DynaML: ML + JVM + Scala

Join the chat at https://gitter.im/DynaML/Lobby Build Status codecov status

DynaML is a Scala & JVM Machine Learning toolbox for research, education & industry.

Plot3d Plot2d


  • Interactive. Don't want to create Maven/sbt project skeletons every time you want to try out ideas? Create and execute scala worksheets in the DynaML shell. DynaML comes packaged with a customized version of the Ammonite REPL, with auto-complete, file operations and scripting capabilities.

  • End to End. Create complex pre-processing pipelines with the data pipes API, train models (deep nets, gaussian processes, linear models and more), optimize over hyper-parameters, evaluate model predictions and visualise results.

  • Enterprise Friendly. Take advantage of the JVM and Scala ecosystem, use Apache Spark to write scalable data analysis jobs, Tensorflow for deep learning, all in the same toolbox.

Getting Started

Platform Compatibility

Currently, only *nix and OSX platforms are supported.

DynaML is compatible with Scala 2.11


Easiest way to install DynaML is cloning & compiling from the github repository. Please take a look at the installation instructions in the user guide, to make sure that you have the pre-requisites and to configure your installation.

CIFAR in 100 lines

Below is a sample script where we train a neural network of stacked Inception cells on the CIFAR-10 image classification task.

import ammonite.ops._
import io.github.mandar2812.dynaml.pipes.DataPipe
import io.github.mandar2812.dynaml.tensorflow.data.AbstractDataSet
import io.github.mandar2812.dynaml.tensorflow.{dtflearn, dtfutils}
import io.github.mandar2812.dynaml.tensorflow.implicits._
import org.platanios.tensorflow.api._
import org.platanios.tensorflow.api.learn.layers.Activation
import org.platanios.tensorflow.data.image.CIFARLoader
import java.nio.file.Paths

val tempdir = home/"tmp"

val dataSet = CIFARLoader.load(

val tf_dataset = AbstractDataSet(
  dataSet.trainImages, dataSet.trainLabels, dataSet.trainLabels.shape(0),
  dataSet.testImages, dataSet.testLabels, dataSet.testLabels.shape(0))

val trainData =

println("Building the model.")
val input = tf.learn.Input(

val trainInput = tf.learn.Input(UINT8, Shape(-1))

val relu_act = DataPipe[String, Activation](tf.learn.ReLU(_))

val architecture = tf.learn.Cast("Input/Cast", FLOAT32) >>
    channels = 3,  Seq.fill(4)(10), 
    relu_act)(layer_index = 1) >>
    channels = 40, Seq.fill(4)(5), 
    relu_act)(layer_index = 2) >>
  tf.learn.Flatten("Layer_3/Flatten") >>
  dtflearn.feedforward(256)(id = 4) >>
  tf.learn.ReLU("Layer_4/ReLU", 0.1f) >>
  dtflearn.feedforward(10)(id = 5)

val trainingInputLayer = tf.learn.Cast("TrainInput/Cast", INT64)

val loss = 
  tf.learn.SparseSoftmaxCrossEntropy("Loss/CrossEntropy") >> 
    tf.learn.Mean("Loss/Mean") >> 
    tf.learn.ScalarSummary("Loss/Summary", "Loss")

val optimizer = tf.train.Adam(0.1)

val summariesDir = Paths.get((tempdir/"cifar_summaries").toString())

val (model, estimator) = dtflearn.build_tf_model(
  architecture, input, trainInput, trainingInputLayer,
  loss, optimizer, summariesDir, dtflearn.max_iter_stop(500),
  100, 100, 100)(
  trainData, true)

def accuracy(predictions: Tensor, labels: Tensor): Float =

val (trainingPreds, testPreds): (Option[Tensor], Option[Tensor]) =
    Tensor, Output, DataType, Shape, Output,
    Tensor, Output, DataType, Shape, Output,
    Tensor, Tensor](
    data = tf_dataset,
    pred_flags = (true, true),
    buff_size = 20000)

val (trainAccuracy, testAccuracy) = (
  accuracy(trainingPreds.get, dataSet.trainLabels),
  accuracy(testPreds.get, dataSet.testLabels))

print("Train accuracy = ")

print("Test accuracy = ")

Support & Community