End-to-end Kubeflow on GCP

Kubeflow on Google Cloud Platform

This guide walks you through an end-to-end example of Kubeflow on Google Cloud Platform (GCP). By working through the guide, you’ll learn how to deploy Kubeflow on Kubernetes Engine (GKE), train an MNIST machine learning model for image classification, and use the model for online inference (also known as online prediction).

Introductions

Overview of GCP and GKE

Google Cloud Platform (GCP) is a suite of cloud computing services running on Google infrastructure. The services include compute power, data storage, data analytics, and machine learning.

The Cloud SDK is a set of tools that you can use to interact with GCP from the command line, including the gcloud command and others.

Kubernetes Engine (GKE) is a managed service on GCP where you can deploy containerized applications. You describe the resources that your application needs, and GKE provisions and manages the underlying cloud resources automatically.

Here’s a list of the primary GCP services that you use when following this guide:

The model and the data

This tutorial trains a TensorFlow model on the MNIST dataset, which is the hello world for machine learning.

The MNIST dataset contains a large number of images of hand-written digits in the range 0 to 9, as well as the labels identifying the digit in each image.

After training, the model can classify incoming images into 10 categories (0 to 9) based on what it’s learned about handwritten images. In other words, you send an image to the model, and the model does its best to identify the digit shown in the image. Prediction UI

In the above screenshot, the image shows a hand-written 7. This image was the input to the model. The table below the image shows a bar graph for each classification label from 0 to 9, as output by the model. Each bar represents the probability that the image matches the respective label. Judging by this screenshot, the model seems pretty confident that this image is a 7.

The overall workflow

Here’s an overview of what you accomplish by following this guide:

  • Setting up Kubeflow in a GKE cluster.

  • Testing the code locally using a Jupyter notebook.

  • Training the model:

  • Using the model for prediction (inference):

    • Saving the trained model to Cloud Storage.
    • Using TensorFlow Serving to serve the model.
    • Running a simple web app to send a prediction request to the model and display the result.

Let’s get started!

Set up your environment

Download the project files

To simplify this tutorial, you can use a set of prepared files that include a TensorFlow application for training your model, a web UI to send prediction requests and display the results, and the Docker files to build runnable containers for the training and prediction applications. The project files are in the Kubeflow examples repository on GitHub.

Clone the project files and go to the directory containing the MNIST example:

cd ${HOME}
git clone https://github.com/kubeflow/examples.git
cd examples/mnist
WORKING_DIR=$(pwd)

As an alternative to cloning, you can download the Kubeflow examples repository zip file.

Set up your GCP account and SDK

Follow these steps to set up your GCP environment:

  1. Select or create a project on the GCP Console.
  2. Make sure that billing is enabled for your project. See the guide to modifying a project’s billing settings.
  3. Install the Cloud SDK.

Notes:

  • As you work through this tutorial, your project uses billable components of GCP. To minimise costs, follow the instructions to clean up your GCP resources when you’ve finished with them.
  • This guide assumes you want to manage your GCP environment on your own server rather than in the Cloud Shell environment. If you choose to use the Cloud Shell, some of the components are pre-installed in your shell.

Install kubectl

Run the following Cloud SDK command to install the kubectl command-line tool for Kubernetes:

gcloud components install kubectl

Install kustomize

Kubeflow makes use of kustomize to help manage deployments.

Make sure you have kustomize version 2.0.3 or later. See the kustomize installation guide for help with installing kustomize.

Set up some handy environment variables

Set up the following environment variables for use throughout the tutorial:

  1. Set your GCP project ID. In the command below, replace <YOUR-PROJECT-ID> with your project ID:

    export PROJECT=<YOUR-PROJECT-ID>
    gcloud config set project ${PROJECT}
    
  2. Set the zone for your GCP configuration. Choose a zone that offers the resources you need. See the guide to GCP regions and zones.

    • Ensure you have enough Compute Engine regional capacity. By default, the GKE cluster setup described in this guide requires 16 CPUs.
    • If you want a GPU, ensure your zone offers GPUs.

