开源软件名称：

开源软件地址：

开源软件介绍：

MindSpore Operator

Experimental notice: This project is still experimental and only serves as a proof of concept for running MindSpore on Kubernetes. The current version of ms-operator is based on an early version of PyTorch Operator and TF Operator. Right now MindSpore supports running LeNet with MNIST dataset on a single node, while distributed training examples are conducted using Volcano.

• MindSpore Operator
  • Introduction of MindSpore and ms-operator
    • MindSpore docker image
    • Design
    • Overview of MindSpore in Kubeflow ecosystem
  • MindSpore CPU example
    • Prerequisites
    • Steps
  • Distributed GPU example using Volcano
    • Volcano prerequistites
    • MindSpore GPU example
  • Future Work
  • Community
  • Contributing
  • License

Introduction of MindSpore and ms-operator

MindSpore is a new open source deep learning training/inference framework thatcould be used for mobile, edge and cloud scenarios. MindSpore is designed toprovide development experience with friendly design and efficient execution forthe data scientists and algorithmic engineers, native support for Ascend AIprocessor, and software hardware co-optimization.

This project contains the specification and implementation of MSJob customresource definition. We will demonstrate running a walkthrough of creatingms-operator, as well as MNIST training job on Kubernetes with MindSpore-cpu image(x86 CPU build version) on a single node. More completed features will bedeveloped in the coming days.

This project defines the following:

• The ms-operator
• A way to deploy the operator
• MindSpore LeNet MNIST training example
• MindSpore distributed GPU example using Volcano
• etc.

MindSpore docker image

Please refer to MindSpore docker image introductionfor details.

Design

The yaml file we used to create our MNIST training job is defined as follows:

apiVersion: v1kind: Podmetadata:  name: msjob-mnistspec:  containers:  - image: mindspore/mindspore-cpu:0.1.0-alpha    imagePullPolicy: IfNotPresent    name: msjob-mnist    command: ["/bin/bash", "-c", "python /tmp/test/MNIST/lenet.py"]    volumeMounts:      - name: training-result        mountPath: /tmp/result      - name: ms-mnist        mountPath: /tmp/test  restartPolicy: OnFailure  volumes:    - name: training-result      emptyDir: {}    - name: ms-mnist      hostPath:        path: /root/gopath/src/gitee.com/mindspore/ms-operator/examples/

Overview of MindSpore in Kubeflow ecosystem

The high-level view of how MindSpore fits in the ecosystem of Kubeflow and itscomponents.

MindSpore CPU example

Prerequisites

• Ubuntu: 16.04.6 LTS
• Helm and Tiller: v2.9.0
• go: go1.12.1
• docker: v18.06.1-ce
• Kubernetes: v1.14.0

Steps

First, pull the ms-operator image from Docker Hub:

docker pull mindspore/ms-operator:latest

Or you can build the ms-operator image on local machine:

go build -ldflags '-w -s' -o ms-operator cmd/ms-operator.v1/main.godocker build -t mindspore/ms-operator .

After the installation, check the image status using docker images command:

REPOSITORY                        TAG                   IMAGE ID            CREATED             SIZEmindspore/ms-operator             latest                4a17028de3d3        5 minutes ago       97.8MB

The MindSpore image we download from docker hub is 0.1.0-alpha version:

REPOSITORY                        TAG                   IMAGE ID            CREATED             SIZEmindspore/mindspore-cpu           0.1.0-alpha           ef443be923bc        3 hours ago         1.05GB

MindSpore supports heterogeneous computing including multiple hardware andbackends (CPU, GPU, Ascend), the device_target of MindSpore isAscend by default but we will use the CPU version here.

