Liu Songtao

@amsword About 6 days on 4 V100, with 64 batchsize. I have not tried to optimize many training parameters such as epochs and lr, and I think 200 epochs are...

Yes, but I didn't make it work for this model.

Then you have to reduce your batchsize or choose a small model like yolox-tiny or yolox-s

Hi, thx for your exps. Actually, we did several exps for torch.cuda.amp too. The main reason for the significant dropping comes from [here](https://github.com/Megvii-BaseDetection/YOLOX/blob/main/yolox/models/yolo_head.py#L494-L496). It is illegal to conduct BCE with...

Ok, I will update a demo within several days. Since I am a little busy recently, I suggest you refer to test_RFB.py and write on your own. Feel free to...

@kaishijeng This is the error only occurred when you use python 2+, We only support python 3+ now.

@kaishijeng Emm......m, in progress.....

@BestSongEver Your location loss is abnormal so I just you to check the annotation of your own data. For training, the box format is [xmin, ymin, xmax, ymax]

@BestSongEver you may try a small lr like 0.001 or 0.0005 in your case.

@shwoo93 Would you please offer the lr and batchsize? In my experiments, larger batchsize would get higher accuracy. The VOC2007(80.5%) is obtained by lr=0.004, batch_size = 32, and I also...