Join GitHub today
GitHub is home to over 28 million developers working together to host and review code, manage projects, and build software together.Sign up
[MRG+1] FIX consistency of memory layout for linear CD solver #5337
This should fix #5013.
This PR fixes several related consistency issues in the handling of the memory layout in models using the linear coordinate descent solvers (Gram or not).
Also I made the expectation w.r.t. memory layout explicit in the Cython prototypes directly which should prevent re-introducing regressions in the future.
@arthurmensch I would appreciate a review on this as I touched a lot of code you changed when introducing the ability to skip input checks. Especially if you have your benchmark scripts at hand it would be great if you could check that I do not re-introduce unwanted redundant input checks.
BTW, I still get a memory error reported by valgrind when I use openblas from ubuntu 14.04 but I think this is in openblas it-self and it does not seem to cause an error at our level. There might be a bug in openblas but I think this PR fixes the issue reported in #5013.
referenced this pull request
Oct 2, 2015
I see no performance regression on dictionary learning, so
Looking back on
if hasattr(precompute, '__array__') and ( fit_intercept and not np.allclose(X_mean, np.zeros(n_features)) or normalize and not np.allclose(X_std, np.ones(n_features))):
despite being important for performance. But that is another issue.