Installing AutoSklearn with conda

I really wanted to try out the examples of using AutoSklearn over at Machine Learning Mastery. But when I ran pip install auto-sklearn, an error occurred:

Cython.Compiler.Errors.CompileError: sklearn/ensemble/_hist_gradient_boosting/splitting.pyx

Looking into this issue suggests a possible incompatibility between an older version of scikit-learn (required by AutoSklearn) and newer Cython (currently Cython-3.0.11 in Slackware). At the time of writing, AutoSklearn was last updated on Sep 20, 2022 (see here). Therefore, making it work likely requires older versions of several packages, and perhaps even Python itself.

This is what finally pushed me to figure out how to install and set up a Python virtual environment. And not just any virtual environment, but conda, a virtual environment manager, which handles not only Python packages but Python itself! According to David Cañones at WhiteBox, “the conda dependency resolver is more robust” than pip. This sounds like it could make life a lot easier!

I recommend reading the whole guide in the last link. In a nutshell, here’s what I did:

• To install conda, download and run https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh.
• Run source ~/.bashrc (the guide above suggests opening a new terminal window) and confirmed that conda was initialized by seeing that my shell prompt started with (base).
• Run these commands:

conda create -n automl
conda activate automl
conda install conda-forge::auto-sklearn

That’s it!

Note that I did not run conda install python because that installs the latest Python, and I didn’t want to spend time figuring out what version of Python is compatible with AutoSklearn from 2022. That single conda install command automatically figured out that it needs python-3.9.20, among lots of other dependencies. I think that’s a good example of robust dependency resolution!

And yes, I ran the AutoSklearn example from Machine Learning Mastery and got some nice results.

>>> print(model.sprint_statistics())
auto-sklearn results:
  Dataset name: b1aa9d5c-a18f-11ef-a3d0-5c5f67894b53
  Metric: accuracy
  Best validation score: 0.956522
  Number of target algorithm runs: 85
  Number of successful target algorithm runs: 55
  Number of crashed target algorithm runs: 30
  Number of target algorithms that exceeded the time limit: 0
  Number of target algorithms that exceeded the memory limit: 0

Enjoy!