Counting 1,899 Big Data & Machine Learning Frameworks, Toolsets, and Examples...
Suggestion? Feedback? Tweet @stkim1

Last Commit
Oct. 18, 2017
Apr. 1, 2017

TensorFlow API for Scala

Build Status Chat Room

This library is a Scala API for It attempts to provide most of the functionality provided by the official Python API, while at the same type being strongly-typed and adding some new features. It is a work in progress and a project I started working on for my personal research purposes. Much of the API should be relatively stable by now, but things are still likely to change. That is why there is no official release of this library yet.

Please refer to the main website for documentation and tutorials. Here are a few useful links:

Main Features

  • Easy manipulation of tensors and computations involving tensors (similar to NumPy in Python):

    val t1 = Tensor( 1.2, 4.5)
    val t2 = Tensor(-0.2, 1.1)
    t1 + t2 == Tensor(1.0, 5.6)
  • High-level API for creating, training, and using neural networks. For example, the following code shows how simple it is to train a multi-layer perceptron for MNIST using TensorFlow for Scala. Here we omit a lot of very powerful features such as summary and checkpoint savers, for simplicity, but these are also very simple to use.

    import org.platanios.tensorflow.api._
    // Load and batch data using pre-fetching.
    val dataSet = MNISTLoader.load(Paths.get("/tmp"))
    val trainImages = DatasetFromSlices(dataSet.trainImages)
    val trainLabels = DatasetFromSlices(dataSet.trainLabels)
    val trainData =
    // Create the MLP model.
    val input = Input(UINT8, Shape(-1, 28, 28))
    val trainInput = Input(UINT8, Shape(-1))
    val layer = Flatten() >> Cast(FLOAT32) >> 
        Linear(128, name = "Layer_0") >> ReLU(0.1f) >>
        Linear(64, name = "Layer_1") >> ReLU(0.1f) >>
        Linear(32, name = "Layer_2") >> ReLU(0.1f) >>
        Linear(10, name = "OutputLayer")
    val trainingInputLayer = Cast(INT64)
    val loss = SparseSoftmaxCrossEntropy() >> Mean()
    val optimizer = GradientDescent(1e-6)
    val model = Model(input, layer, trainInput, trainingInputLayer, loss, optimizer)
    // Create an estimator and train the model.
    val estimator = Estimator(model)
    estimator.train(trainData, StopCriteria(maxSteps = Some(1000000)))

    And by changing a few lines to the following code, you can get checkpoint capability, summaries, and seamless integration with TensorBoard:

    loss = loss >> tf.learn.ScalarSummary("Loss")                  // Collect loss summaries for plotting
    val summariesDir = Paths.get("/tmp/summaries")                 // Directory in which to save summaries and checkpoints
    val estimator = Estimator(model, Configuration(Some(summariesDir)))
      trainData, StopCriteria(maxSteps = Some(1000000)),
        SummarySaverHook(summariesDir, StepHookTrigger(100)),      // Save summaries every 1000 steps
        CheckpointSaverHook(summariesDir, StepHookTrigger(1000))), // Save checkpoint every 1000 steps
      tensorBoardConfig = TensorBoardConfig(summariesDir))         // Launch TensorBoard server in the background

    If you now browse to while training, you can see the training progress:

  • Low-level graph construction API, similar to that of the Python API, but strongly typed wherever possible:

    import org.platanios.tensorflow.api._
    val inputs = tf.placeholder(FLOAT32, Shape(-1, 10))
    val outputs = tf.placeholder(FLOAT32, Shape(-1, 10))
    val predictions = tf.createWith(nameScope = "Linear") {
      val weights = tf.variable("weights", FLOAT32, Shape(10, 1), tf.zerosInitializer)
      tf.matmul(inputs, weights)
    val loss = tf.sum(tf.square(predictions - outputs))
    val optimizer = tf.train.AdaGrad(1.0)
    val trainOp = optimizer.minimize(loss)
  • Numpy-like indexing/slicing for tensors. For example:

    tensor(2 :: 5, ---, 1) // is equivalent to numpy's 'tensor[2:5, ..., 1]'
  • Efficient interaction with the native library that avoids unnecessary copying of data. All tensors are created and managed by the native TensorFlow library. When they are passed to the Scala API (e.g., fetched from a TensorFlow session), we use a combination of weak references and a disposing thread running in the background. Please refer to tensorflow/src/main/scala/org/platanios/tensorflow/api/utilities/Disposer.scala, for the implementation.


Funding for the development of this library has been generously provided by the following sponsors:

cmu_logo nsf_logo afosr_logo
CMU Presidential Fellowship National Science Foundation Air Force Office of Scientific Research
awarded to Emmanouil Antonios Platanios Grant #: IIS1250956 Grant #: FA95501710218