[MAINT] Updated pinned versions of dependencies in setup.cfg

[chiuhoward]

Limited manual test coverage in conda environment with python == 3.11. numpy2 breaks AFQ-Insight.
Closes #11.

Changes:

• Updated pinned versions for scipy, dipy, matplotlib, numpy, pandas, seaborn, statsmodels, and scikit-learn.
• Changed imports for deprecated sklearn functions in serial_bagging.
• Change contributing.md references for repo URL from richford to tractometry

[chiuhoward]

If I’m understanding the failing tests correctly, it has to do with base_estimator being deprecated since 1.4.0? Maybe I could pin it to 1.3.2 as a short-term fix until the refactoring comes in. I don’t know enough to know that we can just replace all instances of base_estimator with estimator, maybe the parameters it takes also change...

scikit-learn/scikit-learn#23819

[arokem]

@chiuhoward : I see a few different errors here:

• TypeError: OneHotEncoder.__init__() got an unexpected keyword argument 'sparse'

• AttributeError: module 'numpy' has no attribute 'object'.

• TypeError: _fit_and_score() missing 1 required keyword-only argument: 'score_params'

• TypeError: BaggingClassifier.__init__() got an unexpected keyword argument 'base_estimator'

• TypeError: BaggingRegressor.__init__() got an unexpected keyword argument 'base_estimator'

• TypeError: AdaBoostClassifier.__init__() got an unexpected keyword argument 'base_estimator'

I think that the three latter ones are essentially the same issue that you identified. But there are also the three others to contend with. I am going to try to fetch your branch, and see whether I can diagnose some of these.

For future reference, it's easier to work with PRs (and will also make your life easier) if you make the PR from a separate branch, and not from your main branch.

[chiuhoward]

I've tried reading the documentation:

1. I think OneHotEncoder needs sparse_output instead of sparse, but from the test log it seems like it's being imported by neurocombat rather than AFQ-Insight? I can't find any mention of OneHotEncoder in the AFQ-Insight repo (https://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.OneHotEncoder.html)

2. I think def _fit_and_score_ckpt is where we introduce score_params as an additional kwarg, but I don't know how/where to concretely implement this. https://github.com/scikit-learn/scikit-learn/blob/6e9039160f0dfc3153643143af4cfdca941d2045/sklearn/model_selection/_validation.py#L740

[arokem]

We do depend on neurocombat_sklearn, which has not been updated in a while...

I wonder whether this is something that the nilearn community may be interested in taking up and maintaining in the long run.

[arokem]

Maybe for now our best option is to vendorize neurocombat_sklearn into our code-base and fix it up here. It's MIT-licensed and less than 400 lines of code, with only sklearn and numpy dependencies, so shouldn't be too bad to bring over.

[chiuhoward]

closed by #20