Qiskit
Allow passing numpy array to `PyDiGraph::extend_from_edge_list`
What is the expected enhancement?
Per current api, one way to extend an existing graph is to use the PyDiGraph::extend_from_edge_list. This method expects a list of tuples. The enhancement would be extend this method so that it accepts a numpy array. This could possibly improve performance if the underlying memory array is directly accessed.
This issue came up in a discussion about the performance of qiskit.transpiler.CouplingMap object initialization: https://github.com/Qiskit/qiskit/issues/11919
I think this should be an OK addition, but it will need to be a separate method.
We had issues with conversions from ndarray -> Rust types before and that is why we have from_complex_adjacency_matrix and from_adjacency_matrix. So let's try to avoid those and just create a new method. Upstream users will need to adapt.
Ivan: an alternative is to accept a Py<PyAny> in the slot and implement the conversion logic using the PyO3 traits manually (with a little bit of type checking). That's more complex from the implementation side, but potentially more convenient for users?
Ivan: an alternative is to accept a
Py<PyAny>in the slot and implement the conversion logic using the PyO3 traits manually (with a little bit of type checking). That's more complex from the implementation side, but potentially more convenient for users?
I think it is possible but I would like to keep the method straightforward. Nominally, I don't want to deal with two things:
- Handling people passing Python objects that are
Sequence[tuple[int, int]]but are not lists. I do not want to accidentally cast that to NumPy - Handling the different NumPy types. For even the slightest mismatch we get stuff like
TypeError: argument 'matrix': type mismatch: from=uint8, to=float64(this is like #411 but I called adjacency matrix withuint8from NumPy)
So I'd rather implement a separate function and document well the specific ndarray format and type we want