Getting Started

dissolvestruct is a distributed solver library for structured output prediction, based on Apache Spark.

The library is based on the primal-dual BCFW solver, allowing approximate inference oracles, and distributes this algorithm using the recent communication efficient CoCoA scheme. The interface to the user is the same as in the widely used SVMstruct in the single machine case.

Obtaining dissolvestruct

Checkout the project repository

$ git clone

Build the solver package first:

$ cd dissolve-struct-lib
$ sbt publish-local

Followed by building the example package (fat jar):

$ cd dissolve-struct-examples
$ sbt assembly

The fat jar produced by this build is placed in target/scala-2.10/DissolveStructExample-assembly-0.1-SNAPSHOT.jar of the dissolve-struct-examples directory. We’ll now refer to this as <examples-jar-path>.

(you might have to install sbt first, which can be done by running brew install sbt on a mac or sudo apt-get install sbt on Ubuntu, or follow the instructions available here)

Alternative: The binaries for both the solver and the examples package can be obtained at the releases page. But be warned, these binaries might not be up-to-date since the project is still in the development stage.

Running the examples

Obtain datasets

Obtain the datasets using:

$ python helpers/

(you might have to brew install wget first if on a mac. Additionally, pip -r requirements.txt to obtain the python dependencies.)

Executing through command line

Running the examples locally follows the format:

spark-1.X/bin/spark-submit \
	--class <class> \
	--master local \
	--driver-memory 2G \

PS: We recommend using the latest Spark version.

For example, the Binary classification example can be run within the dissolve-struct-examples directory using:

spark-1.X/bin/spark-submit \
	--class "ch.ethz.dalab.dissolve.examples.binaryclassification.COVBinary" \
	--master local \
	--driver-memory 2G \
Executing within Eclipse

To ease debugging and development, all examples can also directly be run within Eclipse by Run As | Scala Application. This does not require Spark binaries. See the section below how to set up the environment.

Within Eclipse, Spark can only be run in local mode since all the interactions need to be visible to Eclipse. In order to enable this, the SparkContext needs to be initialized by setting the master to local:

val conf = new SparkConf()

Setting up a development environment

We recommend using Eclipse for Scala, though a similar setup can also be done in Intellij IDEA. To create an Eclipse Scala project for our purposes, the following simple sbt command can be used. This generates the respective .classpath files needed.

cd dissolve-struct-lib
sbt eclipse

Similarly, for dissolve-struct-examples package too:

cd dissolve-struct-examples
sbt eclipse

The resulting projects from above can now be imported individually into Eclipse using: File | Import | Existing Projects into Workspace. Make sure you have search for nested projects checked, so you’ll have the choice to select both the dissolve-struct-lib and dissolve-struct-examples projects, if desired.

Currently Scala 2.10.4 is required by Spark. If Eclipse defaults to Scala 2.11 instead, this can cause an error “cross-compiled with an incompatible version of Scala”. The correct version needs to be set for both the projects by: Project Properties | Scala Compiler | Setting "Scala Installation" to "Latest 2.10 bundle". Alternatively, we recommend directly working with Eclipse IDE for Scala 2.10.4 from

Implementing your own application

This section assumes the reader’s familiarity with Structured SVMs and the notation. Please refer to the BCFW paper in the references section to learn more.

dissolvestruct is designed to solve most generic structured prediction tasks. In order to do this, the solver requires 3 items:

1. Functions

Our oracle-based solver requires 4 functions, which follow the same interface as SVMstruct.

  1. Joint Feature Map

    The joint feature map \[ \phi: \mathcal{X} \times \mathcal{Y} \rightarrow \mathbb{R}^d \] which encodes the input/output pairs.

  2. Structured Loss

    The structured loss function \( \Delta(Y_i, Y) \), over which the Bayes risk is minimized.

  3. Maximization Oracle

    An oracle which computes the most violating constraint by solving: \[ \hat{Y} = \arg \max_{Y \in \mathcal{Y}_i} \Delta(Y_i, Y) - \langle w, \psi_i(X_i, Y) \rangle \] where \[ \psi_i(X_i, Y) = \phi(X_i, Y_i) - \phi(X_i, Y) \]

  4. Prediction function (optional)

    A prediction function that computes:

    \[ \hat{Y} = \arg \max_{Y \in \mathcal{Y}_i} \langle w, \phi(X_i, Y) \rangle \]

    If no prediction function is defined, the solver will call the maximization oracle with the \(\Delta\) term set to zero.

These 4 functions need to be implemented by using a class/object which mixes-in the trait DissolveFunctions.

trait DissolveFunctions[X, Y] extends Serializable {

  def featureFn(x: X, y: Y): Vector[Double]

  def lossFn(yPredicted: Y, yTruth: Y): Double

  def oracleFn(model: StructSVMModel[X, Y], x: X, y: Y): Y

  def predictFn(model: StructSVMModel[X, Y], x: X): Y

2. Data

The data need to be in the form of an RDD consisting of LabeledObject[X,Y] objects. Refer to ch.ethz.dalab.dissolve.regression.LabeledObject for the definition.

3. Parameters

The parameters for the solver can be set using SolverOptions:

val solverOptions: SolverOptions[X, Y] = new SolverOptions()
solverOptions.lambda = 0.01    // regularization parameter

With these three items, a model can be trained as follows:

val trainer: StructSVMWithDBCFW[X, Y] =
	new StructSVMWithDBCFW[X, Y](trainDataRDD,

val model: StructSVMModel[X, Y] = trainer.trainModel()


We’d love to hear back on what you think, your experience or any remark on dissolvestruct. You can contact the authors Martin, Aurelien or Tribhuvanesh.


The CoCoA algorithmic framework is described in the following paper:

see also the binary classification and regression spark code here.

The (single machine) BCFW algorithm for structured prediction is described in the following paper:


This library is free and distributed under the Apache 2.0 License.