How do you handle variably-dimensioned problems?

[johannahaffner]

Thank you for the clean interface! It is really useful to me.

To validate software in JAX, I need to port benchmark problems to JAX proper. To verify that the problems I ported are correct, I'm currently testing them against pycutest.

My tests fail for problems with variable dimensions, which makes sense - the pycutest version always seems to have the default number of variables, while I created instances with all supported numbers of variables. Looking at the source code for the CUTEst problem, I could not immediately see if variably-dimensioned problems are supported at all, and if they are then how specifically this is done.

Best

Johanna

[lindonroberts]

PyCUTEst supports variable-dimensional problems, as long as the underlying SIF file (used in the main Fortran CUTEst package) supports the specific choice. You do this by specifying parameters for the SIF file. Not all parameter values are allowed by CUTEst.

For a given problem, you can check which parameter combinations are allows using the pycutest.print_available_sif_params function (documentation). Alternatively, you can look at the raw SIF file in your $MASTSIF directory.

Once you know what parameter values you want, you can compile in PyCUTEst using the sifParams optional argument for pycutest.import_problem.

There is an example in the documentation for the problem ARGLALE:

pycutest.print_available_sif_params('ARGLALE')

produces a list of the allowed optional parameters

Parameters available for problem ARGLALE:
N = 10 (int)
N = 50 (int)
N = 100 (int)
N = 200 (int) [default]
M = 20 (int, .ge. N)
M = 100 (int, .ge. N)
M = 200 (int, .ge. N)
M = 400 (int, .ge. N) [default]
End of parameters for problem ARGLALE

This says that you can specify one/both of parameters N and M. Both are integer parameters, with defaults N=200 and M=400. The allowed values for N are N=10, 50, 100, 200, and for M are M=20, 100, 200, 400, with the added constraint that you must have M >= N. (if you read ARGLALE.SIF, it turns out N is the dimension of the problem in this case)

So, if you want to use the default values N=200 and M=400 then you can do

problem = pycutest.import_problem('ARGLALE')

but if you want to use a different allowed choices, you can provide a dictionary of the desired parameter values, such as:

problem = pycutest.import_problem('ARGLALE', sifParams={'N':100, 'M':200})

[johannahaffner]

Amazing, thank you!

[jfowkes]

Re-opening for others that may have the same question.