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