Join GitHub today
GitHub is home to over 40 million developers working together to host and review code, manage projects, and build software together.
Sign upENH safe_sparse_dot work on 3D+ arrays #14538
Conversation
| @@ -117,29 +117,40 @@ def density(w, **kwargs): | |||
| def safe_sparse_dot(a, b, dense_output=False): | |||
| """Dot product that handle the sparse matrix case correctly | |||
| Uses BLAS GEMM as replacement for numpy.dot where possible | |||
| to avoid unnecessary copies. | |||
This comment has been minimized.
This comment has been minimized.
| if dense_output and hasattr(ret, "toarray"): | ||
| ret = ret.toarray() | ||
| return ret | ||
| if a.ndim > 2 or b.ndim > 2: |
This comment has been minimized.
This comment has been minimized.
rth
Aug 1, 2019
Member
Does this really work for ND arrays? I would restrict it to 3D, which is what we are using, and raise a NotImplementedError otherwise. We are not testing 4D+ case anyway.
This comment has been minimized.
This comment has been minimized.
jeremiedbb
Aug 1, 2019
Author
Member
Yes it does :)
I updated the tests to ensure that. I don't see any reason to raise an error since it's implemented and works.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
This comment has been minimized.
jeremiedbb
Aug 2, 2019
Author
Member
At least in coordinate descent (see diff) between 3D dense and sparse.
| ret = ret.toarray() | ||
| return ret | ||
| if a.ndim > 2 or b.ndim > 2: | ||
| if sparse.issparse(a): |
This comment has been minimized.
This comment has been minimized.
| b_2d = b_.reshape((b.shape[-2], -1)) | ||
| ret = a @ b_2d | ||
| ret = ret.reshape(a.shape[0], *b_.shape[1:]) | ||
| elif sparse.issparse(b): |
This comment has been minimized.
This comment has been minimized.
|
I wouldn't have been sure about supporting >=3d, but I'm happy with this. I'd like to change the title to not say FIX... it's a DOC fix because the documented behaviour is unreasonable ;) Add what's new? |
jeremiedbb
changed the title
This comment has been minimized.
This comment has been minimized.
|
@rth I think I adressed your comments |
|
Minor comment otherwise LTGM. |
| sparse matrix. | ||
| |Fix| The documented behavior of the ``dense_output`` parameter of | ||
| :func:`utils.safe_sparse_dot` was not the intended behavior and now matches | ||
| the actual behavior of the function. |
This comment has been minimized.
This comment has been minimized.
rth
Sep 9, 2019
Member
So the second part is a documentation fix? Not sure it needs a what's new entry?
This comment has been minimized.
This comment has been minimized.
jeremiedbb
Sep 10, 2019
Author
Member
I was not sure about that. It's true that the behavior did not change, only it's documentation. I removed it.
rth
changed the title
rth
merged commit 61add4d
into
scikit-learn:master
jeremiedbb commentedAug 1, 2019
•
edited
safe_sparse_dotdocstring was not accurate regardingdense_outputparameter. If only one of them is sparse, it would always return a dense array.I don't think it was an intended behavior anyway since I saw several
which would break if the output was sparse. Moreover, it's very unlikely that dense @ sparse has many zeros. So I changed the docstring to match the behavior.
It did not handle the case when one of the arrays is 3D+ and the other is sparse.
When arrays are 1D or 2D, matmul is preferred over dot, and is a little bit faster. I replace dot by @ when applicable.
Function was untested.