No description provided.
yeesh must've slipped on my keyboard. thanks
I don't see these failures locally. I'll try to install 1.6.1 locally. I assume matrix_rank is available in our minimum numpy.
matrix_rank is available in numpy 1.6, but at some point Matthews change the defaults for the threshold. I don't remember if this happened after 1.6, but I think so.
(I have problems looking at TravisCI because I'm using internet explorer.)
Is 1.6 our minimum version? We list 1.5 on the web...
Looks like its behavior changed in 1.7 or 1.8 numpy/numpy@78b7693
Thanks. I can replicate one of these failures with 1.6.0. Odd. So I guess this PR is a non-starter? Is it worth writing a wrapper to adjust the tolerance?
The current numpy matrix_rank implementation could be copypasted into statsmodels, with the intention that it would be deleted when the required numpy version reaches 1.7
We'll see if that works.
if numpy 1.10 is released this will break everyone's version string comparisons ('1.10' < '1.8' < '1.9')
'1.10' < '1.8' < '1.9'
Yeah we need to switch to LooseVersion. I was going to do it in another PR.
Apparently LooseVersion doesn't work for all of the alpha, beta, release candidate, development versions, etc. that are used by numpy and scipy, so this was recently written.
Ah, thanks for the pointer, I hadn't seen that.
MAINT: Deprecate rank in favor of np.linalg.matrix_rank
COMPAT: Add np.linalg.matrix_rank for <= 1.7.1
ENH: Use NumpyVersion for version testing