Distributed TensorFlow on Spark

First presented at the 2016 Spark Summit East: [Slide deck] (http://www.slideshare.net/arimoinc/distributed-tensorflow-scaling-googles-deep-learning-library-on-spark-58527889), [Presentation video] (https://www.youtube.com/watch?v=-QtcP3yRqyM), [Blog post] (https://arimo.com/machine-learning/deep-learning/2016/arimo-distributed-tensorflow-on-spark/)

TensorSpark productionalized in yarn-cluster mode

This latest version contains modifications/improvements that are mostly relevant to someone interested in taking TensorSpark to production in yarn-cluster mode (tested with a Hortonworks distribution [HDP 2.4] with CPU machines). For other deployment and machine types, the earlier version as of [Commit #62] (https://github.com/adatao/tensorspark/tree/2eae6732709884f08e800efa24653340f2f7997b) might still be a better option.

Summary of changes since [Commit #62] (https://github.com/adatao/tensorspark/tree/2eae6732709884f08e800efa24653340f2f7997b)

There are few minor improvements (see commits for details) and the following 2 major changes:

• tensorspark.py: Reading the testset from the HDFS instead (Avoiding the need to put the testset on local disk; we are putting training and test sets at the same location on the HDFS)
• parameterwebsocketclient.py: Find the machine that gets the Spark Driver in yarn-cluster mode (either way, there are some configs to be done here)

To run

1. zip pyfiles.zip ./parameterwebsocketclient.py ./parameterservermodel.py ./mnistcnn.py ./mnistdnn.py ./moleculardnn.py ./higgsdnn.py
2. spark-submit \
   --master yarn \
   --deploy-mode cluster \
   --queue default \
   --num-executors 3 \
   --driver-memory 20g \
   --executor-memory 60g \
   --executor-cores 8 \
   --py-files ./pyfiles.zip \
   ./tensorspark.py

Partial project layout:
tensorspark/gpu_install.sh - script to build tf from source with gpu support for aws
tensorspark/simplewebsocket.py - simple tornado websocket example
tensorspark/parameterservermodel.py - "abstract" model class that has all tensorspark required methods implemented
tensorspark/dnn.py - specific fully connected models for specific datasets
tensorspark/mnistcnn.py - convolutional model for mnist
tensorspark/parameterwebsocketclient.py - spark worker code
tensorspark/tensorspark.py - entry point and spark driver code