Gael Varoquaux

> I think that we could save it in the metadata of memory object. I mean of the memory entry. What is currently stored in the .json file. It's a...

Did you try using lzma? I seem to remember that it's a pretty fast compression algorithm.

> It depends on the data but generally it compress better with the cost of slower speed. I've already made this mistake: I keep confusing it with another one (lz4,...

> One potential compromise would be to continue using get_data (now coped to here) when the user does not specify a float type, but when a dtype=float is specified, use...

From a cursory look, it seems that @KnorrFG is right: the mean and variance should only be computed on the train set. I bet that some people will be unhappy...

Sounds reasonnable to me. Just to check that I understand things right: this happens when the peak extraction routine fails.

> There are also many cases to consider as well such as: threshold, > min_region_size, n_subjects and components. So, I found warning and empty > returns would be reasonable. But,...

What should be the behavior ?  I frown at adding an option. How about: if the FOV is smaller than 10 in all directions the display of coordinates has a...

> this is fairly easy to do by modifying the matplotlib figure I think that if @TheChymera sends a clean pull request implementing the suggesting functionality, we should take it.

Maybe we should add this to an example.