ENH: argmax and argmin methods for sparse matrices #6761
Conversation
| mat.sum_duplicates() | ||
|
|
||
| line_size = mat.shape[axis] | ||
| ret = np.empty(mat.shape[1 - axis], dtype=int) |
There was a problem hiding this comment.
ret_size, line_size = mat._swap(mat.shape)
ret = np.zeros(ret_size, dtype=int)
| line_size = mat.shape[axis] | ||
| ret = np.empty(mat.shape[1 - axis], dtype=int) | ||
|
|
||
| for i in range(ret.shape[0]): |
There was a problem hiding this comment.
I don't think we can avoid the loop entirely, but we can at least vectorize the first condition:
nz_lines, = np.diff(mat.indptr) > 0
for i in nz_lines:
p, q = mat.indptr[i:i+2]
data = mat.data[p:q]
# etc...|
|
||
| D2 = D1.transpose() | ||
|
|
||
| classes = [bsr_matrix, coo_matrix, csr_matrix, csc_matrix] |
There was a problem hiding this comment.
When you integrate this with test_base.py, you can add the argmin/argmax tests to the _TestMinMax class and follow the existing tests as a template.
| for axis in [None, 0, 1]: | ||
| mat = spmatrix(D) | ||
| assert_raises(ValueError, mat.argmax, axis=axis) | ||
| assert_raises(ValueError, mat.argmin, axis=axis) |
There was a problem hiding this comment.
Some of these cases actually don't raise an error when mat is a numpy array:
In [1]: x = np.ones((5,0))
In [2]: x.argmax(axis=0)
Out[2]: array([], dtype=int64)
|
@perimosocordiae great advices, I think I handled them now. If you are fine with the updated state, I will move tests to test_base.py. Another thing I forgot to mention. I decided that returning ndarray is more convenient than any form of sparse matrix, because I expect that in majority of situation people will want to work with ndarray eventually. Is it right decision? |
|
Looks good to me. I agree that dense results are reasonable, considering that an argmin/argmax of zero doesn't necessarily indicate missing data. If we want to follow the numpy matrix convention (which spmatrix mimics), the result should be a matrix (row matrix for axis=0, column matrix for axis=1). On the other hand, argmax/argmin are typically then used for indexing, where a flat ndarray is typically the most useful. I'm leaning toward the matrix return type for now, but I could be convinced otherwise. |
|
Maybe I'm wrong on that, but it seems to me that people usually avoid using numpy matrices. Leaving consistency aside, I think having ndarray right away is more practical. At least for me it's very true. Leave to you to decide. |
|
I made sure that the minimum possible index is always returned and moved tests to test_base.py. Could you please make the final decision about whether to return array or matrix? |
|
|
||
| mat = self.spmatrix(D) | ||
|
|
||
| assert_equal(mat.argmax(), argmax) |
There was a problem hiding this comment.
Nitpick: I think it's clearer to have tests of the form:
assert_equal(mat.argmax(), np.argmax(D))
Rather than computing all the expected results first.
8273e40 to
3c85bb1
Compare
OK, maybe later we can change to array everywhere (like for 1.0 release). I changed to matrix for now. |
|
I want this in version 0.19, so merging now. Thanks, @nmayorov! |
The methods were added in gh-6761.
Methods were added in scipygh-6761.
Issue #5883
I think my approach is reasonable and efficient in terms of algorithmic complexity, but maybe it can be done with less Python loops.
For now I included tests as a separate file (instead of test_base.py). Just easier to test and demonstrate what's going on. At the end we can move it to test_base.py.
@perimosocordiae please look when you can.