Jérôme Dockès

> Something else that should be fixed in the same PR is renaming the attributes of DropCols and SelectCols so they end with _, see https://github.com/scikit-learn/scikit-learn/issues/32910 private attribute names don't...

Maybe it could be reworded a bit but we already have a dedicated error message for this: ``` >>> import skrub >>> iter(skrub.var('a')) Traceback (most recent call last): ... TypeError:...

indeed taking the max of scores biases the measured accuracy. @tbng do you want to take a look? regarding the further comments, the point of the Decoder is that it...

"Ben J. Smith" ***@***.***> writes: > Yes I considered forcing it not to do cross-validation. The only option that seemed viable was to use `model_selection.PredefinedSplit(test_fold)`. you could also pass an...

> The Decoder object should be able to take a pipeline and run nested cross-val on the Pipeline's params ? I don't think so because the Decoder averages the coefficients...

> But then we should reintroduce somewhere in the examples a case in which we perform nested cross-val for parameter selection in a pipeline. This is an important use case...

as @ymzayek says the `__init__.py` files are there because the tests are part of the package, and get installed along with the rest of nilearn. the reason that is the...

with the old setuptools configuration (0.10.0 release), if the `__init__.py` files are missing the test packages do not get copied to the installation directory when installing nilearn, so for example...

as a result, I believe [the azure pipelines](https://github.com/nilearn/nilearn/blob/54f4ccda94c6b9c4e4389f0c48804738904f6218/azure-pipelines.yml#L42) tests only succeed by chance because (i) we run them from the root of the repository, which contains the `nilearn` source (a...

one (small) advantage of having the tests as part of the package we forgot to mention in the meeting is that users can run the tests after installing the package...