Joshua Z. Zhang

@zhanghang1989 Can you have a look?

Like @chinakook said please try to remove usage of `.shape` in the network.

1. The existing network and loss don't respect rotated bounding boxes, so the best bet is that we can adjust the bounding boxes to rotated objects. This way is doable,...

The nan's looks suspicious to me, may be you don't have such classes in your validation set?

We are working on that right now.

We will definitely add a tutorial about how to load custom dataset, and how to fine-tune existing models. Stay tuned.

Replace the cross entropy loss in MultiBoxLoss

@Angzz Since you are using ndarray, I would suggest print intermediate variables one by one to evaluate the correctness of these values.

Just a quick examine, you should not use nd.log_softmax together with F.sigmoid() in FocalLoss, the rest looks good

We would be happy to share, but anyone who can contribute is super welcome since we have higher priority stuffs.