    For example, the following commands set the zone to us-central1-c:

    export ZONE=us-central1-c
    gcloud config set compute/zone ${ZONE}
    
  3. If you want a custom name for your Kubeflow deployment, set the DEPLOYMENT_NAME environment variable. Note that the name must be the same as the one you use in later steps of this tutorial when configuring the redirect URI for the OAuth client credentials. If you don’t set this environment variable, your deployment gets the default name of kubeflow:

    export DEPLOYMENT_NAME=kubeflow
    

Deploy Kubeflow

Follow the instructions in the guide to deploying Kubeflow on GCP, taking note of the following:

  • Set up OAuth client credentials and Cloud Identity-Aware Proxy (IAP) as prompted during the deployment process. So, do not choose the deployment option to skip IAP. IAP ensures you can connect securely to the Kubeflow web applications.
  • When setting up the authorized redirect URI for the OAuth client credentials, use the same value for the <deployment_name> as you used when setting up the DEPLOYMENT_NAME environment variable earlier in this tutorial.
  • Use the Kubeflow deployment UI as a quick way to set up a Kubeflow deployment on GCP. The getting-started guide describes how to use the deployment UI. If you want more control over the configuration of your deployment you can use the kfctl binary instead of the UI. The binary is also described in the getting-started guide.
  • Choose Kubeflow version v0.4.1 or later.

The following screenshot shows the Kubeflow deployment UI with hints about the value for each input field:

Prediction UI

When the cluster is ready, you can do the following:

  1. Connect your local kubectl session to the cluster:

    gcloud container clusters get-credentials \
        ${DEPLOYMENT_NAME} --zone ${ZONE} --project ${PROJECT}
    
  2. Switch to the kubeflow namespace to see the resources on the Kubeflow cluster:

    kubectl config set-context $(kubectl config current-context) --namespace=kubeflow
    
  3. Check the resources deployed in the kubeflow namespace:

    kubectl get all
    
  4. Access the Kubeflow UI, which becomes available at the following URI after several minutes:

    https://<deployment-name>.endpoints.<project>.cloud.goog/
    

The following screenshot shows the Kubeflow UI: Prediction UI

Notes:

  • When the deployment has finished, you should have a running cluster in the cloud ready to run your code. You can interact with the cluster either by using kubectl or by going to the GKE page on the GCP Console.

  • While the deployment is running, you can watch your resources appear on the GCP console:

  • It can take 10-15 minutes for the URI to become available. Kubeflow needs to provision a signed SSL certificate and register a DNS name.

    • If you own/manage the domain or a subdomain with Cloud DNS then you can configure this process to be much faster.
    • While you wait you can access Kubeflow services by using kubectl proxy and kubectl port-forward to connect to services in the cluster.

Create a Cloud Storage bucket

The next step is to create a Cloud Storage bucket to hold your trained model

Cloud Storage is a scalable, fully-managed object/blob store. You can use it for a range of scenarios including serving website content, storing data for archival and disaster recovery, or distributing large data objects to users via direct download. This tutorial uses Cloud Storage to hold the trained machine learning model and associated data.

Use the gsutil mb command to create a storage bucket. Your bucket name must be unique across all of Cloud Storage. The following commands create a bucket in the us-central1 region, which corresponds to the us-central1-c zone used earlier in the tutorial:

export BUCKET_NAME=${PROJECT}-${DEPLOYMENT_NAME}-bucket
gsutil mb -c regional -l us-central1 gs://${BUCKET_NAME}

(Optional) Test the code in a Jupyter notebook

The sample you downloaded contains all the code you need. If you like, you can experiment with and test the code in a Jupyter notebook.

The Kubeflow deployment includes services for spawning and managing Jupyter notebooks.

  1. Follow the Kubeflow notebooks setup guide to create a Jupyter notebook server and open the Jupyter UI. Accept the default settings when configuring your notebook server. The default configuration gives you a standard CPU image with a recent version of TensorFlow.

  2. Create a new notebook by clicking New > Python 2 on the Jupyter dashboard.

    You can read about using notebooks in the Jupyter documentation.

  3. Copy the code from your sample model at ${WORKING_DIR}/model.py and paste the code into a cell in your Jupyter notebook.

