AXNN

AXNN is an open source Android SDK for running on-device inference on neural network models trained with XNN framework. It provides hardware acceleration by using RenderScript Android framework in order to run inference on device GPU, or parallelized across multiple CPU cores if GPU is not available.

You can find here the list of all supported neural network layers.

🚨 CAUTION: 🚨 Starting with Android 12 the RenderScript APIs are deprecated. They will continue to function, but device and component manufacturers will stop providing hardware acceleration support over time. I plan on releasing new AXNN SDK which uses OpenGL for GPU acceleration.

Setup

Make sure Maven Central repository is added in your project build.gradle file:

buildscript {
    repositories {
        mavenCentral()
    }
}

allprojects {
    repositories {
        mavenCentral()
    }
}

Include AXNN in your application build.gradle file:

dependencies {
    implementation 'com.github.markorakita:axnn-rs:1.0.0'
}

Getting Started

You can use NeuralNetFactory class to create some of the well known neural network architectures:

RenderScript rsContext = RenderScript.create(getApplicationContext());

NeuralNetRS neuralNet = NeuralNetFactory.createAlexNetDnn(m_rsContext);

Or you can create neural network from scratch:

RenderScript rsContext = RenderScript.create(getApplicationContext());

NeuralNetRS neuralNet = new NeuralNetRS(rsContext);
    
InputLayerRS inputLayer = new FeaturesInputLayerRS(rsContext, /*number of features:*/ 40);
neuralNet.addLayer(inputLayer);
    
StandardLayerRS standardLayer1 = new StandardLayerRS(rsContext, inputLayer.getActivationNumChannels(),
        inputLayer.getActivationDataWidth(), inputLayer.getActivationDataHeight(),
        /*number of neurons:*/ 128, ActivationFunction.ActivationFunctionType.ReLU);
neuralNet.addLayer(standardLayer1);
    
StandardLayerRS standardLayer2 = new StandardLayerRS(rsContext, standardLayer1.getActivationNumChannels(),
        standardLayer1.getActivationDataWidth(), standardLayer1.getActivationDataHeight(),
        /*number of neurons:*/ 10, ActivationFunction.ActivationFunctionType.Linear);
neuralNet.addLayer(standardLayer2);
    
SoftMaxLayerRS softMaxLayer = new SoftMaxLayerRS(rsContext, standardLayer2.getActivationDataBufferSize());
neuralNet.addLayer(softMaxLayer);
    
OutputLayerRS outputLayer = new OutputLayerRS(softMaxLayer.getActivationDataBufferSize(),
        OutputLayerRS.LossFunctionType.CrossEntropy);
neuralNet.addLayer(outputLayer);

Before you can use the created neural network for inference, you need to load trained XNN model (converted into AXNN format):

try (InputStream modelInputStream = rsContext.getApplicationContext().getAssets().open("model.xnnm"))
{
    // This can't be called from UI thread!
    neuralNet.loadModel(modelInputStream, true);
}
catch (IOException exc)
{
    // Handle exception...
}

You can then use network for inference by calling some of these functions:

try
{
    // These can't be called from UI thread!
    ClassificationResult imageResult = neuralNet.classifyImage(image); // Bitmap image
    ClassificationResult featuresResult = neuralNet.classifyFeatures(features); // float[] features
}
catch (Exception exc)
{
    // Handle exception...
}

You can also use network to extract features, which can then be fed into some other network for classification (used for transfer learning models):

try
{
    // These can't be called from UI thread!
    float[] imageFeatures = featuresNeuralNet.extractFeatures(image); // Bitmap image
    ClassificationResult result = classificationNeuralNet.classifyFeatures(imageFeatures);
}
catch (Exception exc)
{
    // Handle exception...
}