Skip to content

Latest commit

 

History

History
256 lines (183 loc) · 9.85 KB

mlx-workshop.md

File metadata and controls

256 lines (183 loc) · 9.85 KB
Raw

MLX Workshop

This workshop demonstrates how a data scientist can use MLX to analyze a dataset, train a model, and deploy it on Kubernetes using MLX.

We will make use of Project CodeNet, a set of approximately 14 million code samples, each of which is an intended solution to one of 4000 coding problems. For more information about Project CodeNet take a looks at the article Kickstarting AI for Code: Introducing IBM’s Project CodeNet

The specific dataset, notebook, and model we will use can be found here:

Table of Contents

Prerequisites

For this workshop you will need access to a Kubernetes cluster with MLX deployed, or a local deployment of MLX on KIND (Kubernetes in Docker).

There are only minor differences in how to navigate the Kubernetes resources depending on the selected deployment option, with the most notable difference being the host and port of the MLX (UI) server.

For this workshop we assume a single-user local MLX deployment on localhost port 3000 as described in the MLX on KIND deployment guide.

Login to the MLX Web UI

Open the MLX Web UI in your browser and login as the admin user:

http://localhost:3000/login

Once you are logged in, you should see the Settings menu item at the bottom of the left-hand navigation.

Prepare the Dataset to be Mounted

For this workshop we will use a small subset of the CodeNet dataset containing about 2200 random samples across 10 languages to train a language classification model.

Datasets Featured Page

Click on the Datasets menu item and select the Project CodeNet Language Classifier dataset.

  • The DESCRIPTION tab shows the details of the dataset
  • The DATASET YAML tab shows the metadata required by MLX, i.e. for creating a PVC
  • The LAUNCH tab allows users to create a Persistent Volume Claim (PVC) containing the dataset
  • The RELATED ASSETS tab shows other assets that make use of this dataset

Datasets Launch Screen

We will click on the LAUNCH tab to start a pipeline run which will download the dataset and create the PVC.

  • Enter a Run Name of your choice or leave the default value
  • Enter the Namespace for the PVC to be created in. The namespace will vary depending on your MLX deployment. For single-user deployments of MLX, like MLX on KIND, enter kubeflow (for multi-user deployments enter mlx)
  • Click SUBMIT to start the pipeline run

Datasets Run Graph

Now the Kubeflow Pipeline run view should appear, and get updated as the pipeline execution progresses. You can click on a task in the pipeline graph and follow the Logs.

Datasets Run Output

Once the pipeline is finished, click on the second task of the pipeline and select the Input/Output tab. Under the Output Parameters, dataset-status, find the name of the PVC that was created: "codenet-langclass", which should be the same as the ID of the dataset. Select and Copy it to the clipboard. We will need that in the next step.

Run the Notebook

We will use the Project CodeNet Language Classification notebook to train a Keras model that should be able to detect the language of a given snippet of source code.

Notebooks Featured Page

Click on the Notebooks menu item and select the Project CodeNet Language Classification notebook.

  • The DETAILS tab shows the details of the notebook
  • The LAUNCH tab allows users to run the notebook in a pipeline
  • The YAML DEFINITION tab shows the metadata required by MLX to run the notebook in a pipeline
  • The NOTEBOOK CODE tab displays the notebook source code

Notebooks Launch Screen

Click on the LAUNCH tab to start a pipeline run with this notebook

  • Enter a Run Name of your choice or leave the default value
  • Enter the Dataset PVC name from the Dataset run output: "codenet-langclass"
  • Enter a Mount Path to mount the PVC to the Kubernetes Pod that is running the notebook: "/tmp/data"
  • Click SUBMIT to start the pipeline run

Notebooks Run Graph

Now the Kubeflow Pipeline run view should appear, and get updated as the pipeline execution progresses. You can click on the task in the pipeline graph and follow the Logs.

Notebooks Run Logs

The pipeline run will generate an HTML version of the notebook containing the input and newly generated output cells of the notebook. Once the pipeline is finished, scroll to the end of the Logs and find the line that reads PUT /mlpipeline/notebooks/project-codenet-language-classification/Project_CodeNet_LangClass.html i.e.:

[D 02:30:16.327] http://10.96.180.205:9000 "PUT /mlpipeline/notebooks/project-codenet-language-classification/Project_CodeNet_LangClass.html HTTP/1.1" 200

This is where the notebook HTML was stored in the Minio object storage.

To download the notebook output we require access to Minio from our local browser. For MLX on KIND we can achieve that by forwarding the Minio service port:

kubectl port-forward -n kubeflow svc/minio-service 9000:9000

Now we should be able to download the notebook by pasting the URL in our browser's address bar:

http://localhost:9000/mlpipeline/notebooks/project-codenet-language-classification/Project_CodeNet_LangClass.html

Open the notebook in a browser tab and scroll close to the bottom to find out about the accuracy of our trained model and compare it to the accuracy achieved when the original notebook was created.

Notebooks Language Classification Test

Serve the Model

We have created a containerized version of the CodeNet Language Classification model that can be deployed on Kubernetes to allow inferencing via curl, or, directly from a web browser using a simple Swagger UI to the inferencing service.

Models Featured Page

Click on the Models menu item and select the CodeNet Language Classification model.

  • The DESCRIPTION tab shows the details of the model with links to the original dataset
  • The LAUNCH tab allows users to (train or) serve the model on Kubernetes
  • The YAML DEFINITION tab shows the metadata required by MLX
  • The SAMPLE SERVING CODE tab displays a generated sample pipeline to serve the model

Models Launch Tab

Click on the LAUNCH tab to start the pipeline run to serve the model

  • The Launch Type is Serving
  • The Platform is Kubernetes
  • Enter a Run Name or leave the default
  • Click SUBMIT to start the pipeline run

Models Run Output

Now the Kubeflow Pipeline run view should appear, and get updated as the pipeline execution progresses. You can click on the task in the pipeline graph and follow the Logs. Once the second step of the pipeline is completed, you can find the name of the deployed model "codenet-language-classification" and the URL where the deployment can be accessed.

For MLX on KIND we need to make the port available on localhost by running the following command:

kubectl port-forward svc/codenet-language-classification 5000:5000

Now we should be able to access the UI page for the deployed model by pasting localhost:5000 into our browser's address bar:

http://localhost:5000/

Expand the first section called model and click on POST /model/predict and the click on the Try it out button on the very right of the Parameters line.

Models Swagger UI

In order to test the classification model, we need to provide a source code file in any of the 10 supported languages like C/C++, Java, JavaScript or Python. You could either use a source file from one of your local projects or download the CodeNet Sample Dataset for Language Classification and extract it to a local folder.

On the model prediction input dialog, click Browse and select a source file

Open Source File

Now click Execute to get the model's prediction.

The Response should look similar to this:

Response body

{
  "status": "ok",
  "predictions": [
    {
      "language": "Haskell",
      "probability": 0.9999926090240479
    }
  ]
}

Conclusion

In this workshop we demonstrated how to use MLX to download a dataset into a Persistent Volume Claim on Kubernetes and how to mount that dataset to a pod running a notebook. We also showed how MLX can be used to deploy pretrained containerized models on Kubernetes.

For comments, suggestions or bug reports, please do to https://github.com/machine-learning-exchange/mlx/issues