0

I'm trying to calculate the dot product between a scipy parse matrix and a numpy array.

First I was using a numpy matrix, which you can see in the following code:

def power_iteration(matrix, n):
    b_k = np.random.rand(matrix.shape[1])
    for _ in range(n):
        b_k = np.dot(matrix, b_k)

    return b_k

Here the matrix is a numpy matrix and no error occurs.

If you pass a scipy sparse matrix as a parameter, the following error occurs: ValueError: shapes (6762,6762) and (1,6762) not aligned: 6762 (dim 1) != 1 (dim 0)

I have changed

b_k = np.random.rand(matrix.shape[1])

into

b_k = np.random.rand(matrix.shape[1], 1)

which makes the dot product work, but doesn't return the right b_k shape. The shape I need is: (6762,)

Edit: so far I've tried to reshape like this:

b_k = np.reshape(b_k, (matrix.shape[1],))

but this transforms the shape (6762, 1) into (1, 6762), instead of (6762,)

Any tips? Thanks!

Lucas Belpaire
  • 1
  • 1
  • 5
  • Practice this on small arrays and matrix, and one loop at a time. Pay attention to `b_k` is - is it a 1d `ndarray`, or a 2d sparse matrix? A sparse matrix is always 2d. What does a dot with such a matrix produce? – hpaulj Apr 20 '19 at 18:04
  • @hpaulj The b_k is a 2d matrix with a shape of (6762, 1) in this case. After each dot product b_k still has a shape of (6762, 1), which is logical. I'm trying to convert the (6762,1) shape into (6762,) using reshape. I'm basing myself on this answer: https://stackoverflow.com/questions/17869840/numpy-vector-n-1-dimension-n-dimension-conversion But it doesn't seem to work – Lucas Belpaire Apr 20 '19 at 18:10
  • What is the `type` and `dtype` of `b_k` after one dot product? Is `matrix` is sparse, then `matrix.dot(b_k)`, and `np.dot(matrix.b_k)` will produce different objects. Mixing sparse and dense array has to be done with care. And if the intermediate `b_k` is `np.matrix` or `sparse` matrix, it can't be reshaped to 1d. – hpaulj Apr 20 '19 at 19:11
  • I've spotted a problem, the type of matrix is already . I had passed this as a parameter into the power iteration function: (d*matrix_s + (1 - d)*matrix_t), here d is a scalar and matrix_t is also a numpy.matrix. This means that before the first dot product, b_k is a numpy array and after the first it is a numpy matrix. – Lucas Belpaire Apr 20 '19 at 19:30

1 Answers1

0

Seems like in order to use np.dot on sparse matrix you'd need to convert it to dense matrix first with matrix.toarray(). Also see https://docs.scipy.org/doc/scipy/reference/sparse.html#matrix-vector-product

Anna Slastnikova
  • 1,260
  • 9
  • 9