  4. Run the cell in the notebook. You should see output directly beneath the notebook cell, something like this:

    INFO:tensorflow:TF_CONFIG {}
    INFO:tensorflow:cluster=None job_name=None task_index=None
    INFO:tensorflow:Will export model
    Extracting /tmp/data/train-images-idx3-ubyte.gz
    Extracting /tmp/data/train-labels-idx1-ubyte.gz
    Extracting /tmp/data/t10k-images-idx3-ubyte.gz
    Extracting /tmp/data/t10k-labels-idx1-ubyte.gz
    WARNING:tensorflow:Using temporary folder as model directory: /tmp/tmpfNskyK
    INFO:tensorflow:Using config: {'_save_checkpoints_secs': None, '_session_config': None, '_keep_checkpoint_max': 5, '_task_type': 'worker', '_global_id_in_cluster': 0, '_is_chief': True, '_cluster_spec': <tensorflow.python.training.server_lib.ClusterSpec object at 0x7f3e975ea550>, '_evaluation_master': '', '_save_checkpoints_steps': 1000, '_keep_checkpoint_every_n_hours': 10000, '_service': None, '_num_ps_replicas': 0, '_tf_random_seed': None, '_master': '', '_device_fn': None, '_num_worker_replicas': 1, '_task_id': 0, '_log_step_count_steps': 100, '_model_dir': '/tmp/tmpfNskyK', '_train_distribute': None, '_save_summary_steps': 100}
    Train and evaluate
    INFO:tensorflow:Running training and evaluation locally (non-distributed).
    INFO:tensorflow:Start train and evaluate loop. The evaluate will happen after every checkpoint. Checkpoint frequency is determined based on RunConfig arguments: save_checkpoints_steps 1000 or save_checkpoints_secs None.
    INFO:tensorflow:Calling model_fn.
    INFO:tensorflow:Done calling model_fn.
    INFO:tensorflow:Create CheckpointSaverHook.
    INFO:tensorflow:Graph was finalized.
    INFO:tensorflow:Running local_init_op.
    INFO:tensorflow:Done running local_init_op.
    INFO:tensorflow:Saving checkpoints for 0 into /tmp/tmpfNskyK/model.ckpt.
    INFO:tensorflow:loss = 2.3082201, step = 1
    INFO:tensorflow:global_step/sec: 6.81158
    INFO:tensorflow:loss = 2.0083964, step = 101 (14.683 sec)
    INFO:tensorflow:Saving checkpoints for 200 into /tmp/tmpfNskyK/model.ckpt.
    INFO:tensorflow:Calling model_fn.
    INFO:tensorflow:Done calling model_fn.
    INFO:tensorflow:Starting evaluation at 2019-02-02-02:35:00
    INFO:tensorflow:Graph was finalized.
    INFO:tensorflow:Restoring parameters from /tmp/tmpfNskyK/model.ckpt-200
    INFO:tensorflow:Running local_init_op.
    INFO:tensorflow:Done running local_init_op.
    INFO:tensorflow:Evaluation [1/1]
    INFO:tensorflow:Finished evaluation at 2019-02-02-02:35:00
    INFO:tensorflow:Saving dict for global step 200: accuracy = 0.8046875, global_step = 200, loss = 0.79056215
    INFO:tensorflow:Saving 'checkpoint_path' summary for global step 200: /tmp/tmpfNskyK/model.ckpt-200
    INFO:tensorflow:Performing the final export in the end of training.
    INFO:tensorflow:Calling model_fn.
    INFO:tensorflow:Done calling model_fn.
    INFO:tensorflow:Signatures INCLUDED in export for Eval: None
    INFO:tensorflow:Signatures INCLUDED in export for Classify: None
    INFO:tensorflow:Signatures INCLUDED in export for Regress: None
    INFO:tensorflow:Signatures INCLUDED in export for Predict: ['serving_default', 'classes']
    INFO:tensorflow:Signatures INCLUDED in export for Train: None
    INFO:tensorflow:Restoring parameters from /tmp/tmpfNskyK/model.ckpt-200
    INFO:tensorflow:Assets added to graph.
    INFO:tensorflow:No assets to write.
    INFO:tensorflow:SavedModel written to: /tmp/tmpfNskyK/export/mnist/temp-1549074900/saved_model.pb
    INFO:tensorflow:Loss for final step: 0.70332366.
    Training done
    Export saved model
    INFO:tensorflow:Calling model_fn.
    INFO:tensorflow:Done calling model_fn.
    INFO:tensorflow:Signatures INCLUDED in export for Eval: None
    INFO:tensorflow:Signatures INCLUDED in export for Classify: None
    INFO:tensorflow:Signatures INCLUDED in export for Regress: None
    INFO:tensorflow:Signatures INCLUDED in export for Predict: ['serving_default', 'classes']
    INFO:tensorflow:Signatures INCLUDED in export for Train: None
    INFO:tensorflow:Restoring parameters from /tmp/tmpfNskyK/model.ckpt-200
    INFO:tensorflow:Assets added to graph.
    INFO:tensorflow:No assets to write.
    INFO:tensorflow:SavedModel written to: mnist/temp-1549074900/saved_model.pb
    Done exporting the model
    

    The above output indicates that the program retrieved the sample training data then trained the model for 200 steps, reaching a final accuracy level of 0.70332366.

