On Fri, Feb 3, 2012 at 7:40 AM, Tony Yu
I'd like to start a discussion on how to handle versions of dependencies. The current docs suggest that we require Numpy 1.4, but at some point we'd probably want to increment this. In fact, I forgot that the dtype._convert pull request (PR #99) uses the function `promote_types` from numpy 1.6. (We need to either bump up the numpy version now or add a compatibility function.)
I think it is reasonable if the latest release of skimage relies on the latest stable release of numpy and matplotlib.
In addition, we should add a minimum (optional) version of Matplotlib for use in examples (this has come up in a few PRs already).
I recommend that we bump NumPy to 1.6 and Matplotlib to 1.0 (release 2011-01-03). The "subplots" functionality in 1.0 is really, really handy.
I run most of my scientific software from github, so I have no problem with running the bleeding edge, but this is certainly more difficult for others. How do other packages increment versions of dependencies?
True; but I think that, if you're running old versions of numpy and scipy, running an older version of skimage should be fine. If you need the latest, then it is reasonable to upgrade those packages as well. Backward compatibility comes at a cost, unfortunately! Regards Stéfan