Pass-by-reference is not supported for Sparse types

[ecolss]

As described in the passy-by-reference doc, I wonder:

1. is there any particular reason not to support ref for sparse types?
2. since not supported, if I have sparse matrix constructed in python via scipy.sparse.csr_matrix, then have to pass it by value to c++, would this bring much overhead if I have to do a mini-batch processing on the data?

[jagerman]

The main obstacle is that numpy's sparse storage is fundamentally different than Eigen's, so it isn't possible to have one simply reference the other's data. For dense types, all you need is a chunk of data and strides to map between them. That isn't to say something better can't be done, just that no one has looked into it.

The overhead is going to be the effort involved in copying it, which will typically come down to the number of non-sparse values that need to be copied from one side to a new type on the other side.

[PhilipDeegan]

Please forgive my ignorance, but it seems a bit strange to me to require Eigen for numpy.

I'm having trouble with scipy csr_matrix at the moment and could use a simple example - the tests are not so obvious

[PhilipDeegan]

I've gotten something working

built like

used like