rai-toolbox-template

Responsible AI Toolbox Project Template

Description

This template demonstrates how to use Microsoft's Responsible AI Toolbox suite of tools in a Domino workspace.

It covers:

• How to create a Domino compute environment for running the Responsible AI Toolbox.
• Sample code for computing data that can be ingested into the dashboard.
• Code snippets for creating a link to open the dashboard inside of a Domino workspace.

To learn more about the tools included in the Responsible AI Toolbox and how to use them to analyze models and dataset, you can reference these resources:

• Microsoft's responsibleaitoolbox.ai website.
• The responsible-ai-toolbox GitHub project. In particular:
  • Tour.
  • Getting Started.
  • Notebook examples.

Environment

Environment compatibility requirements

To use the Responsible AI Toolbox, the raiwidgets Python package must be installed.

In addition to that, depending on which tools you want to use, you may need to install additional packages. See the responsible-ai-toolbox documentation for details.

Please notice that the packages on the Responsible AI Toolbox may require older versions of Python packages such as numpy, pandas, scikit-learn, scipy, ipython and others. These are requirements of the Python packages that are part of responsible-ai-toolbox, and not requirements imposed by Domino. That means that the toolbox may not be compatible with compute environments that have recent versions of these packages. In order to run the toolbox, it will be necessary to install the responsible-ai-toolbox packages on a compute environment that has older, compatible versions of the required packages, or a minimal set of packages.

Environment creation instructions

To create a Domino compute environment that meets the requirements:

1. On the Environments page, click on Create Environment.
2. On the Name field enter the environment name (e.g. "raiwidgets").
3. On the Base Environment / Image group, select "Start from a custom base image" and enter this image URI on the FROM field:

quay.io/domino/compute-environment-images:ubuntu20-py3.9-domino5.9-minimal

4. On the Visibility field, choose between "Private" (the default) or "Globally Accessible".
5. Click the Customize before building button.
6. On the Dockerfile Instructions field, enter:

# Install the required package for displaying the dashboard
USER root
RUN pip install raiwidgets==0.33.0
# Install additional, optional packages as required
RUN pip install responsibleai-text==0.2.5
RUN pip install responsibleai-vision==0.3.5
# Set the default user
USER ubuntu

7. On the Pluggable Workspace Tools field, enter:

jupyterlab:
  title: "JupyterLab"
  iconUrl: "/assets/images/workspace-logos/jupyterlab.svg"
  start: [  "/opt/domino/workspaces/jupyterlab/start" ]
  httpProxy:
    internalPath: "/{{ownerUsername}}/{{projectName}}/{{sessionPathComponent}}/{{runId}}/{{#if pathToOpen}}tree/{{pathToOpen}}{{/if}}"
    port: 8888
    rewrite: false
    requireSubdomain: false

8. At the bottom of the page, click on Build.
9. Wait for the build status to become a green checkmark ("Succeeded"). Click on the Build Logs link in case of an error.

Workspace

To create a JupyterLab workspace to run the dashboard:

1. Open the project and navigate to its Workspaces page.
2. Click on Create New Workspace.
3. On the Workspace Environment field, select the compute environment created above (e.g. "raiwidgets").
4. Select the JupyterLab IDE.
5. Click the Launch button.

After the JupyterLab interface opens, browse the examples folder and run the provided notebook.

At the bottom of the notebook, after it finishes running, there will be a link to open the Responsible AI Dashboard.

Dashboard link

The Responsible AI Dashboard runs as a Flask app on the same container that is running the Domino workspace. It will listen on a port in the range 8704 to 8993.

In order to be able to access the dashboard, the appropriate link must be assembled. Jupyter and JupyterLab have the ability to reverse proxy requests by appending the /proxy/{port} path to the URL.

See examples/link-snippet.ipynb for a code snippet that can be added to a Jupyter or JupyterLab notebook to display a clickable link that will open the dashboard on a new browser tab, assuming that the dashboard Python variable contains the Dashboard object instance.

The link will open on the same Domino workspace session, and access control will be enforced so that only authorized users can view the dashboard.

See examples/responsibleaidashboard-housing-decision-making.ipynb for a full example.