Open korakoe opened 1 year ago
I'm very sorry about this, I just inherited inference_video.py, and didn't maintain all its functions on GMFSS, it is known that --montage
and --fp16
should not work properly, --UHD
is also de facto obsolete state.
And I have doubts about the effect of using FP16 to infer GMF. If I just want to try GMFSS_union under 8G video memory, I can add a mixed precision modification to do this.
using the --fp16 command results in this error:
RuntimeError: Input type (float) and bias type (struct c10::Half) should be the same