How to convert a numpy array dtype=object to a sparse matrix?

Question

I have an numpy array of dtype = object containing multiple other arrays for elements and I need to convert it to a sparse matrix.

Ex:

a = np.array([np.array([1,0,2]),np.array([1,3])])
array([array([1, 0, 2]), array([1, 3])], dtype=object)

I have tried the solution given by Convert numpy object array to sparse matrix with no success.

In [45]: M=sparse.coo_matrix(a)
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-45-d75020bb3a38> in <module>()
----> 1 M=sparse.coo_matrix(a)

/home/arturcastiel/.local/lib/python3.6/site-packages/scipy/sparse/coo.py in __init__(self, arg1, shape, dtype, copy)
    183                     self._shape = check_shape(M.shape)
    184 
--> 185                 self.row, self.col = M.nonzero()
    186                 self.data = M[self.row, self.col]
    187                 self.has_canonical_format = True

ValueError: The truth value of an array with more than one element is ambiguous. Use a.any() or a.all()

As it was explained on the comments, this is actually a jagged array. In essence, this array represents a graph that I have to convert to sparse matrix so I can use the scipy.sparse.csgraph.shortest_path routine.

Thus,

np.array([np.array([1,0,2]),np.array([1,3])])

should become something such as:

(1,1) 1
(1,2) 0
(1,3) 2
(2,1) 1
(2,2) 3

Answer 1

I'd consider using a dok_matrix if your arrays have a lot of omissed trailing zeros:

In [98]: dok = sparse.dok_matrix((2, 3), dtype=np.int64)

In [99]: for r_num, row in enumerate(a):
    ...:     for col_num, el in enumerate(row):
    ...:         dok[r_num, col_num] = el 
    ...:         

In [100]: dok.toarray()
Out[100]: 
array([[1, 0, 1],
       [1, 3, 0]], dtype=int64)

Answer 2

You can't. This error arises when it tries to find the nonzero elements of a. A sparse matrix just stores the nonzero elements of a matrix. Try

np.nonzero(a)  

If your array contained lists instead of arrays, it would work - sort of:

In [615]: a = np.array([[1,0,1],[1,3]])                                              
In [616]: np.nonzero(a)                                                              
Out[616]: (array([0, 1]),)

In [618]: sparse.coo_matrix(a)                                                       
Out[618]: 
<1x2 sparse matrix of type '<class 'numpy.object_'>'
    with 2 stored elements in COOrdinate format>
In [619]: print(_)                                                                   
  (0, 0)    [1, 0, 1]
  (0, 1)    [1, 3]

Note this is a (1,2) shaped array, with 2 nonzero elements, both of which are the lists (objects) of the original.

But coo format does little processing. It can't for example be converted to csr for computations:

In [622]: _618.tocsr()                                                               
---------------------------------------------------------------------------
TypeError: no supported conversion for types: (dtype('O'),)

If the array wasn't jagged, it could be made into a useful sparse matrix:

In [623]: a = np.array([[1,0,1],[1,3,0]])                                            
In [624]: a                                                                          
Out[624]: 
array([[1, 0, 1],
       [1, 3, 0]])

In [626]: sparse.coo_matrix(a)                                                       
Out[626]: 
<2x3 sparse matrix of type '<class 'numpy.int64'>'
    with 4 stored elements in COOrdinate format>
In [628]: print(_)                                                                   
  (0, 0)    1
  (0, 2)    1
  (1, 0)    1
  (1, 1)    3

note that the 0 values have been omitted. In large useful sparse matrices, more than 90% of the elements are zero.

===

Here's a way of constructing a sparse matrix from your array of arrays. I build the row,col,data attributes of a coo format matrix from the individual arrays in a.

In [630]: a = np.array([np.array([1,0,1]),np.array([1,3])])                          
In [631]: row, col, data = [],[],[]                                                  
In [632]: for i,n in enumerate(a): 
     ...:     row.extend([i]*len(n)) 
     ...:     col.extend(np.arange(len(n))) 
     ...:     data.extend(n) 
     ...:                                                                            
In [633]: row,col,data                                                               
Out[633]: ([0, 0, 0, 1, 1], [0, 1, 2, 0, 1], [1, 0, 1, 1, 3])
In [634]: M = sparse.coo_matrix((data, (row,col)))                                   
In [635]: M                                                                          
Out[635]: 
<2x3 sparse matrix of type '<class 'numpy.int64'>'
    with 5 stored elements in COOrdinate format>
In [636]: print(M)                                                                   
  (0, 0)    1
  (0, 1)    0
  (0, 2)    1
  (1, 0)    1
  (1, 1)    3
In [637]: M.A                                                                        
Out[637]: 
array([[1, 0, 1],
       [1, 3, 0]])

An alternative to is to pad a to make a 2d numeric array, and make the sparse one from that. Padding a jagged list/array has been asked before, with various solutions. This is one of the easier ones to remember and use:

In [658]: alist = list(zip(*(itertools.zip_longest(*a,fillvalue=0))))                                                                            
In [659]: alist                                                                      
Out[659]: [(1, 0, 1), (1, 3, 0)]
In [661]: sparse.coo_matrix(alist)                                                   
Out[661]: 
<2x3 sparse matrix of type '<class 'numpy.int64'>'
    with 4 stored elements in COOrdinate format>
In [662]: _.A                                                                        
Out[662]: 
array([[1, 0, 1],
       [1, 3, 0]])