Pytorch conv operation does not support fp16 operations on cpu. Remember to always check for compability before spending time on specific task.

I am just one step before deployment semantic segmentation task, starting to use pytorch profiler to check for eventual bootlenecks.

Spent some time on optimization, when just after switch to cpu, perform model.half() , prediction return errors.

https://github.com/open-mmlab/mmdetection/issues/2951

It is not new, caused by my lack of awareness. Pytorch profiler as a next nice thing to write about.

All opinions are my own.
s3nh