It breaks on old matplotlib though.

Yes it does break on really old matplotlib, but not on the matplotlib LTS.
According to our installation doc, we don't even mention a minimum matplotlib version requirement.

So I guess the right way would be to raise the minimum requirements on our python 2 CI, and then merge this one. If we do that, maybe we could also raise the min versions for other dependencies such as skimage?

According to our installation doc, we don't even mention a minimum matplotlib version requirement.

See README: for running the examples Matplotlib >= 1.4 is required.

I think we need to fix this one before deprecating python2, because normed will be deprecated in matplotlib 3.1. At that time, 0.20.X is likely to be the latest stable version.

So I guess the right way would be to raise the minimum requirements on our python 2 CI

Apart from raising the minimal version, we have other solutions, see e.g., https://github.com/scikit-learn/scikit-learn/blob/master/examples/decomposition/plot_sparse_coding.py (introduce different solutions for different versions).

@qinhanmin2014 I see. So our website doesn't mention a min dependency, but our README (which is mirrored on PyPi) does. Hmm, do we have a strong attachment to those requirements? I'd say since these changes don't change anything regarding the usage of our API, and they're only examples, we can just change the README to move the min requirement a bit upward.

Otherwise I can make the examples support both matplotlib APIs.

Also, the README says matplotlib >= 1.4, and I guess moving to 2.1.2 or 2.2.3 would be reasonable since matplotlib lists:

• 3.0.2 Stable version.
• 2.2.3 LTS LTS version.
• 3.x Latest git master (unstable)
• 2.1.2 Older stable version.

That said, Ubuntu 14.04 has matplotlib 1.3.1, and 16.04 matplotlib 1.5.1, I guess.

I think the min dependency comes from Ubuntu 14.04. We'll need more consensus to change the min dependency (not so easy I think).

@adrinjalali Do you have time to investigate when (which version) did matplotlib introduce the density parameter?

I think, for 0.21 we can use 16.04 LTS as the oldest "supported" Ubuntu. That's what we were planning to do for numpy and scipy in #12184

So it might make sense to do the same for matplotlib and scikit-image.

I think, for 0.21 we can use 16.04 LTS as the oldest "supported" Ubuntu. That's what we were planning to do for numpy and scipy in #12184

@rth But I'd rather hurry this one into 0.20.2, see my comment above:
I think we need to fix this one before deprecating python2, because normed will be deprecated in matplotlib 3.1. At that time, 0.20.X is likely to be the latest stable version.

@adrinjalali Do you have time to investigate when (which version) did matplotlib introduce the density parameter?

Sorry for the stupid mistake. density parameter is introduced in matplotlib 2.1, which is indicated in the deprecation message.

Hmm, @rth the minimal version of matplotlib for Ubuntu 16.04 is 1.5.1? Do you have time to check it (and provide the URL)?

But I'd rather hurry this one into 0.20.2

Sure

the minimal version of matplotlib for Ubuntu 16.04 is 1.5.1?

it's 1.5.1 for matplotlib (https://packages.ubuntu.com/xenial/python3-matplotlib) and 0.10.1 for scikit-image (https://packages.ubuntu.com/xenial/python3-skimage).

Thanks @rth, so seems that we need version comparison here even if we don't want to include the change in 0.20.2. Do you want to complete it? @adrinjalali

Sure, I'll fix the examples here.