AttributeError: 'Matrix' object has no attribute 'data' (uBlas -> Eigen)

[Andreas Hellander]

After upgrading to 1.6 from 1.5 (switching from uBLAS to Eigen backend), I get this error (where 'M' is an output of dolfin.assemble):

Did you intentionally introduce this change or is this a bug?

/usr/local/pyurdme/pyurdme/pyurdme.pyc in create_system_matrix(self) 636 --> 638 rows, cols, vals = M.data() 639 SM = scipy.sparse.csr_matrix((vals, cols, rows)) 640 vols = SM.sum(axis=1)

AttributeError: 'Matrix' object has no attribute 'data'

[Prof Garth Wells]

This is a deliberate change. Matrix is an abstract wrapper class, so can't make any assumptions as to the underlying data storage.

To use the data() member function, use an EigenMatrix rather than a Matrix. See ChangeLog for more information.

[Christian Clason]

Or cast it explicitly to the backend type (if it supports data access, as Eigen does): rows, cols, vals = as_backend_type(M).data().

[Andreas Hellander]

Thanks. And this, rows, cols, vals = as_backend_type(M).data() will work also with 1.5 and the uBLAS backend? (I want to update my API in a backwards compatible way)

I must have missed the deprecation warning to remove the uBLAS backend in the 1.5 release.

[Christian Clason]

Yes, this also works for the uBLAS backend.

The change is mentioned in the ChangeLog:

- Remove GenericMatrix/Vector::data() and GenericMatrix/Vector::data()
    (to use backends that support data(), cast first to backend type, e.g.
    A = A.as_backend_type()

As far as I can tell, there's no release notes on the webpage yet -- but this is something that should indeed be mentioned prominently at the very top (under "Backward-incompatible changes" or "Major API changes").

[Martin Sandve Alnæs]

Removing milestone: 1.6 (automated comment)