If you want to play more with the code, try adjusting the number of training steps by setting TF_TRAIN_STEPS to a different value, such as 2000, or experiment with adjusting other parts of the code.

Prepare to run your training application on GKE

When you downloaded the project files into your ${WORKING_DIR} directory at the start of the tutorial, you downloaded the TensorFlow code for your training application. The code is in a Python file, model.py, in your ${WORKING_DIR} directory.

The model.py program does the following:

  • Downloads the MNIST dataset and loads it for use by the model training code.
  • Offers a choice between two models:

    • A two-layer convolutional neural network (CNN). This tutorial uses the CNN, which is the default model in model.py.
    • A linear classifier, not used in this tutorial.
  • Defines TensorFlow operations to train and evaluate the model.

  • Runs a number of training cycles.

  • Saves the trained model to a specified location, such as your Cloud Storage bucket.

Build the container for your training application

To deploy your code to Kubernetes, you must first build your local project into a Docker container image and push the image to Container Registry so that it’s available in the cloud.

  1. Create a version tag from the current UNIX timestamp, to be associated with your model each time it runs:

    export VERSION_TAG=$(date +%s)
    
  2. Set the path in Container Registry that you want to push the image to:

    export TRAIN_IMG_PATH=gcr.io/${PROJECT}/${DEPLOYMENT_NAME}-train:${VERSION_TAG}
    
  3. Build the Docker image for your working directory:

    docker build $WORKING_DIR -t $TRAIN_IMG_PATH -f $WORKING_DIR/Dockerfile.model
    

    The container is tagged with its eventual path in Container Registry, but it hasn’t been uploaded to Container Registry yet.

    If everything went well, your program is now encapsulated in a new container.

  4. Test the container locally:

    docker run -it ${TRAIN_IMG_PATH}
    

    You may see some warnings from TensorFlow about deprecated functionality. Then you should see training logs start appearing in your output, similar to these:

    Train and evaluate
    INFO:tensorflow:Running training and evaluation locally (non-distributed).
    INFO:tensorflow:Start train and evaluate loop. The evaluate will happen after 1 secs (eval_spec.throttle_secs) or training is finished.
    INFO:tensorflow:Calling model_fn.
    INFO:tensorflow:Done calling model_fn.
    INFO:tensorflow:Create CheckpointSaverHook.
    INFO:tensorflow:Graph was finalized.
    2019-02-02 04:17:20.655001: I tensorflow/core/platform/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
    INFO:tensorflow:Running local_init_op.
    INFO:tensorflow:Done running local_init_op.
    INFO:tensorflow:Saving checkpoints for 1 into /tmp/tmph861eL/model.ckpt.
    INFO:tensorflow:loss = 2.3235848, step = 1
    INFO:tensorflow:Saving checkpoints for 4 into /tmp/tmph861eL/model.ckpt.
    INFO:tensorflow:Loss for final step: 2.2987146.
    INFO:tensorflow:Calling model_fn.
    INFO:tensorflow:Done calling model_fn.
    INFO:tensorflow:Starting evaluation at 2019-02-02-04:17:21
    INFO:tensorflow:Graph was finalized.
    INFO:tensorflow:Restoring parameters from /tmp/tmph861eL/model.ckpt-4
    INFO:tensorflow:Running local_init_op.
    INFO:tensorflow:Done running local_init_op.
    INFO:tensorflow:Evaluation [1/1]
    INFO:tensorflow:Finished evaluation at 2019-02-02-04:17:22
    INFO:tensorflow:Saving dict for global step 4: accuracy = 0.1640625, global_step = 4, loss = 2.2869625
    INFO:tensorflow:Calling model_fn.
    INFO:tensorflow:Done calling model_fn.
    INFO:tensorflow:Create CheckpointSaverHook.
    INFO:tensorflow:Graph was finalized.
    INFO:tensorflow:Restoring parameters from /tmp/tmph861eL/model.ckpt-4
    INFO:tensorflow:Running local_init_op.
    INFO:tensorflow:Done running local_init_op.
    INFO:tensorflow:Saving checkpoints for 5 into /tmp/tmph861eL/model.ckpt.
    INFO:tensorflow:loss = 2.3025892, step = 5
    INFO:tensorflow:Saving checkpoints for 7 into /tmp/tmph861eL/model.ckpt.
    INFO:tensorflow:Loss for final step: 2.2834966.
    ...
    
