Skip to content

Latest commit

 

History

History
 
 

mnist

Table of Contents generated with DocToc

MNIST on Kubeflow

This example guides you through the process of taking an example model, modifying it to run better within Kubeflow, and serving the resulting trained model.

Prerequisites

Before we get started there are a few requirements.

Deploy Kubeflow

Follow the Getting Started Guide to deploy Kubeflow.

Local Setup

You also need the following command line tools:

To run the client at the end of the example, you must have requirements.txt intalled in your active python environment.

pip install -r requirements.txt

NOTE: These instructions rely on Github, and may cause issues if behind a firewall with many Github users.

Modifying existing examples

Many examples online use models that are unconfigurable, or don't work well in distributed mode. We will modify one of these examples to be better suited for distributed training and model serving.

Prepare model

There is a delta between existing distributed mnist examples and what's needed to run well as a TFJob.

Basically, we must:

  1. Add options in order to make the model configurable.
  2. Use tf.estimator.train_and_evaluate to enable model exporting and serving.
  3. Define serving signatures for model serving.

The resulting model is model.py.

Build and push model image.

With our code ready, we will now build/push the docker image.

DOCKER_URL=docker.io/reponame/mytfmodel:tag # Put your docker registry here
docker build . --no-cache  -f Dockerfile.model -t ${DOCKER_URL}

docker push ${DOCKER_URL}

Preparing your Kubernetes Cluster

With our data and workloads ready, now the cluster must be prepared. We will be deploying the TF Operator, and Argo to help manage our training job.

In the following instructions we will install our required components to a single namespace. For these instructions we will assume the chosen namespace is kubeflow.

kubectl config set-context $(kubectl config current-context) --namespace=kubeflow

Training your model

Local storage

Let's start by runing the training job on Kubeflow and storing the model in a local storage.

Fristly, refer to the document to create Persistent Volume(PV) and Persistent Volume Claim(PVC), the PVC name (${PVC_NAME}) will be used by pods of training and serving for local mode in steps below.

Enter the training/local from the mnist application directory.

cd training/local

Give the job a name to indicate it is running locally

kustomize edit add configmap mnist-map-training --from-literal=name=mnist-train-local

Point the job at your custom training image

kustomize edit set image training-image=$DOCKER_URL

Optionally, configure it to run distributed by setting the number of parameter servers and workers to use. The numPs means the number of Ps and the numWorkers means the number of Worker.

../base/definition.sh --numPs 1 --numWorkers 2

Set the training parameters, such as training steps, batch size and learning rate.

kustomize edit add configmap mnist-map-training --from-literal=trainSteps=200
kustomize edit add configmap mnist-map-training --from-literal=batchSize=100
kustomize edit add configmap mnist-map-training --from-literal=learningRate=0.01

To store the the exported model and checkpoints model, configure PVC name and mount piont.

kustomize edit add configmap mnist-map-training --from-literal=pvcName=${PVC_NAME}
kustomize edit add configmap mnist-map-training --from-literal=pvcMountPath=/mnt

Now we need to configure parameters and telling the code to save the model to PVC.

kustomize edit add configmap mnist-map-training --from-literal=modelDir=/mnt
kustomize edit add configmap mnist-map-training --from-literal=exportDir=/mnt/export

You can now submit the job

kustomize build . |kubectl apply -f -

And you can check the job

kubectl get tfjobs -o yaml mnist-train-local

And to check the logs

kubectl logs mnist-train-local-chief-0

Using GCS

In this section we describe how to save the model to Google Cloud Storage (GCS).

