scipy.sparse.bmat fails if all submatrices have same number of rows

[andreasbuhr]

scipy.sparse.bmat fails if all submatrices have same number of rows

Reproducing code example:

import numpy as np
import scipy
import scipy.sparse

blocks = [
[ np.ones((2,3)), np.ones((2,4))],
[ np.ones((2,3)), np.ones((2,4))]
]

scipy.sparse.bmat(blocks)

Error message:

Traceback (most recent call last):
  File "bug.py", line 12, in <module>
    scipy.sparse.bmat(blocks)
  File "/usr/lib/python2.7/dist-packages/scipy/sparse/construct.py", line 544, in bmat
    blocks = np.asarray(blocks, dtype='object')
  File "/usr/lib/python2.7/dist-packages/numpy/core/numeric.py", line 482, in asarray
    return array(a, dtype, copy=False, order=order)
ValueError: could not broadcast input array from shape (2,3) into shape (2)

Scipy/Numpy/Python version information:

('0.17.0', '1.11.0', sys.version_info(major=2, minor=7, micro=12, releaselevel='final', serial=0))

[pv]

That error is from your own code, before the execution reaches scipy.sparse.bmat

[andreasbuhr]

yes, sorry, was a typo while reducing the reproducer. Now the code is as intended, after edit

[pv]

Workaround: pass the input as an object array >> blocks = np.zeros([2, 2], dtype='object') >> blocks[0,0] = np.ones((2,3)) ... >> scipy.sparse.bmat(blocks)

[perimosocordiae]

Reproduced with scipy 1.0.0 and numpy 1.13.3. It seems that np.array has trouble handling the nested lists containing submatrices, despite the dtype=object argument.

[pv]

bmat could be ok just doing `n = len(blocks); m = len(blocks[0])`, maybe as a fallback...

[andreasbuhr]

I worked around it by appending a row of matrices with zero rows each to "blocks".

[perimosocordiae]

I'm not familiar with how np.array coercion logic works, but this seems like a bug to me.