  5. When you see log entries similar to those above, your model training is working. You can terminate the container with Ctrl+c.

Next, upload the container image to Container Registry so that you can run it on your GKE cluster.

  1. Run the following command to authenticate to Container Registry:

    gcloud auth configure-docker --quiet
    
  2. Push the container to Container Registry:

    docker push ${TRAIN_IMG_PATH}
    

    The push may take a few minutes to complete. You should see Docker progress updates in your command window.

  3. Wait until the process is complete, then you should see your new container image listed on the Container Registry page on the GCP console.

Prepare your training component to run on GKE

  1. Enter the training/GCS directory:

    cd ${WORKING_DIR}/training/GCS
    
  2. Give the job a different name (to distinguish it from your job which didn’t use Cloud Storage):

    kustomize edit add configmap mnist-map-training   --from-literal=name=mnist-train-dist
    
  3. Configure your custom training image:

    kustomize edit set image training-image=${TRAIN_IMG_PATH}
    
  4. Optionally configure it to run distributed by setting the number of parameter servers and workers to use. The numPs means the number of Ps (parameter server) and the numWorkers means the number of worker:

    ../base/definition.sh --numPs 1 --numWorkers 2
    
  5. 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
    
  6. Configure parameters and save the model to Cloud Storage:

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

Check the permissions for your training component

You need to ensure that your Python code has the required permissions to read/write to your Cloud Storage bucket. Kubeflow solves this by creating a service account within your project as a part of the deployment. You can verify this by listing your service accounts:

gcloud iam service-accounts list | grep ${DEPLOYMENT_NAME}

Kubeflow granted this service account the right permissions to read and write to your storage bucket. Kubeflow also added a Kubernetes secret named user-gcp-sa to your cluster, containing the credentials needed to authenticate as this service account within the cluster:

kubectl describe secret user-gcp-sa

To access your storage bucket from inside the train container, you must set the GOOGLE_APPLICATION_CREDENTIALS environment variable to point to the JSON file contained in the secret. Set the variable by passing the following parameters:

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
kustomize edit add configmap mnist-map-training --from-literal=GOOGLE_APPLICATION_CREDENTIALS=/var/secrets/user-gcp-sa.json

Train the model on GKE

Now you are ready to run the TensorFlow training job on your cluster on GKE.

Apply the container to the cluster:

kustomize build . |kubectl apply -f -

When the command finishes running, there should be a new workload on the cluster, with a name like train-<VERSION_TAG>-chief-0.

You can see the workloads on the GKE Workloads page on the GCP console. To see the logs, click the train--chief-0 workload, then click Container logs.

View your trained model on Cloud Storage

When training is complete, you should see the model data pushed into your Cloud Storage bucket, tagged with the same version number as the container that generated it. To explore, click your bucket name on the Cloud Storage page on the GCP Console.

The output from the training application includes the following:

  • A set of checkpoints that you can use to resume training from a given point later.
  • An export directory that holds the trained model in a format that the TensorFlow Serving component can read. Read on to see how to serve your model for prediction using TensorFlow Serving.

Serve the trained model

Now you can put your trained model on a server and send it prediction requests.

  1. Enter the serving/GCS directory:

    cd $WORKING_DIR/serving/GCS
    
  2. Set a different name for the tf-serving:

    kustomize edit add configmap mnist-map-serving   --from-literal=name=mnist-gcs-dist
    
  3. Set your model path:

    kustomize edit add configmap mnist-map-serving   --from-literal=modelBasePath=${EXPORT_DIR} 
    
  4. Deploy it, and run a service to make the deployment accessible to other pods in the cluster:

    kustomize build . |kubectl apply -f -
    
  5. You can check the deployment by running:

    kubectl describe deployments mnist-gcs-dist
    
  6. The service should make the mnist-gcs-dist deployment accessible over port 9000:

    kubectl describe service mnist-gcs-dist
    

    You can see the mnist-gcs-dist service on the GKE Services page. If you click through to see the service details, you can see that it listens for connections within the cluster on port 9000.