Storing the model in GCS has the advantages:

  • The model is readily available after the job finishes

  • We can run distributed training

    • Distributed training requires a storage system accessible to all the machines

Enter the training/GCS from the mnist application directory.

cd training/GCS

Set an environment variable that points to your GCP project Id

PROJECT=<your project id>

Create a bucket on GCS to store our model. The name must be unique across all GCS buckets

BUCKET=distributed-$(date +%s)
gsutil mb gs://$BUCKET/

Give the job a different name (to distinguish it from your job which didn't use GCS)

kustomize edit add configmap mnist-map-training --from-literal=name=mnist-train-dist

Optionally, if you want to use your custom training image, configurate that as below.

kustomize edit set image training-image=$DOCKER_URL:$TAG

Next we configure it to run distributed by setting the number of parameter servers and workers to use. The numPs means the number of Ps and the numWorkers means the number of Worker.

../base/definition.sh --numPs 1 --numWorkers 2

Set the training parameters, such as training steps, batch size and learning rate.

kustomize edit add configmap mnist-map-training --from-literal=trainSteps=200
kustomize edit add configmap mnist-map-training --from-literal=batchSize=100
kustomize edit add configmap mnist-map-training --from-literal=learningRate=0.01

Now we need to configure parameters and telling the code to save the model to GCS.

MODEL_PATH=my-model
kustomize edit add configmap mnist-map-training --from-literal=modelDir=gs://${BUCKET}/${MODEL_PATH}
kustomize edit add configmap mnist-map-training --from-literal=exportDir=gs://${BUCKET}/${MODEL_PATH}/export

In order to write to GCS we need to supply the TFJob with GCP credentials. We do this by telling our training code to use a Google service account.

If you followed the getting started guide for GKE then a number of steps have already been performed for you

  1. We created a Google service account named ${DEPLOYMENT}-user

    • You can run the following command to list all service accounts in your project

      gcloud --project=${PROJECT} iam service-accounts list
      
  2. We stored the private key for this account in a K8s secret named user-gcp-sa

    • To see the secrets in your cluster

      kubectl get secrets
      
  3. We granted this service account permission to read/write GCS buckets in this project

    • To see the IAM policy you can do

      gcloud projects get-iam-policy ${PROJECT} --format=yaml
      
    • The output should look like the following

       bindings:
       ...
       - members:
         - serviceAccount:${DEPLOYMENT}-user@${PROJEC}.iam.gserviceaccount.com
           ...
         role: roles/storage.admin
         ...
       etag: BwV_BqSmSCY=
       version: 1
      

To use this service account we perform the following steps

  1. Mount the secret user-gcp-sa into the pod and configure the mount path of the secret.

    kustomize edit add configmap mnist-map-training --from-literal=secretName=user-gcp-sa
    kustomize edit add configmap mnist-map-training --from-literal=secretMountPath=/var/secrets
    
    • Note: ensure your envrionment is pointed at the same kubeflow namespace as the user-gcp-sa secret
  2. Next we need to set the environment variable GOOGLE_APPLICATION_CREDENTIALS so that our code knows where to look for the service account key.

    kustomize edit add configmap mnist-map-training --from-literal=GOOGLE_APPLICATION_CREDENTIALS=/var/secrets/user-gcp-sa.json
    
    • If we look at the spec for our job we can see that the environment variable GOOGLE_APPLICATION_CREDENTIALS is set.

       kustomize build .
      
       apiVersion: kubeflow.org/v1beta1
       kind: TFJob
       metadata:
         ...
       spec:
         tfReplicaSpecs:
           Chief:
             replicas: 1
             template:
               spec:
                 containers:
                 - command:
                   ..
                   env:
                   ...
                   - name: GOOGLE_APPLICATION_CREDENTIALS
                     value: /var/secrets/user-gcp-sa.json
                   ...
                 ...
           ...
      

You can now submit the job

kustomize build . |kubectl apply -f -

And you can check the job status

kubectl get tfjobs -o yaml mnist-train-dist

And to check the logs

kubectl logs -f mnist-train-dist-chief-0

Using S3

To use S3 we need to configure TensorFlow to use S3 credentials and variables. These credentials will be provided as kubernetes secrets and the variables will be passed in as environment variables. Modify the below values to suit your environment.

