Komputo

Komputo is an experimental library to provide a Kotlin idiomatic API to build, load and execute computation graphs, initially using Tensorflow but potentially on other backends.

It relies on Tensorflow for Java and it contains an experimental DSL to build and execute computation graphs in a structured and type-safe manner:

val model = trainableModel {

    sequential(inputShape = I[28, 28]) {
        conv2d(
            filterCount = 32,
            kernelSize = I[3, 3],
            activation = RELU
        )
        conv2d(
            filterCount = 64,
            kernelSize = I[3, 3],
            activation = RELU
        )
        maxPool2D(windowSize = I[2, 2])
        flatten()
        dropout(0.25)
        dense(128, activation = RELU)
        dropout(0.5)
        dense(10)
    }

    training {
        loss = crossEntropyWithLogits
        optimizer = Adagrad()
    }

}

session(model) {

    train {

        dataset = trainDataset.mapLabels { it.toOneHot(10) }

        epochs = 5
        batchSize = 128

        verbose()

    }
    
}

Examples using the MNIST dataset are available on the test folder. The MNIST dataset has to be available at the data directory, it can be downloaded from http://yann.lecun.com/exdb/mnist/.

To use GPU, compile Tensorflow for Java and set the native library path with the -Djava.library.path=./path_to_tensorflow JVM argument.

TomasVolker/komputo

Komputo