Simple ML neural network

[Christoph Hart]

I realise we already hashed this discussion out, and people might be sick of it. But IMO the NAM trainer has a really intuitive GUI trainer which allows for different sized networks, at various sample rates.

The current state is that I will not add another neural network engine to HISE because of bloat but try to add compatibility of NAM files to RTNeural as suggested in this issue:

Add support for NAM files · Issue #143 · jatinchowdhury18/RTNeural

Hi Jatin, how hard would it be to add support for parsing the NAM file format? https://github.com/sdatkinson/NeuralAmpModelerCore Just from a quick peek at both sources the required layers are almost there (except for the wavenet layer w...

GitHub (github.com)

There seems to be some motivation by other developers to make this happen but it‘s not my best area of expertise and I have a few other priorities at the moment.

[Christoph Hart]

@Dan-Korneff said in Simple ML neural network:

@Christoph-Hart it should be a pytorch model.
This is the training script I'm testing with. It uses RTneural as a backend as well:
https://github.com/GuitarML/Automated-GuitarAmpModelling

Have you tried running it through this script?

Automated-GuitarAmpModelling/simple_modelToKeras.py at next · AidaDSP/Automated-GuitarAmpModelling

Contribute to AidaDSP/Automated-GuitarAmpModelling development by creating an account on GitHub.

GitHub (github.com)

[Dan Korneff]

@Christoph-Hart I just read the thread and found the link. Gonna give this a go first thing this morning.

[Dan Korneff]

@Christoph-Hart
https://github.com/AidaDSP/Automated-GuitarAmpModelling

There is already a script in Automated-GuitarAmpModelling named modelToKeras.
"a way to export models generated here in a format compatible with RTNeural"

I'll give both a try and report back.

I'm seeing that there is also a script to convert NAM dataset.

"NAM Dataset
Since I've received a bunch of request from the NAM community, I leave some infos here. Since the NAM models at the moment are not compatible with the inference engine used by rt-neural-generic (RTNeural), you can't use them with our plugin directly. But you can still use our training script and the NAM Dataset, so that you will be able to use the amplifiers that you are using on NAM with our plugin. In the end, training is 10mins on a Laptop with CUDA."

[Dan Korneff]

Learning curve is high on this one.
I've written a config script, prepared the audio files into a dataset, trained the model with dist_model_recnet.ph.
The model_utils.py script complained about how output_shape was being accessed, so I made a little tweak there.
In the end, it was able to convert the model to keras, but the layer dimensions are exporting as null.
Time for more beer and research

[Christoph Hart]

@Dan-Korneff yeah I tried to write the wavenet layer today for RTNeural, by porting it over from the NAM codebase, but I don't know either framework (or anything about writing inference engines lol), so it wasn't very fruitful.

Let me know if you get somewhere then we'll try to load it into the HISE neural engine.

[aaronventure]

@Dan-Korneff DId you have any luck running the colab for training? I upload input.wav and target.wav and get an error

[resonant]

I don't know if it helps, but Karanyi Sounds (uses HISE) also does machine learning using HISE Neural Network with colab, they shared this photo today.

As I see, if this Neural implementation is done very well, it will be really popular among developers. Lots of people would love to use this latest technology in their software.

[Dan Korneff]

@Christoph-Hart It seems like so many projects are abandoned, even if it's relatively new. Still researching

[Dan Korneff]

@aaronventure The tech is evolving so much that the scripts on google collab break just about every time there is an update. I got the google collab script to work for Proteus, but moved to local processing cause my GPU is better than the ones provided by google.

[Dan Korneff]

@resonant Invite them to the conversation

[ccbl]

@Dan-Korneff In this instance the dataset refers to the input output audio pairs that NAM uses for it's training, not the resulting model. Basically they're saying they added info in their training script that can detect the NAM audio pairs and train and Aida-X model based on those.

[ccbl]

@resonant that's awesome. Would love to pick their brains and see if we can get it up and running for the rest of us.

[Dan Korneff]

Here's where I'm at with the process:
https://gitlab.korneff.co/publicgroup/hise-neuralnetworktrainingscripts

I have used the scripts from https://github.com/AidaDSP/Automated-GuitarAmpModelling/tree/aidadsp_devel as a starting point.

This will allow you to create a dataset from your input/output audio file, train the model from the dataset, and then convert the model to Keras so you can use it RTNeural.

@Christoph-Hart The final model is making HISE crash. I thought I was doing something wrong because some of the values for "shape" are null, but I've downloaded other files created with the source script and they are null in the same places.

Here's one for example:
JMP Low Input.json

It's possible that the script is not formatting the json properly, but I don't know what a correct model looks like to compare to.

