xautoml 0.1.4

A JupyterLab extension to visualize and explain AutoML optimization procedures.

XAutoML: A Visual Analytics Tool for Establishing Trust in Automated Machine Learning

XAutoML is an interactive visual analytics tool for explaining AutoML optimisation procedures and ML pipelines constructed by AutoML. It combines interactive visualizations with established techniques from explainable AI (XAI) to make the complete AutoML procedure transparent and explainable. We integrate XAutoML with Jupyter to enable experienced users to extend the visual analytics with advanced ad-hoc visualizations based on information extracted from XAutoML.

Currently, XAutoML supports only

• auto-sklearn
• dswizard
• FLAML
• Optuna
• scikit-learn

but we plan to add support for further AutoML systems. You can find a video introducing XAutoML on YouTube.

Important: XAutoML requires JavaScript support. This implies that embedded Jupyter Notebooks in VS Code, PyCharm, ... may not work properly. Please use XAutoML in a browser for now.

Install

Create a new environment with python >= 3.7 and make sure swig is installed either on your system or inside the environment.

Install swig

• You can either install swig via conda (conda install swig)
• Or follow the official documentation to install it

To install the extension, execute:

pip install xautoml

Usage

XAutoML currently only works with JupyterLab. You can find ready to use Notebook examples in the examples folder.

cd examples
jupyter lab

To use XAutoML, three steps are necessary:

1. Perform an optimization in one of the supported AutoML frameworks
2. Import the RunHistory of the optimizer via the corresponding adapter
3. Create the XAutoML visualization

from xautoml.main import XAutoML
from xautoml.adapter import import_sklearn
from xautoml.util.datasets import openml_task
from sklearn.model_selection import RandomizedSearchCV

# 1) Perform AutoML optimization
random_search = RandomizedSearchCV(...).fit(...)

# 2) Use Adapter to create RunHistory
rh = import_sklearn(random_search)

# 3) Create Visualization
X_test, y_test = openml_task(31, 0, test=True)
main = XAutoML(rh, X_test, y_test)
main.explain()

Checking Only Specific Aspects

If you want to see only specific information in XAutoML, you can include only specific visualizations. For example, if you only want to validate the model behaviour, you can exclude all technical ML details by using

main.explain(include={'overview', 'candidate:domain', 'ensemble'})

Standalone Visualizations

Instead of rendering the complete visualization, only specific standalone views can be rendered. For example, if you are only interested in checking the leaderboard, only it can be rendered using

main.explain_leaderboard()

For basically all views a stand-alone view is available. Check the documentation of the various XAutoML().render_*(...) methods for more details.

Docker

XAutoML is also provided as a Docker container. The container is based on the popular Jupyter Docker Stack scipy-notebook. It contains a ready to use Jupyter Lab with the installed XAutoML extension. In addition, the /home/jovyan/automl contains ready to use XAutoML examples for all supported AutoML libraries. To use the docker container, simply run

docker run -it -p 8888:8888 mz93/xautoml

Uninstall

To remove the extension, execute:

pip uninstall xautoml

Troubleshoot

If you are seeing the frontend extension, but it is not working, check that the server extension is enabled:

jupyter server extension list

If the server extension is installed and enabled, but you are not seeing the frontend extension, check the frontend extension is installed:

jupyter labextension list

If the installation failed with the following exception

[...]
    Running setup.py install for pyrfr ... error
    ERROR: Command errored out with exit status 1:
    [...]
    swig.exe -python -c++ -modern -py3 -features nondynamic -I./include -o pyrfr/regression_wrap.cpp pyrfr/regression.i
    error: command 'swig.exe' failed: No such file or directory
[...]

verify that you have swig installed (see Installation above).

Contributing

Development install

Note: You will need NodeJS to build the extension package.

The jlpm command is JupyterLab's pinned version of yarn that is installed with JupyterLab. You may use yarn or npm in lieu of jlpm below.

# Clone the repo to your local environment
# Change directory to the xautoml directory
# Install package in development mode
pip install -e .
# Link your development version of the extension with JupyterLab
jupyter labextension develop . --overwrite
# Server extension must be manually installed in develop mode
jupyter server extension enable xautoml
# Rebuild extension Typescript source after making changes
jlpm install
jlpm run build

You can watch the source directory and run JupyterLab at the same time in different terminals to watch for changes in the extension's source and automatically rebuild the extension.

# Watch the source directory in one terminal, automatically rebuilding when needed
jlpm run watch
# Run JupyterLab in another terminal
jupyter lab

With the watch command running, every saved change will immediately be built locally and available in your running JupyterLab. Refresh JupyterLab to load the change in your browser (you may need to wait several seconds for the extension to be rebuilt).

By default, the jlpm run build command generates the source maps for this extension to make it easier to debug using the browser dev tools. To also generate source maps for the JupyterLab core extensions, you can run the following command:

jupyter lab build --minimize=False

Development uninstall

# Server extension must be manually disabled in develop mode
jupyter server extension disable xautoml
pip uninstall xautoml

In development mode, you will also need to remove the symlink created by jupyter labextension develop command. To find its location, you can run jupyter labextension list to figure out where the labextensions folder is located. Then you can remove the symlink named xautoml within that folder.

Release new version

Increase version number in package.json and upload the latest build to pypi.

pip install build
python -m build -s
python -m twine upload dist/*

Citation

If you are using XAutoML, please cite it as

@article{Zoller2022,
    author = "Z{\"{o}}ller, Marc-Andr{\'{e}} and Titov, Waldemar and Schlegel, Thomas and Huber, Marco F.",
    title = "{XAutoML: A Visual Analytics Tool for Establishing Trust in Automated Machine Learning}",
    journal = "arXiv preprint arXiv: 2202.11954",
    volume = "1",
    year = "2022",
    pages = "1-34",
    url = "http://arxiv.org/abs/2202.11954",
    eprint = "2202.11954",
    archivePrefix = "arXiv",
    arxivId = "2202.11954"
}