIT License

898 stars 163 forks source link

ENet prediction become worse after 100 epoch #17

Closed guanqp closed 3 years ago

guanqp commented 3 years ago

I use this project to train ENet on cityscapes dataset(640*480). The loss decase normally, but predict result become worse when trainning for more than 50 epochs... from left to right, 10epoch, 50epoch, 80epoch. and the training loss vs epochs looks fine:

Can anyone give me some advice? Thanks!

guanqp commented 3 years ago

Looks like the train process does not evaluate the validation loss and mIoU, so the model over fited.