why not use fp16 these two lines?

Open parryppp opened this issue 8 months ago • 1 comments

https://github.com/deepinsight/insightface/blob/786c4a8327398aecb4cad0cb83ebcefc12b9d3cb/recognition/arcface_torch/backbones/iresnet.py#L160

Apr 19 '25 10:04 parryppp

probably because this is the most critical layer, it converts the output of the convolutions to a class probability so it needs high precision

May 02 '25 12:05 pyphan1