Larger num_levels is not supported about torch-ngp HOT 4 CLOSED

Hi,
Can you run the newest example code on the fox dataset? The default num_levels is 16 and it should be able to work.
As for the aabb_scale, you could enlarge the --bound parameter to achieve a similar effect (the scene is assumed to be bounded in Box[-bound, bound], so a larger bound allows the ray to travel further. For the fox dataset, the default bound=2 should work fine).

from torch-ngp.

Thanks for the prompt reply! Sorry, there is a typo in the last message (num_levels=16 -> num_levels=32). I tried to enlarge it to 32, but got the -inf/inf issue.

For the bound, thanks and I will try that! it seems that the current implementation is doing uniform sampling in Box[-bound, bound]? However, in the original instant-ngp, they are doing uniform in unit cube, and exponential sampling in outside range?

from torch-ngp.

1. I tried 32 in the fox experiment and still didn't met inf, could you provide a reproducible example? (I just fixed a maybe related issue #11, you could try the latest version)
2. You are right, the current simple pytorch rendering function only performs uniform sampling. A better rendering that matches the original repo is still under development.

from torch-ngp.

Closed for now.

from torch-ngp.