[Christoph Hart]

@Dan-Korneff That's the JSON from the sine generator example:

{
  "layers": "SineModel(\n  (network): Sequential(\n    (0): Linear(in_features=1, out_features=8, bias=True)\n    (1): Tanh()\n    (2): Linear(in_features=8, out_features=4, bias=True)\n    (3): Tanh()\n    (4): Linear(in_features=4, out_features=1, bias=True)\n  )\n)",
  "weights": {
    "network.0.weight": [
      [
        1.046385407447815
      ],
      [
        1.417808413505554
      ],
      [
        0.9530450105667114
      ],
      [
        1.118412375450134
      ],
      [
        -2.003693819046021
      ],
      [
        1.485351920127869
      ],
      [
        -1.323277235031128
      ],
      [
        -1.482439756393433
      ]
    ],
    "network.0.bias": [
      -0.4485535621643066,
      -1.284180760383606,
      1.995141625404358,
      -1.036547422409058,
      0.2926304638385773,
      0.4770179986953735,
      0.3244697153568268,
      0.4108103811740875
    ],
    "network.2.weight": [
      [
        -1.791297316551208,
        -0.3762974143028259,
        -0.3934035897254944,
        0.1596113294363022,
        0.5510663390159607,
        -1.115586280822754,
        0.678738534450531,
        1.327430963516235
      ],
      [
        0.3413433432579041,
        1.86607301235199,
        -0.217528447508812,
        2.568317174911499,
        0.3797312676906586,
        -0.1846907883882523,
        0.04422684758901596,
        -0.0883311927318573
      ],
      [
        0.3113365173339844,
        0.8516308069229126,
        -0.6042391061782837,
        0.9669480919837952,
        -1.354665994644165,
        0.1234097927808762,
        -1.171357274055481,
        -0.9616029858589172
      ],
      [
        -0.5073869824409485,
        -0.7385743856430054,
        0.3118444979190826,
        -0.9642266035079956,
        1.899434208869934,
        -0.1497718989849091,
        1.684132099151611,
        0.895214855670929
      ]
    ],
    "network.2.bias": [
      -0.6971003413200378,
      0.3228396475315094,
      -0.6209602355957031,
      0.1816271394491196
    ],
    "network.4.weight": [
      [
        -0.9233435988426208,
        1.108147859573364,
        -0.8966623544692993,
        0.394584596157074
      ]
    ],
    "network.4.bias": [
      0.06727132201194763
    ]
  }
}

So apparently it doesn't resolve the python code for defining the layer composition but uses a single string that is parsed. That's the output of a custom python script I wrote and run on a model built with TorchStudio, but if your model is "the standard" way, I'll make sure that it loads correctly too as these things look like syntactic sugar to me.

[Christoph Hart]

Here's the link to the tutorial again:

hise_tutorial/NeuralNetworkExample/Scripts/python at master · christophhart/hise_tutorial

The Tutorial project for HISE. Contribute to christophhart/hise_tutorial development by creating an account on GitHub.

GitHub (github.com)

But I realized your example looks more or less like the Tensorflow model in this directory. Which method are you using for loading the model?

[Lindon]

@Dan-Korneff said in Simple ML neural network:

@resonant Invite them to the conversation

@resonant
yeah, maybe - they asked me to get re-involved with them on some ML stuff as they were a bit stuck.....it didnt go anywhere, so they may well still be stuck or they found someone else to do the coding for them....your call

[Dan Korneff]

@Christoph-Hart The structure does look like TensorFlow, but the script is using torch libraries to create the model:

torch
torch.optim
torch.cuda
torch.optim.Adam
torch.optim

I was using this example code to load the model:

const var pt = Engine.createNeuralNetwork("PytorchNetwork");;

// Load the model layout & weights that were exported as JSON
const var modelJSON = pythonRoot.getChildFile("model_keras.json").loadAsObject();

// Load the model & weights:
pt.loadPytorchModel(modelJSON);

when I load the model into the Neural node, HISE checks out.

I've run through the MNIST dataset tutorial a couple times to get a basic idea of how TorchStudio works, but I'm not sure how to adapt the scripts to work there.

[ccbl]

I asked Jatin to have a quick look through the thread to see if he could see any issues, he just had this to say.

"Hmmm, it seems to me that the model JSON file that is being loaded into the "Neural Node" is structured as a TensorFlow-style JSON file, but it's being loaded with the HISE's loadPytorchModel() method? I don't really know what the Neural Node does internally, so idk how much I can help beyond that."

[Dan Korneff]

Using loadTensorFlowModel() was indeed the solution. I'll try to make some tutorials on training and loading models this weekend.