dok_matrix no longer allows update

[amillb]

In previous versions, scipy allowed an update on sparse.dok_matrix. This seems to be no longer the case with v1.0.0.

With version 0.19.0, no error is thrown, and the matrix is updated accordingly.

The docs imply that the update method is still supported. If this isn't a bug, is there a suggested approach to update a dok_matrix?

Reproducing code example:

from scipy import sparse
m = sparse.dok_matrix((10, 10))
m.update({(1,1):'a'})

Error message:

NotImplementedErrorTraceback (most recent call last)
<ipython-input-15-b3e12160bd9e> in <module>()
      1 from scipy import sparse
      2 m = sparse.dok_matrix((10, 10))
----> 3 m.update({(1,1):'a'})
      4 

/Users/amb/Library/Enthought/Canopy/edm/envs/User/lib/python2.7/site-packages/scipy/sparse/dok.pyc in update(self, val)
    115     def update(self, val):
    116         # Prevent direct usage of update
--> 117         raise NotImplementedError("Direct modification to dok_matrix element "
    118                                   "is not allowed.")
    119 

NotImplementedError: Direct modification to dok_matrix element is not allowed.

Scipy/Numpy/Python version information:

('1.0.0', '1.14.0', sys.version_info(major=2, minor=7, micro=13, releaselevel='final', serial=0))

[WarrenWeckesser]

The change was made in #7673

@perimosocordiae, can you comment on this?

[...] is there a suggested approach to update a dok_matrix?

You can assign a value to a dok_matrix:

In [20]: m = sparse.dok_matrix((5, 10))

In [21]: m[1, 2] = -10.0

In [22]: m.A
Out[22]: 
array([[  0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.],
       [  0.,   0., -10.,   0.,   0.,   0.,   0.,   0.,   0.,   0.],
       [  0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.],
       [  0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.],
       [  0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.,   0.]])

[amillb]

You can assign a value to a dok_matrix:

Thanks. However, the update function is convenient for assigning multiple values at once from a dictionary. I suppose that I could iterate through that dictionary, but that's more cumbersome.

[perimosocordiae]

The update method sets values directly into the dict that backs a dok_matrix, a consequence of subclassing dict. We removed access to it intentionally because it bypasses the sanity checking that normal assignment provides, and makes it easy to accidentally put the matrix into an inconsistent state.

The recommended way to bulk-add elements to a DOK matrix is either (a) use fancy indexing to assign like you would for a dense array/matrix, or (b) loop over your values and use scalar-index assignment.

If you really must have the speed of a raw dict.update, you can use the internal dok_matrix._update method, which bypasses the checks and just inserts whatever you give it.

[WarrenWeckesser]

Wouldn't it be possible to implement an update method that acted like dict.update but also did the appropriate sanity checking?

[perimosocordiae]

Sure, that seems reasonable.