Liu Songtao

please provide more information about your error.....I have no idea what happened on your machine

The version of you cuda, pytorch, apex and so on. And the details of your error information.... I know you use the default script, but obviously the error is not...

I didn't test the code without distributed training, so even with single gpu, I suggest you keep the distributed training. And your batchsize is only one, which could significantly depress...

@AlexeyAB Hi, I just do this for simplicity. I think the essential part of rfb is the distribution of RFs, not BN. But I haven't done any experiments to verify...

40 mAP on VOC is quite low actually. Our experiments on VOC get about 84 mAP on VOC2007. If you don't have multi-gpu resources, I suggest you use mobilenet backbone...

Thanks for your suggestion! 1.5x longer training (450-500 epochs) will consume about 2 weeks on 10 2080ti, which is unavailable for me now. The purpose of training the ASFF* model...

@kaishijeng use config/yolov3_mobile.cfg instead. I forget to update this cfg file last night.

Hi, our experiments on RetinaNet show consistent improvements with ASFF, and the results can also be confirmed by another work [EfficientDet](https://arxiv.org/pdf/1911.09070.pdf). I am not familiar with Retinaface, but I suggest...

Sure, and I think the BiFPN is actually very similar to ASFF.

@92ypli #6 See this issue