How come not-copying a numpy array changes the data attribute?

问题

As my MWE below shows, calling np.array(a, copy=False) on an existing array a returns something that behaves exactly as expected, except that the .data attributes seem to differ. How can this be?

>>> a                               # My original array
array([2])
>>> b = np.array(a, copy=False)     # Not-a-copy of the original array
>>> b is a                          # The Python objects seem to be identical
True
>>> b.data is a.data                # But their .data attributes aren't??
False
>>> a.data
<memory at 0x7f82ebd757c8>
>>> b.data
<memory at 0x7f82ebd75888>
>>> b 
array([2])
>>> a
array([2])
>>> a[:] = 3                        # Changing a indeed also changes b
>>> a
array([3])
>>> b
array([3])
>>> a.data
<memory at 0x7f82ebd757c8>
>>> b.data
<memory at 0x7f82ebd75888>

EDIT

While playing around, I even found that the .data attribute changes while looking at it!

>>> a.data is a.data        # a.data isn't equal to itself?!
False
>>> a.data
<memory at 0x7f82ebd75948>
>>> a.data
<memory at 0x7f82ebd75888>  # A different value than a minute ago
>>> a.data
<memory at 0x7f82ebd75948>
>>> a.data
<memory at 0x7f82ebd75888>
>>> a.data
<memory at 0x7f82ebd75948>
>>> a.data
<memory at 0x7f82ebd75888>
>>> a.data
<memory at 0x7f82ebd75948>
>>> a.data
<memory at 0x7f82ebd75888>
>>> a.data
<memory at 0x7f82ebd75948>

回答1:

If you type:

help(a.data)

You will see that it does not return exactly what you expect:

class memoryview(object)
 |  memoryview(object)
 |  
 |  Create a new memoryview object which references the given object.
 |  
 |  Methods defined here:
 |

It creates a memoryview that references that object. If you want the memory address use id:

id(a) == id(b)
True

Note that:

id(a) is id(b)
False

Because id(a) returns an integer and id(b) too and their memory address isn't the same although their values are.

If you want hexadecimal memory:

hex(id(a))

Also, from NumPy doc: numpy.ndarray.data (I really don't know why this .data is useful but it exists)

回答2:

In [33]: a = np.array([2])                                                                             
In [34]: b = np.array(a, copy=False)

A nice human-readable way of checking for shared data buffer is the __array_interface__ dictionary.

In [36]: a.__array_interface__                                                                         
Out[36]: 
{'data': (69508768, False),
 'strides': None,
 'descr': [('', '<i8')],
 'typestr': '<i8',
 'shape': (1,),
 'version': 3}
In [37]: b.__array_interface__                                                                         
Out[37]: 
{'data': (69508768, False),
 'strides': None,
 'descr': [('', '<i8')],
 'typestr': '<i8',
 'shape': (1,),
 'version': 3}

a.data can be used to make a new array, but otherwise isn't very useful. And even this use is too low-level for most purposes:

In [44]: c = np.ndarray(shape=(1,1), dtype=int, buffer=a.data)                                         
In [45]: c                                                                                             
Out[45]: array([[2]])
In [46]: c.__array_interface__                                                                         
Out[46]: 
{'data': (69508768, False),
 'strides': None,
 'descr': [('', '<i8')],
 'typestr': '<i8',
 'shape': (1, 1),
 'version': 3}

来源：https://stackoverflow.com/questions/61102237/how-come-not-copying-a-numpy-array-changes-the-data-attribute

标签

python

arrays

numpy

numpy-ndarray