This is interesting. If I understand right, the idea is that once scipy/numpy/matplotlib make it so importing sub packages happens automagically upon attr lookup, we will remove the commented import lines.

I like the idea that a single import of scipy allows you to be able to use e.g. scipy.sparse without having to import that sub package explicitly.

Could we do the same thing? Instead of having to import approximations separately, let's try to make it so nx.approximations loads the sub package. If we work the kinks out for that, we might learn enough to try to make it work for scipy, numpy, etc.

But that's another PR.
This looks good to me. I haven't searched to make sure all places have been changed. I guess that you have done that.

I like the idea that a single import of scipy allows you to be able to use e.g. scipy.sparse without having to import that sub package explicitly.

I definitely agree. I haven't looked too closely at the __getattr__ lazy importing, but it would be great to have that pattern. I also agree with @dschult that it's an interesting thing to investigate for networkx as well.

@stefanv is looking into lazy imports and said he thinks it should be doable.

This is for another PR as well, but I noticed some of our examples were doing this

from networkx import nx

instead of this

import networkx as nx

I fixed them in #4403. But we should clean up our __init__s and __all__s soon to prevent this from being possible (among other things).

I have a working prototype.

https://gist.github.com/stefanv/fde776cef7464b4e2cc0721bd37d51f4

Using this, I had to modify scikit-image's main __init__.py to include the following:

from .lazy import require

def __getattr__(name):
    return require(f'skimage.{name}')

Then, when you do:

import skimage as ski
ski.filters.window(...)

it lazy loads, and works!

The only problem I've noticed is that it makes IPython's auto-completer very angry, but I suspect we can work with that team to address it.

@rossbar I removed the out of scope changes to the tests (cb08f7d) so it should be more straightforward to review now.

@dschult @rossbar I think this is ready to merge.