[SciPy-User] L-BFGS broken in Ubuntu's scipy builds?

[Lars Buitinck]

Dear all,

I was trying to implement logistic regression using the fmin_l_bfgs_b
from SciPy 0.11/0.12 as shipped by Ubuntu 13.04/13.10 and nothing
seemed to work. Fearing I had done something wrong in implementing the
gradient from Bishop's book, I turned to Fabian Pedregosa's
implementation [1] to see if that would work. I removed the callback
and maxiter options that aren't available in 0.11 from the script,
then tried the following combinations:

* SciPy 0.11.0 (Ubuntu), NumPy 1.7.1 (Ubuntu): ABNORMAL_TERMINATION_IN_LNSRCH
* SciPy 0.12.0 (Ubuntu), NumPy 1.7.1 (Ubuntu): ABNORMAL_TERMINATION_IN_LNSRCH
* SciPy 0.12.0 (fresh build), NumPy ad5bded (GitHub): converges
* SciPy 0.13.0rc1, NumPy ad5bded (GitHub): converges

I also tried a freshly built SciPy 0.11.0 from GH but it failed to
load something from LAPACK. I got the same error from my own code. A
numerical gradient check pointed out no obvious flaws in my code.

Is this a known issue with Ubuntu's SciPy build? I saw a bunch of
commits related to line search issues in the Git log, could any of
those be related?

[1] http://nbviewer.ipython.org/github/fabianp/pytron/blob/master/doc/benchmark_logistic.ipynb


Regards,
Lars Buitinck