We should probably add a utility function in python-graphblas to help convert expressions to objects. In python-graphblas, we frequently use A._expect_type_message(x, Matrix, ...) (and sometimes output_type(x)) to do this internally, which also raises informative error message if applicable. I think a simpler function would be nice. Should it respect "autocompute" config option?

• A = gb.ensure(A, gb.Matrix)
  • This is explicit, but what if we want Matrix or TransposedMatrix? Need to pass in both?
• A = gb.core.ensure_matrix(A), ensure_vector, ensure_scalar
  • Returns a Matrix or TransposedMatrix, computing if necessary
  • Can have allow_transpose= keyword argument (default True)

This can also have compute=None that defers to the "autocompute" config if unspecified.

I'll give this more thought and will try to give a concrete proposal of my preference.

Alright, here's my proposal, ensure_type(x, types): python-graphblas/python-graphblas#312

I like the ensure_type proposal. Let's merge that in python-graphblas and then use it here.

What can we do about the return type of generalized_degree? Our native return type is Matrix, but networkx is a dict of dicts, {row: {col: val}}. Perhaps we should pass the current function dispatch name to convert_to_nx.

For generalized_degree, we probably need to take the approach of returning a Graph from graphblas-algorithms, and then return a dict of dicts from the interface.

I know the goal is to avoid translation costs when using dispatching in "real" code, so returning a Graph from k_truss is ideal. However, if someone has code today which calls generalized_degree, they don't get back an nx.Graph, so the idea of passing the returned value from generalized_degree to k_truss (for example) isn't an option. They would expect exactly a dict of dicts.

I think we need to follow suit and return a dict of dicts so their code "just works".

But we should also raise an issue with NetworkX and ask why that is the return type. If there is a good reason, okay keep it as is. But if not, having a Graph object feels so much more practical.

Sorry, I misspoke. The NetworkX dict for generalized_degree is of the form {node: {degree: count}}, where node is the generic node label, degree is an integer that reflects the degree of a neighbor, and count is an integer of how many neighbors have that degree. So, this isn't quite a graph. It's more like a NodeMap where the values are dicts.

Our native return time is naturally a Matrix; see the magic here (from_values does heavy lifting):

So, I added a simple fix to matrix_to_dicts in #27, because our current conversion was wrong for non-default node ids. Perhaps we can also add matrix_to_nodemap conversion that can be treated as a dict of dicts and would work for generalized_degrees.

The more I've thought about it, the more I think we should add the function dispatch name to both convert_to_nx and convert_from_nx. This is pragmatic. In an ideal world where we have infinite time, we probably wouldn't need this. But, I think this will also be useful for the cugraph plugin.

Here's another argument for adding function dispatch name to test convert functions:

Some NetworkX functions/algorithms return numpy array or scipy.sparse array, but we'll want to return GraphBLAS vector or matrix. I think converting them during tests is the best option--we don't want to be forced to make our objects "look" like numpy or scipy.sparse, and it would be awkward (and difficult) to create a wrapper for them to look like numpy or scipy.sparse.

We should use convert_to_nx for tournament_matrix to convert to scipy.sparse.

Codecov Report

Base: 72.80% // Head: 73.63% // Increases project coverage by +0.83% 🎉

Coverage data is based on head (e48cec8) compared to base (4bb7cfb).
Patch coverage: 52.63% of modified lines in pull request are covered.

@@            Coverage Diff             @@
##             main      #25      +/-   ##
==========================================
+ Coverage   72.80%   73.63%   +0.83%     
==========================================
  Files          68       68              
  Lines        2335     2348      +13     
  Branches      430      428       -2     
==========================================
+ Hits         1700     1729      +29     
+ Misses        477      459      -18     
- Partials      158      160       +2     

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.