Feng Wang
@nahidalam Thank you for pointing out the bug in `augment_hsv` function. `cv2` is used in YOLOX, and `cv2.imread` return an BGR image by default. I think this bug might not...
Maybe you should try 0.01 / 64 * 4
Could you make your issue more clear?
@ImSuMyatNoe Just modify the following two lines accroding to your own setting. reference [code](https://github.com/Megvii-BaseDetection/YOLOX/blob/main/exps/example/yolox_voc/yolox_voc_s.py#L45-L46)
negative score in pycocotools means NaN here.
When it got stucked, did your GPU memory value(provided by nvidia-smi) differs on different devices?
It seems that you model is waiting for data. What about train on a single device?
What's the last log info? `init prefetcher` or just `Rank 1 initialization finished`?
Padding has not much effect on mAP, actually. Detection backbone have some requirements(every stage has its own stride, or refer [this](https://github.com/facebookresearch/detectron2/blob/8ba4dd8ded/detectron2/modeling/backbone/backbone.py#L31-L38)) , so padding image to given size is necessry.
what if you remove the cache and rerun the code?