compute_polytope_vertices does not work for large matrices

[paulpacaud]

compute_polytope_vertices does not work for large matrices

[stephane-caron]

compute_polytope_vertices is an application of the more general get_generators() from cdd, where we further assume that the output is a polytope (get_generators works with general polyhedra, which can have both rays and vertices).

[stephane-caron]

It seems indeed that cdd does not find any vertex. It could be that the set is empty, or the problem hits some numerical issue.

One small piece of advice: your NumPy arrays will have an integer datatype, but compute_polytope_vertices assumes floating-point numbers. One advice (not changing the outcome here) is to add dtype=np.float64 to the calls to np.array to make sure all numbers are treated as floating point scalars.

To go further, I'd recommend you dig into what the function is doing, that is, taking the code out and managing the pycddlib objects directly. For instance rather you can try another number type:

pypoman/pypoman/duality.py

Line 137 in e22a7a7

mat = cdd.Matrix(np.hstack([b, -A]), number_type="float")

I don't know whether that's the case but it may also be possible to get more verbose outputs from cdd to see why it returns no generator.

Hoping this helps!

[paulpacaud]

Thank you for your reply. Picking up on what you said, to enumerate vertices of a polytope, what's the difference between compute_polytope_vertices from pypoman and poly.get_generators() from pycddlib ?