Result reproducibility issues depending on Python version about sccoda HOT 3 CLOSED

Hi! Thanks for the detailed description of your problem!

The inconsistency between the summary and the output from credible_effects was due to a bug in the credible_effects function. I could fortunately fix that on the master branch in github already.

As for the difference between the Python versions, there seem to be multiple issues here. The current version of scCODA was developed in Python 3.8, so I recommend to use that version for now. The next update of scCODA will be up-to-date again with the current versions of most dependencies.
For that, I'd like to understand the problems you were facing in your Python 3.9 environment a little better. It would be incredible if you could tell me

• The package versions of scCODA's dependencies that you have installed (especially tensorflow, tensorflow-probability, numpy, pandas, sccoda)
• The full error log you get when trying to run mod.CompositionalAnalysis()

Thanks in advance!

from sccoda.

Thanks for the quick answer !
Indeed, installing with git has solved the problem for Python3.8

Here are the logs for Python 3.9:

And the list of the dependencies installed with conda for Python3.9:
Dependencies_conda_Py3.9.csv

Hope that helps you. You can reach me if you need anything else.

from sccoda.

I just released a new update of scCODA (0.1.4) that now supports the latest versions of tensorflow (you had tensorflow 2.6 installed in your 3.9 environment). Running it in Python 3.9 should also be possible.
This hopefully explains the long duration for model inference, as those new versions of tensorflow defaulted to eager execution (very slow), whereas older versions did not.

Also, you had an old version of scCODA (0.1.1) installed in the dependency list that you sent me. Neither rel_abundance_dispersion_plot nor automatic reference selection were implemented in that version. The latest version has both these features.

I hope this solves all issues!

from sccoda.