Do you mind sharing if adding OHEM will lead to a higher training loss?

soeaver / caffe-model

Caffe models (including classification, detection and segmentation) and deploy files for famouse networks

MIT License

1.28k stars 624 forks source link

Do you mind sharing if adding OHEM will lead to a higher training loss? #64

Open yxchng opened 6 years ago

yxchng commented 6 years ago

I have tried to train object detector using OHEM. My MAP stays the same but the training loss becomes higher. Is it normal? Do you mind sharing your training loss graphs?