[SciPy-User] fmin_cg fmin_bfgs "Desired error not necessarily achieveddue to precision loss"
Wed Nov 23 09:02:09 CST 2011
On 23/11/2011 14:48, email@example.com wrote:
> On Wed, Nov 23, 2011 at 4:41 AM, Dan Stowell
> <firstname.lastname@example.org> wrote:
>> Anyone got any suggestions about this "precision loss" issue, please?
>> I found this message from last year, suggesting that using dot instead
>> of sum might help (yuck):
>> - but no difference here, I still get the optimisation stopping after
>> three iterations with that complaint.
> something is wrong with the gradient calculation
> If I drop fprime in the call to fmin_bfgs, then it converges after 11
> to 14 iterations (600 in the last case)
> fmin also doesn't have any problems with convergence
> (I'm using just float64)
Thanks, you're absolutely right. (Also, plain 'fmin' converges easily.)
I've found the problem now. I was assuming that the function would
preserve the shape of my parameter vector (a column vector), whereas it
was feeding a row vector into my functions, causing wrong behaviour. A
bit of reshape fixed it.
>> Any tips welcome
>> On 19/11/11 19:19, Dan Stowell wrote:
>>> I'm translating a fairly straightforward optimisation code example from
>>> octave. (Attached - it does a quadratic regression, with a tweaked
>>> regularisation function.)
>>> Both fmin_cg and fmin_bfgs give me poor convergence and this warning:
>>> "Desired error not necessarily achieveddue to precision loss"
>>> This is with various regularisation strengths, with normalised data, and
>>> with high-precision data (float128).
>>> Is there something I can do to enable these to converge properly?
>>> (Using ubuntu 11.04, python 2.7.1, scipy 0.8)
>> Dan Stowell
>> Postdoctoral Research Assistant
>> Centre for Digital Music
>> Queen Mary, University of London
>> Mile End Road, London E1 4NS
>> SciPy-User mailing list
> SciPy-User mailing list
Postdoctoral Research Assistant
Centre for Digital Music
Queen Mary, University of London
Mile End Road, London E1 4NS
More information about the SciPy-User