It's quite a tall order actually, we need some time to think about it. You're asking for a pretty complicated thing, and it takes some brain power to come up with a simple, minimal example. Give us some time.

I understand. I really appreciate your time and efforts. I didn't realize that it might be that complicated.

That sounds like a neat solution for my current requirements. Will try to implement it. Also, I wonder, you wrote "We don't exactly do this", what did you mean by that? what is the alternative way to do it?

Thanks for your time and efforts.

I know someone who's using shared memory instead of multiprocessing Queues. But we've found that if you don't need to send more than 2.5GB/s then it's not worth the extra work of managing shaded memory. Multiprocessing Queue should be sufficient for most tasks.

Thanks for the replay. Sure, I will share my PR once it is working.

Ok, I have something working. Can I write the example in the context of training a model with pytorch, or should it be a more generic example that shows how to run and update a rendered scene in a different process?

I guess both would be useful actually! :)

Hi guys, (@Korijn @almarklein @kushalkolar)
I have submitted a pull request with a working example that demonstrates how to train a model and render the results at the same time. Specifically, the example just trains a GNN to predict the Gaussian curvature at each mesh vertex, by overfitting the ground-truth curvature.