Does sample code with one-blob encoding still work?

[shocker-0x15]

Thanks for sharing the great library.

I tried the sample code using the image in the data directory and have confirmed the program instantly learns the image with "HashGrid" encoding (config_hash.json). However, with one-blob encoding (config_oneblob.json), the program constantly produces completely black images.

I tried bisection to find with which revision this issue occurred and it seems c88c1dcf81d89e0b4217ae4aaac34c963bf29d86 (The first revision for v1.2).

Does the sample with one-blob encoding still work in other environments other than me?

My environment:
Windows 10 21H2 (19044.1526)
Visual Studio Community 2022 (17.1.0)
cmake 3.21.4
RTX 3080 10GB
CUDA 11.6
Driver: 511.79

Thanks,

[Tom94]

Hi there, thanks for reporting this!

While everything works fine on my machine (Windows, WSL2, as well as native Linux), other users reported similar issues in the past.

There's a (seemingly unrelated) fix to this problem -- I suspect something funky in the underlying codegen -- which in the past always led to slightly worse performance in the case of 128-neurons wide neural networks.

But, and this stumps me, the fix in question improved performance by ~10% when I tried it just now! I'm not going to question it (the fix simplifies the code considerably), push it to master, and merely ask you to pull to confirm it works now.

[shocker-0x15]

Wow, thanks for your quick response!
I pulled the latest master 1e235e8 and have confirmed it works perfectly with one-blob encoding.

I appreciate your quick fix.