使用 boolean 索引将向量作为对象添加到 numpy 数组 - 2D arrays 有效，1D arrays 无效

Question

I want to assign vectors to a numpy array to make convenient use of the boolean indexing. It worked, until I hit a corner case. Minimal working example:

a = np.array([None] * 2)
b = [np.ones(shape=[1, 3]), np.ones(shape=[2, 4])]
a[True, True] = b  # Works as intended

print(a)

However, when all arrays are 1D, I get an error:

a = np.array([None] * 2)
b = [np.ones(shape=[1, 3]), np.ones(shape=[1, 4])]  # shape of second array changed!
a[True, True] = b  # ValueError: cannot copy sequence with size 2 to array axis with dimension 1

Any ideas how I can avoid this? (I checked on both numpy versions v1.19.5 and v1.21.1)

Answer 1

I solved it. If you use a np.where, the code runs. Not sure why. Example:

a = np.array([None] * 2)
b = [np.ones(shape=[1, 3]), np.ones(shape=[1, 4])]  # shape of second array changed!
a[np.where([True, True])] = b  # Works as intended.

print(a)

Answer 2

In [436]: a[True,True]=b
Traceback (most recent call last):
  File "<ipython-input-436-7c0409496970>", line 1, in <module>
    a[True,True]=b
ValueError: could not broadcast input array from shape (3,) into shape (1,)

Looks like this masked assignment is first converting b to an array:

In [437]: np.array(b)
<ipython-input-437-cbcacfae87b2>:1: VisibleDeprecationWarning: Creating an ndarray from ragged nested sequences (which is a list-or-tuple of lists-or-tuples-or ndarrays with different lengths or shapes) is deprecated. If you meant to do this, you must specify 'dtype=object' when creating the ndarray.
  np.array(b)
Traceback (most recent call last):
  File "<ipython-input-437-cbcacfae87b2>", line 1, in <module>
    np.array(b)
ValueError: could not broadcast input array from shape (3,) into shape (1,)

But (as experienced users know), trying to make an array from arrays that match in the first dimension (1), but not the second, raises this error.

Your in your first case the b arrays differ in the first axis, and np.array(b) produces a 2 element object dtype array.

These assignments don't, apparently, do the np.array(b) :

In [439]: a[:]=b
In [440]: a
Out[440]: array([array([[1., 1., 1.]]), array([[1., 1., 1., 1.]])], dtype=object)
In [441]: a[[0,1]]=b
In [442]: a
Out[442]: array([array([[1., 1., 1.]]), array([[1., 1., 1., 1.]])], dtype=object)

Assigning a 'proper' object dtype array does work:

In [447]: a[[True,True]]=a.copy()