Jérôme Dockès

from office hours discussion we may simply change the version check to make it handle the case where one version has incorrect type and say that they don't match in...

> Besides, `bg_map` and `bg_on_data` didn't feel clear to me the first time I used them (why set `bg_on_data=True` if I already set a `bg_map`?), so maybe there is room...

> @jeromedockes I don't see how we handle the case with two meshes/textures for two hemispheres ? we don't. what I meant was that the issue discussed during the sprint...

It's perfectly fine and quite common to have pandas column names that are not valid identifiers and it's better to always index dataframes with `[]` (or `loc`, `iloc`, `at`) anyway....

IMO the best is to do what you suggest in your first comment -- keep the original name, ie have a column name in the design matrix that is not...

In all cases the background map is shown in the thresholded regions. `bg_on_data` also shows it in the non-thresholded parts by mixing it with the actual statistical map data. so...

> Also, it seems to me that `engine="plotly"` in `plot_surf()` doesn't handle well `bg_on_data=True` as illustrated in images generated in #3171. > I'll investigate this later. I think so too....

we probably don't want to add a dependency, especially as templateflow itself requires pybids and tqdm IIUC, which are not dependencies of nilearn. why couldn't a user use tensorflow directly...

> Would be great to know if others are also OK with that plan of action so we can lay out some clear next steps ! Yes your plan sounds...

> I was unsure whether it would be worth it, given that the plotly engine is still experimental (if I understand correctly). > > If you think it is worth...