Joshua Z. Zhang

@stereomatchingkiss This is currently hardcoded as float32: https://github.com/dmlc/gluon-cv/blob/master/gluoncv/loss.py#L222 See if changing this helps.

YOLO3 might be easier to deal with for fp16. The only thing I found hardcoded as fp32 is here: https://github.com/dmlc/gluon-cv/blob/master/gluoncv/model_zoo/yolo/yolo3.py#L65 You are welcome to modify that and let me know...

visually the pr looks good, do you have any clue where could be the problem?

Thanks for the PR, this might need further investigation whether the change will make the training results better/worse

Agreed, @zhanghang1989 we might be able to modify the operator by allowing it to take NDArray shapes rather than arguments only.

Maybe we can do local distributed training style instead of multithreading

It can support float16, but would require some knowledge of the legacy module way to cast to float16, which is a bit tricky, once I found an elegant way to...

See updates here: https://github.com/apache/incubator-mxnet/pull/17212

@hetong007 split may not support uneven division, but @becauseofAI used workaround which uses `F.slice_axis` with pre-calculated brakets. So I guess it's matching the original implementation? @becauseofAI Can you verify if...

@becauseofAI any update on this?