bug for log1p on csr_matrix

[XinyuHua]

I'm trying to use log1p function on csr_matrix. I use two different ways to construct the same csr_matrix but got different results after applying log1p.

Create the matrix by directly using a python list:

>>> mat_1_raw = [[2, 2, 0, 0, 0], [0, 0, 1, 1, 1]]
>>> mat_1 = csr_matrix(mat_1_raw)
>>> mat_1.log1p().toarray()
array([[ 1.09861229,  1.09861229,  0.        ,  0.        ,  0.        ],
       [ 0.        ,  0.        ,  0.69314718,  0.69314718,  0.69314718]])

Create the matrix by using indices:

>>> indices = [0, 1, 0, 1, 2, 3, 4]
>>> indptr = [0, 4, 7]
>>> data = [1, 1, 1, 1, 1, 1, 1]
>>> mat_2 = csr_matrix((data, indices, indptr), shape=(2, 5))
>>> mat_2.toarray()
array([[ 2.,  2.,  0.,  0.,  0.],
       [ 0.,  0.,  1.,  1.,  1.]])
>>> mat_2.log1p().toarray()
array([[ 1.38629436,  1.38629436,  0.        ,  0.        ,  0.        ],
       [ 0.        ,  0.        ,  0.69314718,  0.69314718,  0.69314718]])

Apparently in the second case there are two separate 1s and two separate 0s added to the same entry in the final csr_matrix, and the corresponding entry in the log1p result is log(1+1) * 2 = 1.39, where it should be log(1 + 2*1) = 1.099.

[XinyuHua]

Solved by adding the following before applying log1p:
>>> mat_2.sum_duplicates()