GLFW error #65542: GLX: GLX extension not found

[compressionist]

Hello everyone,

I'm trying to run your GUI-enabled examples on a remote Ubuntu 20.04 virtual machine with a Quadro RTX 6000 (Lambda Labs entry-level instance). I installed the VNC server, VNC client and XQuartz on my Mac. So I can only open pictures in both tools and play videos in VNC. But unfortunately, when I run any of your examples with GUI support, I get the following error:

:~/instant-ngp$ ./testbed --scene data/image/tokyo.bin
05:24:09 INFO     Loading binary image from data/image/tokyo.bin
05:24:09 SUCCESS  Loaded a half-precision image with 1023x387 pixels.
05:24:09 INFO     Loading network config from: configs/image/base.json
05:24:09 INFO     GridEncoding:  Nmin=16 b=1.25984 F=2 T=2^24 L=16
05:24:09 INFO     Model:         2--[HashGrid]-->32--[FullyFusedMLP(neurons=64,layers=4)]-->3
05:24:09 INFO       total_encoding_params=1419392 total_network_params=7168
05:24:13 ERROR    GLFW error #65542: GLX: GLX extension not found
05:24:13 ERROR    Uncaught exception: GLFW window could not be created.

Maybe somebody of you solved similar issues?

The second question, if I run your examples without GUI, how do I understand that everything is working fine? How and where can I see intermediate and final results?

Best Regards,
Eugene

[Tom94]

Hi there,

I don't think GUI-over-VNC will be easy to make work. You might be able to get some remote OpenGL protocol working, but we don't officially support any of it.

While we're releasing this codebase as a way for other researchers to build upon and extend our results / tools, this is not fully fledged framework. There is no detailed documentation (exploring the source code is encouraged :)) and the scripts that we bundle only showcase basic functionality (especially without the GUI).

By running them with analogous command lines to the README, all you'll be able to see in console is that training doesn't explode (loss goes down), but not much more. scripts/run.py has a few convenience features, such as writing out screenshots or saving/loading snapshots of the trained weights, but nothing too in-depth.

In contrast, the Python bindings (of which you can see example use in scripts/run.py and a full list in src/pythin_api.cu) allow you to programmatically instrument all the features of instant-ngp. A good starting point would be to familiarize yourself with the different features of scripts/run.py (python scripts/run.py --help) and how they are implemented (screenshots, model snapshots, the training loop, etc.), followed by writing your own bespoke script that performs the experiments for you that you desire.

Cheers!