Constructing a Python set from a Numpy matrix

I'm trying to execute the开发者_如何学Go following

>> from numpy import *
>> x = array([[3,2,3],[4,4,4]])
>> y = set(x)
TypeError: unhashable type: 'numpy.ndarray'

How can I easily and efficiently create a set with all the elements from the Numpy array?

If you want a set of the elements, here is another, probably faster way:

y = set(x.flatten())

PS: after performing comparisons between x.flat, x.flatten(), and x.ravel() on a 10x100 array, I found out that they all perform at about the same speed. For a 3x3 array, the fastest version is the iterator version:

y = set(x.flat)

which I would recommend because it is the less memory expensive version (it scales up well with the size of the array).

PPS: There is also a NumPy function that does something similar:

y = numpy.unique(x)

This does produce a NumPy array with the same element as set(x.flat), but as a NumPy array. This is very fast (almost 10 times faster), but if you need a set, then doing set(numpy.unique(x)) is a bit slower than the other procedures (building a set comes with a large overhead).

The immutable counterpart to an array is the tuple, hence, try convert the array of arrays into an array of tuples:

>> from numpy import *
>> x = array([[3,2,3],[4,4,4]])

>> x_hashable = map(tuple, x)

>> y = set(x_hashable)
set([(3, 2, 3), (4, 4, 4)])

The above answers work if you want to create a set out of the elements contained in an ndarray, but if you want to create a set of ndarray objects – or use ndarray objects as keys in a dictionary – then you'll have to provide a hashable wrapper for them. See the code below for a simple example:

from hashlib import sha1

from numpy import all, array, uint8


class hashable(object):
    r'''Hashable wrapper for ndarray objects.

        Instances of ndarray are not hashable, meaning they cannot be added to
        sets, nor used as keys in dictionaries. This is by design - ndarray
        objects are mutable, and therefore cannot reliably implement the
        __hash__() method.

        The hashable class allows a way around this limitation. It implements
        the required methods for hashable objects in terms of an encapsulated
        ndarray object. This can be either a copied instance (which is safer)
        or the original object (which requires the user to be careful enough
        not to modify it).
    '''
    def __init__(self, wrapped, tight=False):
        r'''Creates a new hashable object encapsulating an ndarray.

            wrapped
                The wrapped ndarray.

            tight
                Optional. If True, a copy of the input ndaray is created.
                Defaults to False.
        '''
        self.__tight = tight
        self.__wrapped = array(wrapped) if tight else wrapped
        self.__hash = int(sha1(wrapped.view(uint8)).hexdigest(), 16)

    def __eq__(self, other):
        return all(self.__wrapped == other.__wrapped)

    def __hash__(self):
        return self.__hash

    def unwrap(self):
        r'''Returns the encapsulated ndarray.

            If the wrapper is "tight", a copy of the encapsulated ndarray is
            returned. Otherwise, the encapsulated ndarray itself is returned.
        '''
        if self.__tight:
            return array(self.__wrapped)

        return self.__wrapped

Using the wrapper class is simple enough:

>>> from numpy import arange

>>> a = arange(0, 1024)
>>> d = {}
>>> d[a] = 'foo'
Traceback (most recent call last):
  File "<input>", line 1, in <module>
TypeError: unhashable type: 'numpy.ndarray'
>>> b = hashable(a)
>>> d[b] = 'bar'
>>> d[b]
'bar'

If you want a set of the elements:

>> y = set(e for r in x
             for e in r)
set([2, 3, 4])

For a set of the rows:

>> y = set(tuple(r) for r in x)
set([(3, 2, 3), (4, 4, 4)])

I liked xperroni's idea. But I think implementation can be simplified using direct inheritance from ndarray instead of wrapping it.

from hashlib import sha1
from numpy import ndarray, array

class HashableNdarray(ndarray):
    @classmethod
    def create(cls, array):
        return HashableNdarray(shape=array.shape, dtype=array.dtype, buffer=array.copy())

    def __hash__(self):
        if not hasattr(self, '_HashableNdarray__hash'):
            self.__hash = int(sha1(self.view()).hexdigest(), 16)
        return self.__hash

    def __eq__(self, other):
        if not isinstance(other, HashableNdarray):
            return super().__eq__(other)
        return super().__eq__(super(HashableNdarray, other)).all()

NumPy ndarray can be viewed as derived class and used as hashable object. view(ndarray) can be used for back transformation, but it is not even needed in most cases.

>>> a = array([1,2,3])
>>> b = array([2,3,4])
>>> c = array([1,2,3])
>>> s = set()

>>> s.add(a.view(HashableNdarray))
>>> s.add(b.view(HashableNdarray))
>>> s.add(c.view(HashableNdarray))
>>> print(s)
{HashableNdarray([2, 3, 4]), HashableNdarray([1, 2, 3])}
>>> d = next(iter(s))
>>> print(d == a)
[False False False]
>>> import ctypes
>>> print(d.ctypes.data_as(ctypes.POINTER(ctypes.c_double)))
<__main__.LP_c_double object at 0x7f99f4dbe488>

Adding to @Eric Lebigot and his great post.

The following did the trick for building a tensor lookup table:

a = np.array([[1, 0, 0], [1, 0, 0], [2, 3, 4]])
np.unique(a, axis=0)

output:

array([[1, 0, 0], [2, 3, 4]])

np.unique documentation