np.array() is not working when i gave a list including two ndarray with different shape

Question

This is a code. I made a list including two ndarray with different shape.

d = []

a = np.arange(183).reshape(3,61)
b = np.arange(51).reshape(3,17)

d = [a,b]

np.array(d)

Error is like below.

 File "C:\Program Files\JetBrains\PyCharm 2019.1.1\helpers\pydev\_pydev_imps\_pydev_execfile.py", line 18, in execfile
    exec(compile(contents+"\n", file, 'exec'), glob, loc)
  File "D:/dev/workspace/rl/test/npcopy.py", line 10, in <module>
    np.array(d)
ValueError: could not broadcast input array from shape (3,61) into shape (3)

np.copy() works when two ndarrays' first shpae are different. but if not, it is not working like above.

if I change this code as below,

import numpy as np

d = []

a = np.arange(183).reshape(4, 61)
b = np.arange(51).reshape(3, 17)

d = [a,b]

np.array(d)

it works!! so weird!!

Its expected. You cannot club two matrices of different dimensions into one n dimensional array — Bhaskar, May 21 '19 at 11:48
@DeveshKumarSingh I want it to be list including two ndarrays. — verystrongjoe, May 21 '19 at 11:50
@DeveshKumarSingh The type would be [ndarrry (3,61), ndarray(3,17)] — verystrongjoe, May 21 '19 at 11:51
if that, this code run in function, so that it will be delivered as call by reference and needs to be executed as different value. — verystrongjoe, May 21 '19 at 11:55

score 3 · Answer 1 · answered May 21 '19 at 12:09

3

As matrices are of different dimensions

> a = np.arange(183).reshape(3,61) b = np.arange(51).reshape(3,17)
> d=[np.array(a),np.array(b)] 
>  print(d) for output
> 
> or  d=[a,b]
>  np.concatenate(d, axis=1)

answered May 21 '19 at 12:09

AnkushRasgon

782
1
6
14

score 1 · Accepted Answer · answered May 21 '19 at 17:11

When you try to make an array from arrays there are three possible results:

If the arrays have the same shape, the result is a higher dimension array:

In [295]: np.array((np.zeros((2,3),int),np.ones((2,3),int)))                 
Out[295]: 
array([[[0, 0, 0],
        [0, 0, 0]],

       [[1, 1, 1],
        [1, 1, 1]]])
In [296]: _.shape                                                            
Out[296]: (2, 2, 3)

If the arrays differ in shape, the result could be an object dtype array (similar to a list):

In [298]: np.array((np.zeros((2,3),int),np.ones((3,3),int)))                 
Out[298]: 
array([array([[0, 0, 0],
       [0, 0, 0]]),
       array([[1, 1, 1],
       [1, 1, 1],
       [1, 1, 1]])], dtype=object)    # shape (2,)

But for some combinations of shapes, the result is an error:

In [301]: np.array((np.zeros((2,3),int),np.ones((2,4),int)))                 
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-301-d67e6685581d> in <module>
----> 1 np.array((np.zeros((2,3),int),np.ones((2,4),int)))

ValueError: could not broadcast input array from shape (2,3) into shape (2)

In the error case, the first dimensions match, just as in your first case.

Sometimes to create an object array you have to start with a 'empty' one, and fill it. This is more reliable than the np.array(...) approach.

In [303]: arr = np.empty(2, object)                                          
In [304]: arr[:] = np.zeros((2,3),int),np.ones((2,4),int)                    
In [305]: arr                                                                
Out[305]: 
array([array([[0, 0, 0],
       [0, 0, 0]]),
       array([[1, 1, 1, 1],
       [1, 1, 1, 1]])], dtype=object)
In [306]: arr[:] = np.zeros((2,3),int),np.ones((2,3),int)                    
In [307]: arr                                                                
Out[307]: 
array([array([[0, 0, 0],
       [0, 0, 0]]),
       array([[1, 1, 1],
       [1, 1, 1]])], dtype=object)

thanks for your answer! so this is a numpy bug? and you suggested workaround of it? — verystrongjoe, May 22 '19 at 01:14
I wouldn't describe it as a bug. The normal, expected behavior is the first case. The others are, in a sense ways of making do with bad inputs. I'd like to see `np.array` throw an error in both cases unless an object dtype was specified. I suggested the most reliable way of making an object dtype array. — hpaulj, May 22 '19 at 01:37

score 0 · Answer 3 · answered May 21 '19 at 12:11

0

Instead of using np.copy(), I chose to use copy.deepcopy(). It works even if the first shape of two items in list are same.

answered May 21 '19 at 12:11

verystrongjoe

3,831
9
35
66

np.array() is not working when i gave a list including two ndarray with different shape

3 Answers3

Linked