I think amass_to_pose can be vectorized (not iterating a single pose at a time but all poses in a sequence (a npz file) at a single time.).
I tried this and it gave about >10x boost in speed. (since my graphics card GTX 1660 Ti gave me a CUDA memory allocation error, I calculated in cpu though.).
Hello,
I think
amass_to_pose
can be vectorized (not iterating a single pose at a time but all poses in a sequence (a npz file) at a single time.).I tried this and it gave about >10x boost in speed. (since my graphics card GTX 1660 Ti gave me a CUDA memory allocation error, I calculated in cpu though.).
If it's not a resolved issue, I may issue a PR.
Thanks,
Young