Extension of howto unit tests: PyPy conformity - Githubissues

fhswf / MLPro

MLPro - The Integrative Middleware Framework for Standardized Machine Learning in Python

https://mlpro.readthedocs.io/

Apache License 2.0

12 stars 3 forks source link

Extension of howto unit tests: PyPy conformity #433

Open detlefarend opened 2 years ago

detlefarend commented 2 years ago

Description/Motivation This issue shall ensure MLPro's compatibility to the faster Python derivate PyPy.

Task list

[ ] 1. All howto files: reduction of runtime in unit test mode to an absolute necessary minimum
[ ] 2. Extension of unit test by a second run under PyPy
[ ] 3. Logging of speed gain by using PyPy
[ ] 4. RTD documentation, introduction: hint for PyPy conformity (and speed gain)

Cross references https://www.pypy.org/ https://github.com/fhswf/MLPro/tree/rizkydiprasetya/issue433/.github/workflows

rizkydiprasetya commented 1 year ago

Pytorch cannot be installed in pypy environment.

detlefarend commented 1 year ago

What code is affected except the adaptive function?

rizkydiprasetya commented 1 year ago

SB3 wrapper

detlefarend commented 1 year ago

Could you exclude both from the unit test just to see what's with the rest?

rizkydiprasetya commented 1 year ago

The results are the following without pytorch:

Pypy 3.8 : Success
Pypy 3.9 : Failed due to Scipy Package (I have to check this locally, because the installation of Scipy was succeed)

Note: Some of the packages are not available as binary, so those packages need to be compiled and it take times. This could be solved with pre-compiled docker container. The run can be checked here: https://github.com/fhswf/MLPro/actions/runs/3483988002

detlefarend commented 1 year ago

I've found some articles that suggest using anaconda packages in combination with pypy might be beneficial: https://www.pypy.org/compat.html https://www.pypy.org/posts/2022/11/pypy-and-conda-forge.html

rizkydiprasetya commented 1 year ago

Then, it is in conda environment. Maybe we can leave some note for the user, for pypy, they need to use conda.

rizkydiprasetya commented 1 year ago

with conda-forge, there is another problem with the scikit-learn package. It is successfully installed, but the error says, it is not installed properly by conda. And also another problem with pygame in rl_wp_003 example.

Here is the runner: https://github.com/fhswf/MLPro/actions/runs/3487389868

rizkydiprasetya commented 1 year ago

close this due to dependencies problem in pypy.

detlefarend commented 1 year ago

As discussed in our today's meeting: let' recap this topic and setup a new option for pip to install mlpro with all 3rd-party packages that are compatible to pypy....