Are there any caveats to take into consideration before installing AutoML's SMAC3 using pip inside a regular venv rather than within a conda environment, as recommended by the docs?

The docs suggest using conda for environment management, but in the initial 'Requirements' section they suggest installing swig, the only system dependency that can't be handled by pip, using the OS's package manager. Moreover, the first option literally stated in the docs is to "install SMAC via PyPI" using pip.

I'm not very experienced with the anaconda distribution. Are there any special versions or builds of some packages that smac depends on? I understand that it comes with its own python interpreter implementation or build, too. I'm assuming it respects the CPython standards, but maybe there are some key differences that SMAC3 could depend on?

LE: I fot to mention that I ran the test battery of SMAC both in a conda env and in a venv, and the same few tests fail.

LE2: apparently, there's an option even for swig to be installed via pip as a PyPI package

Are there any caveats to take into consideration before installing AutoML's SMAC3 using pip inside a regular venv rather than within a conda environment, as recommended by the docs?

The docs suggest using conda for environment management, but in the initial 'Requirements' section they suggest installing swig, the only system dependency that can't be handled by pip, using the OS's package manager. Moreover, the first option literally stated in the docs is to "install SMAC via PyPI" using pip.

I'm not very experienced with the anaconda distribution. Are there any special versions or builds of some packages that smac depends on? I understand that it comes with its own python interpreter implementation or build, too. I'm assuming it respects the CPython standards, but maybe there are some key differences that SMAC3 could depend on?

LE: I fot to mention that I ran the test battery of SMAC both in a conda env and in a venv, and the same few tests fail.

LE2: apparently, there's an option even for swig to be installed via pip as a PyPI package

• python
• pip
• anaconda
• conda
• pypi

1 Answer 1

As per this comment by a member of smac's dev team on a github issue on the subject, indeed there is a dependency that is easiest to compile with anaconda to provide cross-compatibility.

However, it's also worth noting that on linux there are prebuilt binaries for that dep, so conda is not actually required, and one can safely install using pip.