Feng Wang
You could masked your loss to train those classes.
What about direct build by the following code: ```python3 from yolox.layers import FastCOCOEvalOp FastCOCOEvalOp().jit_load() ```
> Where should the above code be added? Plz try it in a terminal like `python3` or `ipython`
Plz make sure that ninja is installed in your conda env.
It's simliar to FCOS, you could check `get_assignments` in yolo_head.py for more details.
Your issue is not quiet clear for me. In my opinion, cls and reg loss only care about foreground, but objectness affect both foreground area and background area, code here...
To make the loss more stable and your training process could be more smooth. If your like, you could alos use foreground ema value as scaled number. @zye1996
Hi, normalization is not needed for YOLOX, here are some reason: 1. we didn't use imagenet pretrained model 2. our training time is long enough to fit non-normalization augmentation.
Are you using the latest commit? This bug is fixed in #1416
If you change your batchsize, please check your learning rate carefully. Issues of unproducble performance are often caused by wrong learning rate.