JGNN

A native Java library for Graph Neural Networks.

🌀 Changes from 1.0.0

Up to 30% less memory and running time
Renamed GCNBuilder to FastBuilder
Neighbor attention and message passing
Sort pooling and graph classification

🎯 About

Graph Neural Networks (GNNs) are getting more and more popular; they can reason from relational information and perform inference from small datasets. JGNN provides GNNs to native Java applications and supports cross-platform machine learning, such as on Android, without specific hardware or firmware.

🚀 Features

Cross-platform (written in native java)
Lightweight
Minimal memory footprints
Symbolic model definition
Parallelized batching

⚡ Quickstart

To install the latest working version, you can include it as Maven or Gradle dependency by following the instructions of its JitPack distribution:

As a good out-of-the-box graph neural network for node classification, you can try the following architecture. To create your own graph adjacency matrices, node features, and labels look at the data creation tutorial.

Dataset dataset = new Cora();
Matrix adjacency = dataset.graph().setMainDiagonal(1).setToSymmetricNormalization();
Matrix nodeFeatures = dataset.features();
Matrix nodeLabels = dataset.labels();
Slice nodes = new Slice(0, nodeLabels.getRows()).shuffle(100);
long numClasses = nodeLabels.getCols();

ModelBuilder modelBuilder = new FastBuilder(adjacency, nodeFeatures)
		.config("reg", 0.005)
		.config("hidden", 16)
		.config("classes", numClasses)
		.layer("h{l+1}=relu(h{l}@matrix(features, hidden, reg)+vector(hidden))")
		.layer("h{l+1}=h{l}@matrix(hidden, classes)+vector(classes)")
		.rememberAs("0")
		.constant("a", 0.9)
		.layerRepeat("h{l+1} = a*(dropout(A, 0.5)@h{l})+(1-a)*h{0}", 10)
		.classify();

ModelTraining trainer = new ModelTraining()
		.setOptimizer(new Adam(0.01))
		.setEpochs(300)
		.setPatience(100)
		.setLoss(new CategoricalCrossEntropy())
		.setValidationLoss(new CategoricalCrossEntropy());

Model model = modelBuilder.getModel()
		.init(new XavierNormal())
		.train(trainer,
				nodes.samplesAsFeatures(), 
				nodeLabels, 
				nodes.range(0, 0.6), 
				nodes.range(0.6, 0.8));

Matrix output = model.predict(nodes.samplesAsFeatures()).get(0).cast(Matrix.class);
double acc = 0;
for(Long node : nodes.range(0.8, 1)) {
	Matrix nodeLabels = dataset.labels().accessRow(node).asRow();
	Tensor nodeOutput = output.accessRow(node).asRow();
	acc += nodeOutput.argmax()==nodeLabels.argmax()?1:0;
}
System.out.println("Acc\t "+acc/nodes.range(0.8, 1).size());

👍 Contributing

Feel free to contribute in any way, for example through the issue tracker. Please check out the contribution guidelines when bringing modifications to the code base.

📓 Citation

TBD