Install the msjob crd, ms-operator deployment and pod:

RBAC=true # set false if you do not have an RBAC clusterhelm install ms-operator-chart/ -n ms-operator --set rbac.install=${RBAC} --wait --replace

Using helm status ms-operator command to check generated resources:

LAST DEPLOYED: Tue Mar 24 11:36:51 2020NAMESPACE: defaultSTATUS: DEPLOYEDRESOURCES:==> v1beta1/CustomResourceDefinitionNAME                 AGEmsjobs.kubeflow.org  1d==> v1beta1/DeploymentNAME         DESIRED  CURRENT  UP-TO-DATE  AVAILABLE  AGEms-operator  1        1        1           1          1d==> v1/Pod(related)NAME                          READY  STATUS   RESTARTS  AGEms-operator-7b5b457d69-dpd2b  1/1    Running  0         1d

We will do a MNIST training to check the eligibility of MindSpore running onKubernetes:

cd examples/ && kubectl apply -f ms-mnist.yaml

The job is simply importing MindSpore packages, the dataset is already included in the MNIST_Data folder, executing only one epoch and printing result which should only consume little time. After the job completed, you should be able to check the job status and see the result logs. You can check the source training code in examples/ folder.

kubectl get pod msjob-mnist && kubectl logs msjob-mnist

The expected output is:

NAME          READY   STATUS      RESTARTS   AGEmsjob-mnist   0/1     Completed   0          3h53m============== Starting Training ==============epoch: 1 step: 1, loss is 2.3005836epoch: 1 step: 2, loss is 2.2978227epoch: 1 step: 3, loss is 2.3004227epoch: 1 step: 4, loss is 2.3054247epoch: 1 step: 5, loss is 2.3068798epoch: 1 step: 6, loss is 2.298408epoch: 1 step: 7, loss is 2.3055573epoch: 1 step: 8, loss is 2.2998955epoch: 1 step: 9, loss is 2.3028255epoch: 1 step: 10, loss is 2.2972553

Distributed GPU example using Volcano

The source code of the example can be found here.

Volcano Prerequisites

• Kubernetes: v1.16.6
• NVIDIA Docker: 2.3.0
• NVIDIA/k8s-device-plugin: 1.0.0-beta6
• NVIDIA drivers: 418.39
• CUDA: 10.1

Install Volcano: kubectl apply -f https://raw.githubusercontent.com/volcano-sh/volcano/master/installer/volcano-development.yaml

MindSpore GPU example

Using a modified image which the openssh-server is installed from the official MindSpore GPU image. To check the eligibility of MindSpore GPU's ability to communicate with other processes, we leverage the mpimaster and mpiworker task spec of Volcano. In this example, we launch one mpimaster and two mpiworkers, the python script is taken from MindSpore Gitee README, which is also modified to be able to run parallelly.

cd to example/MindSpore-example/mindspore_gpu folder, then:
pull image: docker pull lyd911/mindspore-gpu-example:0.2.0
to run: kubectl apply -f mindspore-gpu.yaml
to check result: kubectl logs mindspore-gpu-mpimster-0

The expected output should be (2*3) of multi-dimensional array.

Future Work

Kubeflow just announced its first major1.0 release recently with the graduation of a core set of stable applicationsincluding:

• Kubeflow's UI
• Jupyter notebook controller and web app
• Tensorflow Operator(TFJob), and PyTorch Operator for distributed training
• kfctl for deployment and upgrade
• etc.

The MindSpore community is driving to collaborate with the Kubeflow communityas well as making the ms-operator more complex, well-organized and itsdependencies up-to-date. All these components make it easy for machine learningengineers and data scientists to leverage cloud assets (public or on-premise)for machine learning workloads.

MindSpore is also looking forward to enable users to use Jupyter to developmodels. Users in the future can use Kubeflow tools like fairing (Kubeflow’spython SDK) to build containers and create Kubernetes resources to train theirMindSpore models.

Once training completed, users can use KFServingto create and deploy a server for inference thus completing the life cycle ofmachine learning.

Community

• MindSpore Slack - Ask questions and find answers.

Contributing

Welcome contributions. See our Contributor Wiki for more details.

License

Apache License 2.0