-
Notifications
You must be signed in to change notification settings - Fork 312
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Selective Adam #432
Selective Adam #432
Conversation
Help from the community to evaluate the performance and final evaluation metrics for 30K iterations and to test this for multi-GPUs will be very much appreciated. :) |
I was able to test the
The comparison can be unfair here due to different number of gaussians. So I tried the MCMC strategy with 1M gaussian cap and the results for that are:
But here, I think that the densification calculation is slow which hides the speed-up of selective adam. Finally, we can see a drop in metrics. I think the main reason for this is that this is a different optimizer and requires different learning rates which I'm avoiding due to extra complexity. From my old experiments on the outdoor scenes of MN360, changing the learning rates led to better quality. |
This is cool! Do you want to add a doc string to the |
Done! |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM now!
This is the same as the sparse adam described in the Taming3DGS paper. Right now the non-zero radii gaussians are optimized. Hence the flag
visible adam
is used in the command line argument. Any other kind of mask can be passed to the same optimizer backend for different behaviour.This optimizer is compatible with
packed=True
as well. It gives a speed-up as briefly evaluated in the table below. I tried the bicycle scene till 7000 iterations on 3080Ti. Although, it leads to a different number of gaussians and slightly different evaluation metrics. The speed-up should be higher and more noticeable with higher number of iterations (due to higher number of gaussians).PS: I've created a new optimizer directory similar to strategy directory. I expect different optimizers catering to 3DGS to show up as research progresses. So I think this is a good option.