FP 16 Pretrained MODEL

[Choi-YeongJoon]

I am very interested in your work! I ran your model in TensorRT, C++. The fps is so low that I think I need to run a model with fp16 precision. Do you have a pretrained model trained with fp16 precision? Can I possibly get it?

[hisfog]

The ConvNeXt model is not effcient (392.2 GMACs), you can try the KITTI (Efficient-b5) model (70.35 GMACs), which is much more faster and also accurate enough.
For fp16 models, they should be trained from scratch in fp16 precision, I didn't do that training, since this repo is not for industrial use cases.