Distributed TensorFlow

TensorFlow is an open source software library for numerical computation using data flow graphs, and it supports distributed computing, allowing 'data parallel' or 'model parallel' on different servers. This means data scientists can now scale out distributed training to 100s of GPUs using TensorFlow.

• https://www.tensorflow.org
• https://www.tensorflow.org/deploy/distributed

Prerequisites

• Kubernetes cluster v1.12+

Chart Details

This chart will create a TensorFlow cluster, and distribute a computation graph across that cluster.

Installing the Chart

• To install the chart with the release name mnist:

  $ helm install mnist incubator/distributed-tensorflow

• To install with custom values via file:

  $ helm install --values values.yaml mnist incubator/distributed-tensorflow
  

  Below is an example of the custom value file values.yaml with GPU support.

  worker:
    number: 2
    podManagementPolicy: Parallel
    image:
      repository: dysproz/distributed-tf
      tag: 1.6.0-gpu
    port: 9090
    gpuCount: 1
  
  ps:
    number: 2
    podManagementPolicy: Parallel
    image:
      repository: dysproz/distributed-tf
      tag: 1.6.0
      pullPolicy: IfNotPresent
    port: 8080
  
  # optimize for training
  hyperparams:
    batchsize: 20
    learningrate: 0.001
    trainsteps: 10000
  

Notice: you can check the details of docker image from Docker hub

Uninstalling the Chart

• To uninstall/delete the mnist deployment:

   $ helm delete mnist

The command removes all the Kubernetes components associated with the chart and deletes the release.

Configuration

The following table lists the configurable parameters of the Service Distributed Tensorflow chart and their default values.

Parameter	Description	Default
worker.image.repository	TensorFlow Worker Server's image repository	dysproz/distributed-tf
worker.image.tag	TensorFlow Worker Server's image tag	gpu
worker.image.pullPolicy	image pullPolicy for the worker	IfNotPresent
worker.gpuCount	Set the gpu to be allocated and allowed for the Pods	0
worker.env	key-value environment variables	None
ps.image.repository	TensorFlow Parameter Server's image repository	dysproz/distributed-tf
ps.image.tag	TensorFlow Parameter Server's image tag	1.7.0-gpu
ps.image.pullPolicy	image pullPolicy for the ps	IfNotPresent
ps.env	key-value environment variables	None
volumes	List of volumes defined in cluster (in standard k8s format)	host path to /tmp/mnist
volumeMounts	Volumes mounted into Pods (in standard k8s format)	host path volume mounted into /tmp/mnist-log
hyperparams.batchsize	batch size	20
hyperparams.learningrate	learning rate	0.001
hyperparams.trainsteps	train steps	0 (continuous run)
hyperparams.datadir	data directory	None
hyperparams.logdir	logging directory	None
hyperparams.hiddenunits	Hidden units in neural network	None