Deep copy of a np.array of np.array
Beaten by one minute. Indeed, deepcopy is the answer here.
To your second question abut indexing: I have a feeling that you may be better off with a simple list or a dictionary-type data structure here. np.arrays make sense primarily if each array element is of the same type. Of course you can argue that each element in array_of_arrays is another array, but what is the benefit of having them collected in a numpy array instead of a simple list?
list_of_arrays = [np.arange(a*b).reshape(a,b) for (a, b) in pairs]
import numpy as np
import copy
pairs = [(2, 3), (3, 4), (4, 5)]
array_of_arrays = np.array([np.arange(a*b).reshape(a,b) for (a, b) in pairs])
a = copy.deepcopy(array_of_arrays)
Feel free to read up more about this here.
Oh, here is simplest test case:
a[0][0,0]
print a[0][0,0], array_of_arrays[0][0,0]
In [276]: array_of_arrays
Out[276]:
array([array([[0, 1, 2],
[3, 4, 5]]),
array([[ 0, 1, 2, 3],
[ 4, 5, 6, 7],
[ 8, 9, 10, 11]]),
array([[ 0, 1, 2, 3, 4],
[ 5, 6, 7, 8, 9],
[10, 11, 12, 13, 14],
[15, 16, 17, 18, 19]])], dtype=object)
array_of_arrays
is dtype=object
; that means each element of the array is a pointer to an object else where in memory. In this case those elements are arrays of different sizes.
a = array_of_arrays[:]
a
is a new array, but a view of array_of_arrays
; that is, it has the same data buffer (which in this case is list of pointers).
b = array_of_arrays[:][:]
this is just a view of a view. The second [:]
acts on the result of the first.
c = np.array(array_of_arrays, copy=True)
This is the same as array_of_arrays.copy()
. c
has a new data buffer, a copy of the originals
If I replace an element of c
, it will not affect array_of_arrays
:
c[0] = np.arange(3)
But if I modify an element of c
, it will modify the same element in array_of_arrays
- because they both point to the same array.
The same sort of thing applies to nested lists of lists. What array
adds is the view
case.
d = np.array([np.array(x, copy=True) for x in array_of_arrays])
In this case you are making copies of the individual elements. As others noted there is a deepcopy
function. It was designed for things like lists of lists, but works on arrays as well. It is basically doing what you do with d
; recursively working down the nesting tree.
In general, an object array is like list nesting. A few operations cross the object boundary, e.g.
array_of_arrays+1
but even this effectively is
np.array([x+1 for x in array_of_arrays])
One thing that a object array adds, compared to a list, is operations like reshape
. array_of_arrays.reshape(3,1)
makes it 2d; if it had 4 elements you could do array_of_arrays.reshape(2,2)
. Some times that's handy; other times it's a pain (it's harder to iterate).