Deploy and use a Deep Learning (DL) container

A Deep Learning (DL) container is a Docker container with preinstalled data science frameworks, libraries, and tools. A user, such as a data scientist, chooses a single DL container and deploys it. The deployed container has a performance-optimized, consistent environment that helps with quickly prototyping and implementing workflows.

Deploy a DL container

Before using a DL container, you must choose and deploy a container image that runs your Machine Learning (ML) task. Preinstalled frameworks, libraries, and tools provision a DL container.

Deploy a DL container using `kubeconfig` files

Google Distributed Cloud (GDC) air-gapped appliance provides the following kubeconfig files that deploy your DL container:

ADMIN_CLUSTER_KUBECONFIG: This file connects to an admin cluster. Each organization has an admin cluster. The admin cluster hosts some main components, such as the GDC console and the Harbor registry. Each organization has multiple user clusters. The admin cluster controls all of the user clusters within an organization. The ADMIN_CLUSTER_KUBECONFIG_PATH is the file path to the admin cluster kubeconfig file. For example, the path might be ~/home/org-admin-kubeconfig.
USER_CLUSTER_KUBECONFIG: This file connects to a user cluster. You choose a DL container to deploy to this cluster. The USER_CLUSTER_KUBECONFIG_PATH is the file path to the user cluster kubeconfig file. For example, the path might be ~/home/user-vm-1-kubeconfig.

For more information about signing into the UI and the kubectl tool, see Sign in. In order to retrieve the ADMIN_CLUSTER_KUBECONFIG and USER_CLUSTER_KUBECONFIG files, see Get a kubeconfig file.

Download the sample Machine Learning (ML) script and dataset

Download the sample ML script, beginner.ipynb, and dataset, mnist.npz, to run the ML quickstart tutorial. The tutorial demonstrates how to deploy and use a DL container to run ML experiments.

mkdir -p /tmp/datasets
cd /tmp/datasets

wget --no-check-certificate
https://GDCH_Appliance_URL/.well-known/static/dl-container-tutorial/beginner.ipynb

wget --no-check-certificate
https://GDCH_Appliance_URL/.well-known/static/dl-container-tutorial/mnist.npz

Replace GDCH_Appliance_URL with the domain name used to access GDC. When opening any URL for the first time, GDC redirects you to your identity provider login page.

Look up the IP address of the Harbor registry

Before using the sample script and dataset, you must find the DL container image location in the Harbor registry. The Harbor registry is a service that stores private container images.

The first line of sample code sets the path of the KUBECONFIG environment variable, which gets the Harbor registry address. The Harbor registry address provides access to a list of available container images.

In the second line of sample code, the kubectl tool uses the KUBECONFIG environment variable.

In the third line of sample code, the REGISTRY_URL#https:// command removes the prefix https:// from the URL and stores the Harbor registry domain in the REGISTRY_IP environment variable.

In the last line of the sample code, the kubectl tool fetches the password for the admin user.

export KUBECONFIG=ADMIN_CLUSTER_KUBECONFIG_PATH

REGISTRY_URL=$(kubectl get harborcluster harbor -n harbor-system -o=jsonpath='{.spec.externalURL}')
REGISTRY_IP=${REGISTRY_URL#https://}
ADMIN_PASS=$(kubectl -n harbor-system get secret harbor-admin -o jsonpath="{.data.secret}" | base64 -d)

Choose a container image

You must choose a container image to deploy before you can run an ML task. Use the Harbor registry domain and the path in the following table from the Harbor registry to view the list of available container images:

Framework	Processor	Container Image Name
Base	GPU	base-cu113
Base	CPU	base-cpu
TensorFlow Enterprise 2.x	GPU	tf2-gpu
PyTorch	GPU	pytorch-gpu

This table is organized by framework and processor. To choose a DL container image that can process your ML experiment, follow these steps:

Identify the framework, which contains the ML tools.
Choose the processor. You choose the processor based on the kind of ML task to run and the compute intensity of that task. For example, choose one of the GPU processors when you have a compute-intensive ML task, and allocate a GPU resource to the DL container.

Create and deploy a DL container to the user cluster

To create the GDC instance, specify the path to the kubeconfig file of the user cluster. The KUBECONFIG environment variable specifies to which cluster the kubectl tool deploys the DL container. The kubectl apply command deploys the DL container instance.

Replace CONTAINER_IMAGE_NAME with the image selected from the list of images in Choose a container image, and be sure to supply the tag.

export KUBECONFIG=USER_CLUSTER_KUBECONFIG_PATH

kubectl apply -f - <<EOF
apiVersion: v1
kind: Pod
metadata:
  name: dl-container-pod
  namespace: NAMESPACE
spec:
  containers:
  - image: gcr.io/private-cloud-staging/notebooks/deeplearning-platform-release/CONTAINER_IMAGE_NAME:CONTAINER_IMAGE_TAG
    command: ["tail", "-f", "/dev/null"]
    name: training
EOF

Use a deployed DL container

The following topics provide an example of how to use a DL container image to train and use a model to generate predictions.

Copy the tutorial files to the DL container pod

Copy the quickstart tutorial files into your DL container pod. The beginner.ipynb contains the steps to train and use a model to make predictions. The ML training tutorial uses the mnist.npz data set file to train a model.

cd /tmp/datasets

kubectl cp beginner.ipynb dl-container-pod:/tmp
kubectl cp mnist.npz dl-container-pod:/tmp

Run the ML quickstart tutorial

Run the tutorial with the following commands. Use the first line to enter the container pod. After you are in the container pod, change directory to tmp, and run the papermill tool that is packaged in the DL container. The papermill tool runs the tutorial to create a notebook that generates predictions.

Enter an interactive terminal into the DL pod:

kubectl exec -it dl-container-pod -- /bin/bash

In the DL pod context, run the following commands:
```
cd tmp
papermill beginner.ipynb result.ipynb
```
The deployment generates a result.ipynb file in the /tmp directory.
View the content and the prediction outputs from the generated ML model:
```
cat result.ipynb
```

Optional: Delete the DL container pod

After you finish running your experiment in the DL container pod, delete the pod as a best practice:

kubectl delete pod dl-container-pod