Java and Jython/Python interface

Download

If you do not want to install the whole GraphLab C++ installation, you can simply download the Java interface from here: graphlab_java_20101210.tar.gz. We currently provide 64-bit Linux and Mac binary versions of the native library. For other platforms, you need to build your own (see README.txt for instructions).

Consult README.txt for instructions on how to build the Java API and how to run the example applications.

(Java interface is also located as part of the main source distribution under directory extapis).

Java API for GraphLab

• 1. Create Graph. Create a graph encapsulating the data and dependencies. To encapsulate any type of data in a vertex, create a custom subclass of graphlab.Vertex. Similarly for edges, simply subclass graphlab.Edge. Basic graph class is graphlab.SparseGraph. Use methods addVertex() and addEdge() to construct the graph. (You may also create your own implementation of graphlab.Graph, but this is necessary only if you need a highly compressed graph representation or an implicitly created graph).
• 2. Update functions. Implement update functions by creating implementations of abstract class graphlab.UpdateFunction. Interface is similar to the C++ update function, with the exception of currently not supporting shared data
• 3. Run Graphlab. Create instance of graphlab.wrapper.GraphLabJNIWrapper. Use its setX-methods to configure your GraphLab session and provide the graph with setGraph() method. Call start() to execute a GraphLab session. Example:

         GraphLabJNIWrapper jniWrapper = new GraphLabJNIWrapper(updateFunction);
          jniWrapper.setScopeType("edge");
          jniWrapper.setScheduler("multiqueue_fifo");
          jniWrapper.setGraph(graph);
          jniWrapper.scheduleAll();
          jniWrapper.start();
     

Documentation

Examples

Python/Jython API

To use the Jython interface, you only need to install the Java Graphlab API (instructions above). Full GraphLab Java API is available for the Python scripts.

• 1. Script for configuring the GraphLab parameters. This can be empty as well. Global variable graphlab is a handle to the GraphLabJNIWrapper object. For available methods, see GraphLabJNIWrapper API doc.
• 2. Script for creating the graph, or loading it from a file. Object graph is a handle to the graph object. For vertex and edge data, you can use any python data type! Examples:

     graph.addEdge(0,1, 3.2)
     graph.addEdge(i, j, [0.1, 12, "test"])
     graph.addVertex(MyVertexClass())
   

  Vertex and Edge data is stored in field "value" in the corresponding object. Example: graph.getVertex(vertexid).value.
• 3. Script for the update function. Global objects scheduler and scope are provided. They implement the interfaces: graphlab.Scope and graphlab.Scheduler.Example :

      # Python implementation of pagerank update function
      import math
      damping = 0.85
      
      def update(scope, scheduler):
          pvertex = scope.getVertex().value
          sumval = pvertex.rank * pvertex.selfedge + sum([e.value * scope.getNeighbor(e.from).value.rank for e in scope.getInboundEdges()])
          newval = (1-damping)/scope.getNumOfVertices() + damping*sumval
          if (abs(newval-pvertex.rank)>0.00001):
              scheduler.addTaskToOutbound(scope)
          pvertex.rank = newval
      
      update(scope, scheduler)
  
      

• 4. Post processing script. This is invoked when GraphLab completes. Computed graph is stored in object graph.

java -Xmx500m -Djava.library.path=../native/ graphlab.wrapper.PythonGraphlab pagerank_load.py pagerank_config.py pagerank_update.py pagerank_post.py testdata/stoch30x.csv

Examples

Tutorial: Lasso Linear Regression using Python GraphLab