search

zeng-yifei / STAG4D

Official Implementation for STAG4D: Spatial-Temporal Anchored Generative 4D Gaussians
156 stars 7 forks source link

Acceleration the Optimization #19

Open Duotun opened 3 months ago

Duotun commented 3 months ago

Hi,

Sorry to bother you a little bit. I was curious about how to accelerate the optimizations for this 4D generation simply. Increasing batch size doesn't work very well. Thank you!

Best

zeng-yifei commented 3 months ago

Actually, increasing batch size will only prolong the training time.

I am also not very familiar with network optimization. But what I can tell is where the time is spent in the training process. The main time cost lies in the backpropagation of the diffusion network and the hexplane network (about 1/2 time and 1/4 time). If you want to accelerate the optimizations, maybe you can find your way in these two processes.