RTNeural Example

This repository contains an example of how to use the RTNeural library for running inference on a neural network in realtime. The example case shown use here is in the form of an audio plugin, built using the JUCE framework.

Building from source

To build from source, you must have CMake installed.

$ git clone --recursive https://github.com/jatinchowdhury18/RTNeural-example
$ cd RTNeural-example
$ cmake -Bbuild
$ cmake --build build --config Release

Builds will be output to build/RTNeuralExample_artefacts.

Creating the network

The audio plugin functions as a distortion effect, using a neural network to generate the distorted outut signal. The architecture and weights of the neural network are defined in neural_net.py, and then exported to neural_net_weights.json. The plugin then stores the json file as binary data, and loads the neural network at run-time.

C++ Implementation

The C++ implementation is fairly simple, but there are a few code snippets worth drawing attention to. Note that the documentation below describes the neural net implementation using RTNeural's run-time API. The code in plugin/ also contains example code for using RTNeural's compile-time API.

Linking to RTNeural with CMake

CMakeLists.txt:

# Set your choice of RTNeural backends here!
set(RTNEURAL_STL ON CACHE BOOL "Use RTNeural with this backend" FORCE)
add_subdirectory(modules/RTNeural)
include_directories(modules/RTNeural)
...
target_link_libraries(RTNeuralExample PUBLIC
    ...
    RTNeural
)

Note that to choose the backend used by RTNeural, you may instead set RTNEURAL_XSIMD or RTNEURAL_EIGEN.

Adding RTNeural::model as a member variable

plugin/Plugin.h

#include <RTNeural/RTNeural.h>

class RTNeuralExamplePlugin
{
...
private:
    std::unique_ptr<RTNeural::Model<float>> neuralNet[2];
};

Note that we create 2 models, since the plugin accepts 2 channels of audio (stereo).

Loading the network

MemoryInputStream jsonStream (BinaryData::neural_net_weights_json, BinaryData::neural_net_weights_jsonSize, false);
auto jsonInput = nlohmann::json::parse (jsonStream.readEntireStreamAsString().toStdString());
neuralNet[0] = RTNeural::json_parser::parseJson<float> (jsonInput);
neuralNet[1] = RTNeural::json_parser::parseJson<float> (jsonInput);

Preparing the network

neuralNet[0]->reset();
neuralNet[1]->reset();

Running inference

for (int ch = 0; ch < buffer.getNumChannels(); ++ch)
{
    auto* x = buffer.getWritePointer (ch);
    for (int n = 0; n < buffer.getNumSamples(); ++n)
    {
        float input[] = { x[n] };
        x[n] = neuralNet[ch]->forward (input);
    }
}

License

The code in this repository is licensed under the BSD 3-clause license. Enjoy!