Could not broadcast input array from shape (1285)

问题:

I'm trying to follow some example code provided in the documentation for np.linalg.svd in order to compare term and document similarities following an SVD on a TDM matrix. Here's what I've got:

results_t = np.linalg.svd(tdm_t)
results_t[1].shape

yields

(1285,)

Also

results_t[2].shape
(5334, 5334)

So then trying to broadcast these results to create a real S matrix per the classic SVD projection approach, I've got:

S = np.zeros((results_t[0].shape[0], results_t[2].shape[0]), dtype = float)
S[:results_t[2].shape[0], :results_t[2].shape[0]] = results_t[1]

This last line produces the error:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-329-16e79bc97c4b> in <module>()
----> 1 S[:results_t[2].shape[0], :results_t[2].shape[0]] = results_t[1]

ValueError: could not broadcast input array from shape (1285) into shape (1285,5334)

What am I doing wrong here?

回答1:

So according to the error message, the target

S[:results_t[2].shape[0], :results_t[2].shape[0]]

is (1285,5334), while the source

results_t[1]

is (1285,).

So it has to broadcast the source to a shape that matches the target before it can do the assignment. Same would apply if trying to sum two arrays with these shapes, or multiply, etc.

• The first broadcasting step to make the number of dimensions match. The source is 1d, so it needs to be 2d. numpy will do try results_t[1][np.newaxis,:], producing (1, 1285).

• Second step is to expand all size 1 dimensions to match the other. But that can't happen here - hence the error. (1285,5334)+(1,1285)?

======

If you want to assign to a block (or all of S) then use:

 S[:results_t[2].shape[0], :results_t[2].shape[0]] = results_t[1][:,np.newaxis]

To assign r1 to a diagonal of S, use list indexing (instead of slices):

S[range(results_t[1].shape[0]), range(results_t[1].shape[0])] = results_t[1]

or

S[np.diag_indices(results_t[1].shape[0])] = results_t[1]

In this case those ranges have to match results_t[1] in length.

回答2:

(note, looking at Bi Rico's answer, it looks like perhaps the right interpretation was that you aren't actually trying to broadcast with that command? This answer shows how to actually do the broadcasting.)

X = scipy.zeros((5,4))
X
> array([[ 0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.],
       [ 0.,  0.,  0.,  0.]])

Y = scipy.arange(5)
Y
> array([0, 1, 2, 3, 4])
X[:,:]=Y
> could not broadcast input array from shape (5) into shape (5,4)

So instead try

X[:,:]=Y[:,None]
X
> array([[ 0.,  0.,  0.,  0.],
   [ 1.,  1.,  1.,  1.],
   [ 2.,  2.,  2.,  2.],
   [ 3.,  3.,  3.,  3.],
   [ 4.,  4.,  4.,  4.]])

You can get a bit more understanding of the issue from the following

Z = scipy.arange(4)
Z
> array([0, 1, 2, 3])
X[:,:]=Z
X
>array([[ 0.,  1.,  2.,  3.],
       [ 0.,  1.,  2.,  3.],
       [ 0.,  1.,  2.,  3.],
       [ 0.,  1.,  2.,  3.],
       [ 0.,  1.,  2.,  3.]])

The issue here is that it's convinced you're treating Y (or Z) as a row of the array, while you're trying to fit it into a column. Doing Y[:,None] basically forces it to interpret Y as a column.

回答3:

You're using slicing when you should be using indexing. Take for example:

A = np.zeros((4, 5))
end0, end1 = A.shape

# These are both true
A == A[:, :]
A[:, :] == A[:end0, :end1]

# To get the diagonal of an array you want to use range or np.arange
A[range(end0), range(end0)] == A.diagonal()

Some other things you might find useful:

# If U.shape is (a, b) and S.shape is (b,)
U.dot(np.diag(S)) == (U * S)

# IF V.shape (b, a) and S.shape is (b,)
np.diag(S).dot(V) == (S.reshape(b, 1) * V)