Enter the training/S3 from the mnist application directory.

cd training/S3

Give the job a different name (to distinguish it from your job which didn't use S3)

kustomize edit add configmap mnist-map-training --from-literal=name=mnist-train-dist

Optionally, if you want to use your custom training image, configurate that as below.

kustomize edit set image training-image=$DOCKER_URL:$TAG

Next we configure it to run distributed by setting the number of parameter servers and workers to use. The numPs means the number of Ps and the numWorkers means the number of Worker.

../base/definition.sh --numPs 1 --numWorkers 2

Set the training parameters, such as training steps, batch size and learning rate.

kustomize edit add configmap mnist-map-training --from-literal=trainSteps=200
kustomize edit add configmap mnist-map-training --from-literal=batchSize=100
kustomize edit add configmap mnist-map-training --from-literal=learningRate=0.01

Now we need to configure parameters telling the code to save the model to S3, replace ${S3_MODEL_PATH_URI} and ${S3_MODEL_EXPORT_URI} below with real value.

kustomize edit add configmap mnist-map-training --from-literal=modelDir=${S3_MODEL_PATH_URI}
kustomize edit add configmap mnist-map-training --from-literal=exportDir=${S3_MODEL_EXPORT_URI}

In order to write to S3 we need to supply the TensorFlow code with AWS credentials we also need to set various environment variables configuring access to S3.

  1. Define a bunch of environment variables corresponding to your S3 settings; these will be used in subsequent steps

    export S3_ENDPOINT=s3.us-west-2.amazonaws.com  #replace with your s3 endpoint in a host:port format, e.g. minio:9000
    export AWS_ENDPOINT_URL=https://${S3_ENDPOINT} #use http instead of https for default minio installs
    export AWS_ACCESS_KEY_ID=xxxxx
    export AWS_SECRET_ACCESS_KEY=xxxxx
    export AWS_REGION=us-west-2
    export BUCKET_NAME=mybucket
    export S3_USE_HTTPS=1 #set to 0 for default minio installs
    export S3_VERIFY_SSL=1 #set to 0 for defaul minio installs 
    
  2. Create a K8s secret containing your AWS credentials

    kustomize edit add secret aws-creds --from-literal=awsAccessKeyID=${AWS_ACCESS_KEY_ID} \
      --from-literal=awsSecretAccessKey=${AWS_SECRET_ACCESS_KEY}
    
  3. Pass secrets as environment variables into pod

    kustomize edit add configmap mnist-map-training --from-literal=awsSecretName=aws-creds
    kustomize edit add configmap mnist-map-training --from-literal=awsAccessKeyIDName=awsAccessKeyID
    kustomize edit add configmap mnist-map-training --from-literal=awsSecretAccessKeyName=awsSecretAccessKey
    
  4. Next we need to set a whole bunch of S3 related environment variables so that TensorFlow knows how to talk to S3

    kustomize edit add configmap mnist-map-training --from-literal=S3_ENDPOINT=${S3_ENDPOINT}
    kustomize edit add configmap mnist-map-training --from-literal=AWS_ENDPOINT_URL=${AWS_ENDPOINT_URL}
    kustomize edit add configmap mnist-map-training --from-literal=AWS_REGION=${AWS_REGION}
    kustomize edit add configmap mnist-map-training --from-literal=BUCKET_NAME=${BUCKET_NAME}
    kustomize edit add configmap mnist-map-training --from-literal=S3_USE_HTTPS=${S3_USE_HTTPS}
    kustomize edit add configmap mnist-map-training --from-literal=S3_VERIFY_SSL=${S3_VERIFY_SSL}
    
    • If we look at the spec for our job we can see that the environment variables related to S3 are set.

       kustomize build .
      
       apiVersion: kubeflow.org/v1beta1
       kind: TFJob
       metadata:
         ...
       spec:
         tfReplicaSpecs:
           Chief:
             replicas: 1
             template:
               spec:
                 containers:
                 - command:
                   ..
                   env:
                   ...
                   - name: AWS_REGION
                     value: us-west-2
                   - name: BUCKET_NAME
                     value: somebucket
                   ...
                 ...
           ...
      