Send online prediction requests to your model

Now you can deploy the final piece of your system: a web interface that can interact with a trained model server.

Deploy the sample web UI

When you downloaded the project files at the start of the tutorial, you downloaded the code for a simple web UI. The code is stored in the ${WORKING_DIR}/web-ui directory.

The web UI uses a Flask server to host the HTML, CSS, and JavaScript files for the web page. The Python program, mnist_client.py, contains a function that interacts directly with the TensorFlow model server.

The ${WORKING_DIR}/web-ui directory also contains a Dockerfile to build the application into a container image.

(Optional) Build an image and push it to Container Registry

Follow these steps to build an image from your code:

  1. Move back to the project directory:

    cd ${WORKING_DIR}
    
  2. Set the path in Container Registry to push the container image to:

    export UI_IMG_PATH=gcr.io/${PROJECT}/${DEPLOYMENT_NAME}-web-ui
    
  3. Build the Docker image for the web-ui directory:

    docker build ${WORKING_DIR}/web-ui -t ${UI_IMG_PATH}
    
  4. Allow Docker access to your Container Registry:

    gcloud auth configure-docker --quiet
    
  5. Push the container to Container Registry:

    docker push ${UI_IMG_PATH}
    

    The push may take a few minutes to complete. You should see Docker progress updates in your command window.

  6. Wait until the process is complete, then you should see your new container image listed on the Container Registry page on the GCP console.

Deploy the web UI to the cluster

Follow these steps to deploy the web UI to your Kubeflow cluster:

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

  1. Enter the front directory:

    cd front
    
  2. Optionally update the image to the ${UI_IMG_PATH} in the deployment.yaml.

  3. To deploy the web front end to your cluster:

    kustomize build . |kubectl apply -f -
    

    Now there should be a new web UI running in the cluster. You can see the web-ui entry on the GKE Workloads page and on the Services page.

Access the web UI in your browser

Follow these steps to access the web UI in your web browser. It may take a few minutes for the IP address to be available:

  1. Find the IP address assigned to the service:

    kubectl get service web-ui
    
  2. Copy the value shown under EXTERNAL-IP and paste it into your web browser’s address bar. The web UI should appear.

  3. The web UI offers three fields to connect to the prediction server: Connection UI

  4. By default, the fields on the above web page are pre-filled with the details of the TensorFlow server that’s running in the cluster: a name, an address, and a port. You can change them if you used different values:

    • Model Name: mnist - The name that you gave to your serving component.

    • Server Address: mnist-service - You can enter the server address as a domain name or an IP address. Note that this is an internal IP address for the mnist-service service within your cluster, not a public address. Kubernetes provides an internal DNS service, so you can write the name of the service in the address field. Kubernetes routes all requests to the required IP address automatically.

    • Port: 9000 - The server listens on port 9000 by default.

  5. Click Connect. The system finds the server in your cluster and displays the classification results.

As an alternative to the external IP address, you can run the following command to access the UI via kubectl port-forward:

kubectl port-forward -n kubeflow `kubectl get pods --all-namespaces \
    --selector=app=web-ui -o=jsonpath='{.items[0].metadata.name}'` 8080:5000

If you use the port-forwarding option, you can open the UI in your web browser at localhost:8080.

The final product

Below the connect screen, you should see a prediction UI for your MNIST model. Prediction UI

Each time you refresh the page, it loads a random image from the MNIST test dataset and performs a prediction. In the above screenshot, the image shows a hand-written 7. The table below the image shows a bar graph for each classification label from 0 to 9. Each bar represents the probability that the image matches the respective label.

Clean up your GCP environment

Run the following command to delete your deployment and related resources:

gcloud deployment-manager --project=${PROJECT} deployments delete ${DEPLOYMENT_NAME}

Delete your Cloud Storage bucket when you’ve finished with it:

gsutil rm -r gs://${BUCKET_NAME}

Delete the container images uploaded to Container Registry:

// Find the digest id for each container image:
gcloud container images list-tags us.gcr.io/$PROJECT/kubeflow-train
gcloud container images list-tags us.gcr.io/$PROJECT/kubeflow-web-ui

// Delete each image:
gcloud container images delete us.gcr.io/$PROJECT/kubeflow-web-ui:$DIGEST_ID
gcloud container images delete us.gcr.io/$PROJECT/kubeflow-train:$DIGEST_ID

As an alternative to the command line, you can delete the various resources using the GCP Console.