Hardcoded Neural Network does not work as expected

[ccbl]

@Dan-Korneff I'm not sure if the tolerance refers to an absolute loss/ESR value, or whether it actually refers to a threshold of progress between Epochs.

Also, depending on the size of the model, there's basically a hard limit to how close the model can be to the original. The Rat is quite harmonically rich, so it will be on the harder end to model.

[ustk]

@Dan-Korneff Would it be CPU greedy to run several small models at the same time?
I'm thinking about modelling key component of a circuit. Not for reconstructing it though, just for having those models here and there in a DSP...

In fact my question is, since a single component/sub-circuit has an easier behaviour than a full circuit, are the resultant models lighter to run, at least if you run only one of those?

[Dan Korneff]

@ustk I think it should be doable. My plan is to do a Grey Box approach. Use ML for some items, and then analog modeling for others.
Test out the snippet above and see how much CPU it uses.
I'm still in the "trying to make this work" phase so I haven't gotten into measuring or optimizing yet.

[ccbl]

@ustk so basically every NN model uses a number of weights. This results in a certain number of parameters being able to be tweaked by the AI essentially. The more parameters the more CPU is required for processing.

With simpler circuits, or even say individual components, you could probably use very small models. I am actually planning to use the same approach as you, using NNs just for the non linnear stuff. I'm hoping to get it all working at higher sample rates though to reduce aliasing which would stack up quickly if you are chaining NNs together.

[ccbl]

@Dan-Korneff are there any flags I need to enable when compiling HISE to get the RT_Neural stuff compiling into a plugin properly?

[Dan Korneff]

@ccbl I didn't add any flags and it exported correctly.

[JulesV]

@Dan-Korneff Thanks for sharing this.

What is the method to train your models?

[Dan Korneff]

@JulesV https://forum.hise.audio/post/86374

[ccbl]

@Dan-Korneff I followed all the steps in your Gitlab to get the Aida-X trainer up and running, but when I get to the actual training part, it reads all the configs and starts the training process but then fails with

"RuntimeError: cuDNN error: CUDNN_STATUS_NOT_SUPPORTED. This error may appear if you passed in a non-contiguous input.

I tried with both 24bit and 32bit float input files.

[Dan Korneff]

@ccbl I'm still working on the scripts, so they aren't 100% yet. Feel free to dig around and see if you can trace the issue.

[ccbl]

This post is deleted!

[tomekslesicki]

@Dan-Korneff I'm also getting the

 CUDNN_STATUS_NOT_SUPPORTED

error when training with CUDA. With CUDA disabled, the traning goes as expected but is very slow (as expected ;-))

Are you using CUDA to train the model? If so, how are you setting the enviorment?

Like:

conda env config vars set CUBLAS_WORKSPACE_CONFIG=:16:8
conda activate base

...or some other way?

Thank you!

[Dan Korneff]

@tomekslesicki same is happening here. Looks like there's an incompatibility with the latest CUDA driver. I'll have to tweak that

[tomekslesicki]

@Dan-Korneff the solution is to install CUDA Toolkit 11.8 and install pytorch 2.3.0 instead of the current version. Here's the install prompt:

conda install pytorch==2.3.0 torchvision==0.18.0 torchaudio==2.3.0 pytorch-cuda=11.8 -c pytorch -c nvidia