search

taichi-dev / taichi-nerfs

Implementations of NeRF variants based on Taichi + PyTorch
Apache License 2.0
734 stars 49 forks source link

Add an experimental option to use fast backward in the hash encoder #57

Closed Linyou closed 1 year ago

Linyou commented 1 year ago

This PR adds a new option --half_opt that enables the use of a faster backward in the hash encoder, which also allows for float16 acceleration.

Speed up info (autodiff -> half_opt):

erizmr commented 1 year ago

LGTM! The speed-up ratio is impressive!