I'm doing maintenance on my old box, updating things, etc. and I noticed this.
I don't think this check actually works. AFAICT callback is never removed from the kwargs.
Since all the ARMA models pass callback=None up to fit, the lack of reports suggests to me that we don't have anyone running scipy <= 0.12...
Should at least be fixed for 0.5.1 assuming this refactor affects the 0.5.x series.
AFAIK, TravisCI uses scipy 0.9 (Ubuntu LTS)
I jswitched recently to scipy 0.12 and to 0.13.3
Well, all these tests fail for me with scipy 0.12.0.dev-006503d.
is 0.12.0.dev-006503d a released version or an old buggy scipy master?
I have to check again, I'm pretty sure I used 0.12.0 for a while, but I might never have run the full test suite with it. I think at least one of pythonxy Ubuntu or Nipy Debian uses 0.12.0.
It is also possible that I never tested for 0.9 < scipy < 0.12.0, but I saw 0.10 also on one of Ubuntu or Debian test machines, IIRC
Yeah, possible it works in releases, but I don't see callback earlier (I get an error that this keyword doesn't exist), and I don't see that we actually remove the keyword.
maybe this is or isn't related to your issue, but LooseVersion doesn't work completely correctly with numpy or scipy. There's a scipy thing called 'NumpyVersion' that works.
It is also copypasted into other projects like
It would be tempting to just put it into numpy to avoid all this copypasting, but this doesn't work for obvious reasons...