Optimization demo looks broken

[mattloper]

Optimization demo looks broken

[mattloper]

What version of chumpy are you running? Is it from pipy or from latest master on github? For me, ch.demo('optimisation') is working.

[chenliu0831]

I'm using the latest master.

[mattloper]

From my testing, it seems that Mavericks exhibits no problem, Ubuntu 12.04 gives a warning but still works, and Ubuntu 14.04 hangs. I don't know what is causing the problem, or why behavior is different across platforms.

[chenliu0831]

Oh. It actually hangs on my mac. Probably the default numpy and scipy version is different across platform. I am using scipy==0.14.0, numpy==1.8.2 on my mac.

[chenliu0831]

@mattloper Sorry I accidentally deleted the previous comment. In scipy 0.14.0, scipy.sparse.linalg.spsolve is filling nan if singular matrix is detected. See diff line 139 in scipy/scipy@2344e41

[mattloper]

Finally fixed this, by falling back to scipy.sparse.linalg.lsqr when warnings are fired (typically with a singular matrix). The problem is that when any single column of the Jacobian is all-zeros, the predicted hessian is non-invertable. Sometimes such zero-columns indicates user-error (optimizing over vars that don't affect output), other times not.