为什么 Numpy 不能从 Numpy Arrays 的列表中生成数组？

Question

I'm writing some code to group vectors by the angles between them. For example I might have an array of vectors:

vectors = np.array([[1, 0, 0], [1.1, 0, 0], [0, 2, 2]])

With an acceptable angle deviation of 0.1 radians for example. Currently, I'm doing this in a while loop like so:

groups = []
while not vectors.size == 0:
    vector = vectors[0]
    angles = (vectors @ vector)/(np.linalg.norm(vector, axis=1))
    angles = np.arccos(angles/np.linalg.norm(vector))
    group = vectors[angles <= angle]
    groups.append(group)
    vectors = vectors[angles > angle]
return np.array(groups)

I expect this to return a numpy array with the following form:

expected_array = np.array([[[1, 0, 0], [1.1, 0, 0]], [[0, 2, 2]]])

But instead I get the following:

actual_array = np.array([array([[1. , 0. , 0. ], [1.1, 0. , 0. ]]),
                         array([[0. , 2, 2]])])

Why doesn't Numpy notice that the list contains arrays and give me what I expect? Is there a way of making Numpy notice this? Or do you always have to use np.concatenate or something similar to get the desired result?

Answer 1

"I expect this to return a numpy array with the following form:"

In [420]: np.array([[[1, 0, 0], [1.1, 0, 0]], [[0, 2, 2]]])
<ipython-input-420-a1f3305ab5c3>: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([[[1, 0, 0], [1.1, 0, 0]], [[0, 2, 2]]])

Out[420]: array([list([[1, 0, 0], [1.1, 0, 0]]), list([[0, 2, 2]])], dtype=object)

Is that really what you expected? An array that preserves the nesting of the lists?

vstack (or concatenate ) can join the lists/arrays with the lists, to make a 2d array:

In [421]: np.vstack([[[1, 0, 0], [1.1, 0, 0]], [[0, 2, 2]]])
Out[421]: 
array([[1. , 0. , 0. ],
       [1.1, 0. , 0. ],
       [0. , 2. , 2. ]])

Converting those 2 arrays back to lists:

In [422]: _420.tolist()
Out[422]: [[[1, 0, 0], [1.1, 0, 0]], [[0, 2, 2]]]
In [423]: _421.tolist()
Out[423]: [[1.0, 0.0, 0.0], [1.1, 0.0, 0.0], [0.0, 2.0, 2.0]]

The first has 3 levels of nesting, same as the original; the second has only 2.

===

Your code isn't runnable:

In [424]: vectors = np.array([[1, 0, 0], [1.1, 0, 0], [0, 2, 2]])
In [425]: groups = []
     ...: while not vectors.size == 0:
     ...:     vector = vectors[0]
     ...:     angles = (vectors @ vector)/(np.linalg.norm(vector, axis=1))
     ...:     angles = np.arccos(angles/np.linalg.norm(vector))
     ...:     group = vectors[angles <= angle]
     ...:     groups.append(group)
     ...:     vectors = vectors[angles > angle]
     ...: 
Traceback (most recent call last):
  File "<ipython-input-425-e50fafbda1c3>", line 4, in <module>
    angles = (vectors @ vector)/(np.linalg.norm(vector, axis=1))
  File "<__array_function__ internals>", line 5, in norm
  File "/usr/local/lib/python3.8/dist-packages/numpy/linalg/linalg.py", line 2561, in norm
    return sqrt(add.reduce(s, axis=axis, keepdims=keepdims))
AxisError: axis 1 is out of bounds for array of dimension 1

I was hoping to see the list groups , before you tried to make an array from it. I don't feel like debugging your sample.

Answer 2

In short: It can! (But not in the way you want.)

A search around this topic finds this question: Numpy stack with unequal shapes

In which the following is made clear: numpy arrays must be rectangular , as the above example won't produce a rectangular array when added together this doesn't work.

If you try and do the same thing with arrays of the same size it will work . For example:

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

Will produce an array with the shape (2, 3).

However, with arrays of different sizes the numpy array dtype defaults to the object , and all the individual arrays get stored instead.