T-Matrix trimer test fit fails on Windows x64

Yes. The fit status (1 vs. 3) is what caused that test to fail. For a trimer, the fit returns 19 parameters. We check each of these parameters against results obtained previously (the gold standard).

The fit status parameter is certainly important as far as "<5" vs. "=5". When the status equals 5, it means that the fitting algorithm stopped iterating because it reached the maximum allowed number of iterations (the fit did not necessarily converge to a good solution). When the status is less than 5, it means that the fitter had a different (probably better) reason to not do any more iterations.

Does it make sense that different machines would come up with different codes 1-4?

We could relax the test to check that the fit status parameter is simply <5 if it makes sense for different machines to have different codes 1-4 when running the same fit.

I think it makes sense that the fit status could be different for different computers. I think compiler differences might affect how numpy or the c/fortran code does calculations down near the limits of machine precision. This could in turn affect the path that the fitter takes to the solution, and thus potentially the termination conditions it reaches.

Someone who actually knows what the termination conditions means well (Jerome?) should weigh in, but it seems to me that Becca's idea of treating fit status 1-4 as equivalent should be safe, since we are testing that all the parameter values (the results we care about) agree to good precision.

According to the nmpfit code at
http://stsdas.stsci.edu/pyraf/stscidocs/pytools_pkg/pytools_api/pytools.nmpfit.mpfit-class.html
, fit status 3 means that criteria for both fit status 1 (sum of squares
has converged) and status 2 (parameter vector has converged) apply. So
it certainly makes sense to change the test to allow fit status to be
either 1 or 3. We probably also want to allow fit status 2 and 4. Note
that there are other fit failure codes besides status=5.

Could someone make the change and commit to stable?

On 06/09/2011 04:54 PM, Tom Dimiduk wrote:
> I think it makes sense that the fit status could be different for
> different computers. I think compiler differences might affect how
> numpy or the c/fortran code does calculations down near the limits of
> machine precision. This could in turn affect the path that the fitter
> takes to the solution, and thus potentially the termination conditions
> it reaches.
>
> Someone who actually knows what the termination conditions means well
> (Jerome?) should weigh in, but it seems to me that Becca's idea of
> treating fit status 1-4 as equivalent should be safe, since we are
> testing that all the parameter values (the results we care about) agree
> to good precision.
>

I uploaded a fix to the stable branch-- could someone please try running the tests on Windows 7 x64 (EPD 7.0-1) to make sure I they pass? They still pass on Ubuntu.

Test passes now. Thanks!