logicalclocks/hops-tensorflow

HopsYARN Tensorflow Framework.

Hops-TensorFlow

Easy distributed TensorFlow on Hops Hadoop. See our Python package yarntf repository for examples.

Compatibility

Since GPU support was added, Hops-TensorFlow is not compatible with Apache Hadoop. If you want to use Hops-TensorFlow with Apache Hadoop, either use v0.0.1 or remove Hops dependencies from the code.

• Tested with Python 2.7
• yarntf examples adapted for TensorFlow v1.0.0+

Before building

Install the Python package, run pip install yarntf on your machine and your cluster.

How to build

With Hops Hadoop and its dependencies installed: mvn clean install -Pndb

License

Hops-TensorFlow is released under an Apache 2.0 license.

hops-tensorflow's Issues

Client#main is broken

Since this change:

   public Client() throws Exception {
 -    this(new YarnConfiguration());
 +    this.appMasterMainClass = ApplicationMaster.class.getName();
    }

Client#main is broken and therefore yarntf-submit too. The main method uses the empty constructor and now essentially lacks the following code:

    conf = new YarnConfiguration();
    yarnClient = YarnClient.createYarnClient();
    yarnClient.init(conf);
    yarnClient.start();
    opts = createOptions();