Nilearn: Attribute `components_` is deprecated since 0.4.1 but still here

[NicolasGensollen]

As can be seen in these files:

https://github.com/nilearn/nilearn/blob/dbbd221e3633c97464034c5366faa548728c9999/nilearn/decomposition/canica.py#L130-L134

https://github.com/nilearn/nilearn/blob/dbbd221e3633c97464034c5366faa548728c9999/nilearn/decomposition/dict_learning.py#L152-L155

https://github.com/nilearn/nilearn/blob/a36c0442fc8eebe401b310dd9799c0f071080e90/nilearn/decomposition/multi_pca.py#L121-L125

There is also a note here:

https://github.com/nilearn/nilearn/blob/a36c0442fc8eebe401b310dd9799c0f071080e90/examples/03_connectivity/plot_extract_regions_dictlearning_maps.py#L20-L26

We should remove them in 0.8.0.

[NicolasGensollen]

Actually, I'm not really sure what was intended here. I initially thought that components_ was meant to disapear in favor of components_img_, but I realize it could just be a note to encourage users to rely on components_img_ instead of manually unmasking components_. Does someone knows more on this?

[jeromedockes]

it seems this was done in pr #1536 , and here is the discussion about deprecating components_

[jeromedockes]

also, the transform and inverse_transform of the base decomposition:
https://github.com/nilearn/nilearn/blob/dbbd221e3633c97464034c5366faa548728c9999/nilearn/decomposition/base.py#L458

unmask the components every time; they should rely on self.components_img_; they also create a new mapsmasker every time; this should be done in fit

[jeromedockes]

note the Decoder & co., for example, have both coef_ and coef_img_. SearchLight has scores_ which is neither an image nor masked data (see #2673), ... maybe we should make a list of all such objects and make this consistent?

I don't see a big disadvantage of having them both, as the Decoder does. to avoid storing them twice the least frequently used one can be implemented as a property