You can now submit the job

kustomize build . |kubectl apply -f -

And you can check the job

kubectl get tfjobs -o yaml mnist-train-dist

And to check the logs

kubectl logs -f mnist-train-dist-chief-0

Monitoring

There are various ways to monitor workflow/training job. In addition to using kubectl to query for the status of pods, some basic dashboards are also available.

Tensorboard

Using GCS

Enter the monitoring/GCS from the mnist application directory.

cd monitoring/GCS

Configure TensorBoard to point to your model location

kustomize edit add configmap mnist-map-monitoring --from-literal=logDir=${LOGDIR}

Assuming you followed the directions above if you used GCS you can use the following value

LOGDIR=gs://${BUCKET}/${MODEL_PATH}

You need to point TensorBoard to GCP credentials to access GCS bucket with model.

  1. Mount the secret user-gcp-sa into the pod and configure the mount path of the secret.

    kustomize edit add configmap mnist-map-monitoring --from-literal=secretName=user-gcp-sa
    kustomize edit add configmap mnist-map-monitoring --from-literal=secretMountPath=/var/secrets
    
    • Setting this parameter causes a volumeMount and volume to be added to TensorBoard deployment
  2. Next we need to set the environment variable GOOGLE_APPLICATION_CREDENTIALS so that our code knows where to look for the service account key.

    kustomize edit add configmap mnist-map-monitoring --from-literal=GOOGLE_APPLICATION_CREDENTIALS=/var/secrets/user-gcp-sa.json     
    
    • If we look at the spec for TensorBoard deployment we can see that the environment variable GOOGLE_APPLICATION_CREDENTIALS is set.

      kustomize build .
      
       ...
       env:
       ...
       - name: GOOGLE_APPLICATION_CREDENTIALS
         value: /var/secrets/user-gcp-sa.json
      

Using S3

Enter the monitoring/S3 from the mnist application directory.

cd monitoring/S3

Configure TensorBoard to point to your model location

kustomize edit add configmap mnist-map-monitoring --from-literal=logDir=${LOGDIR}

Assuming you followed the directions above if you used S3 you can use the following value

LOGDIR=s3://${BUCKET}/${MODEL_PATH}

You need to point TensorBoard to AWS credentials to access S3 bucket with model.

  1. Pass secrets as environment variables into pod

    kustomize edit add configmap mnist-map-monitoring --from-literal=awsSecretName=aws-creds
    kustomize edit add configmap mnist-map-monitoring --from-literal=awsAccessKeyIDName=awsAccessKeyID
    kustomize edit add configmap mnist-map-monitoring --from-literal=awsSecretAccessKeyName=awsSecretAccessKey
    
  2. Next we need to set a whole bunch of S3 related environment variables so that TensorBoard knows how to talk to S3

    kustomize edit add configmap mnist-map-monitoring --from-literal=S3_ENDPOINT=${S3_ENDPOINT}
    kustomize edit add configmap mnist-map-monitoring --from-literal=AWS_ENDPOINT_URL=${AWS_ENDPOINT_URL}
    kustomize edit add configmap mnist-map-monitoring --from-literal=AWS_REGION=${AWS_REGION}
    kustomize edit add configmap mnist-map-monitoring --from-literal=BUCKET_NAME=${BUCKET_NAME}
    kustomize edit add configmap mnist-map-monitoring --from-literal=S3_USE_HTTPS=${S3_USE_HTTPS}
    kustomize edit add configmap mnist-map-monitoring --from-literal=S3_VERIFY_SSL=${S3_VERIFY_SSL}
    
    • If we look at the spec for TensorBoard deployment we can see that the environment variables related to S3 are set.

      kustomize build .
      
       ...
       spec:
         containers:
         - command:
           ..
           env:
           ...
           - name: AWS_REGION
             value: us-west-2
           - name: BUCKET_NAME
             value: somebucket
           ...
      

Deploying TensorBoard

Now you can deploy TensorBoard

kustomize build . | kubectl apply -f -

To access TensorBoard using port-forwarding

kubectl port-forward service/tensorboard-tb 8090:80

TensorBoard can now be accessed at http://127.0.0.1:8090.

Serving the model

The model code will export the model in saved model format which is suitable for serving with TensorFlow serving.

To serve the model follow the instructions below. The instructins vary slightly based on where you are storing your model (e.g. GCS, S3, PVC). Depending on the storage system we provide different kustomization as a convenience for setting relevant environment variables.

GCS

Here we show to serve the model when it is stored on GCS. This assumes that when you trained the model you set exportDir to a GCS URI; if not you can always copy it to GCS using gsutil.

Check that a model was exported

EXPORT_DIR=gs://${BUCKET}/${MODEL_PATH}/export
gsutil ls -r ${EXPORT_DIR}

The output should look something like

${EXPORT_DIR}/1547100373/saved_model.pb
${EXPORT_DIR}/1547100373/variables/:
${EXPORT_DIR}/1547100373/variables/
${EXPORT_DIR}/1547100373/variables/variables.data-00000-of-00001
${EXPORT_DIR}/1547100373/variables/variables.index

The number 1547100373 is a version number auto-generated by TensorFlow; it will vary on each run but should be monotonically increasing if you save a model to the same location as a previous location.

Enter the serving/GCS from the mnist application directory.

cd serving/GCS

Set a different name for the tf-serving.

kustomize edit add configmap mnist-map-serving --from-literal=name=mnist-gcs-dist

Set your model path

kustomize edit add configmap mnist-map-serving --from-literal=modelBasePath=${EXPORT_DIR} 

Deploy it, and run a service to make the deployment accessible to other pods in the cluster

kustomize build . |kubectl apply -f -

You can check the deployment by running

kubectl describe deployments mnist-gcs-dist

The service should make the mnist-gcs-dist deployment accessible over port 9000

kubectl describe service mnist-gcs-dist

S3

TODO: Add instructions

Local storage

The section shows how to serve the local model that was stored in PVC while training.

Enter the serving/local from the mnist application directory.

cd serving/local

Set a different name for the tf-serving.

kustomize edit add configmap mnist-map-serving --from-literal=name=mnist-service-local

Mount the PVC, by default the pvc will be mounted to the /mnt of the pod.

kustomize edit add configmap mnist-map-serving --from-literal=pvcName=${PVC_NAME}
kustomize edit add configmap mnist-map-serving --from-literal=pvcMountPath=/mnt

Configure a filepath for the exported model.

kustomize edit add configmap mnist-map-serving --from-literal=modelBasePath=/mnt/export

Deploy it, and run a service to make the deployment accessible to other pods in the cluster.

kustomize build . |kubectl apply -f -

You can check the deployment by running

kubectl describe deployments mnist-deploy-local

The service should make the mnist-deploy-local deployment accessible over port 9000.

kubectl describe service mnist-service

Web Front End

The example comes with a simple web front end that can be used with your model.

Enter the front from the mnist application directory.

cd front

To deploy the web front end

kustomize build . |kubectl apply -f -

Connecting via port forwarding

To connect to the web app via port-forwarding

POD_NAME=$(kubectl get pods --selector=app=web-ui --template '{{range .items}}{{.metadata.name}}{{"\n"}}{{end}}')

kubectl port-forward ${POD_NAME} 8080:5000  

You should now be able to open up the web app at http://localhost:8080.

Using IAP on GCP

If you are using GCP and have set up IAP then you can access the web UI at

https://${DEPLOYMENT}.endpoints.${PROJECT}.cloud.goog/${NAMESPACE}/mnist/

Conclusion and Next Steps

This is an example of what your machine learning can look like. Feel free to play with the tunables and see if you can increase your model's accuracy (increasing model-train-steps can go a long way).