我需要在numpy.array中找到唯一的行。

例如:

>>> a # I have
array([[1, 1, 1, 0, 0, 0],
       [0, 1, 1, 1, 0, 0],
       [0, 1, 1, 1, 0, 0],
       [1, 1, 1, 0, 0, 0],
       [1, 1, 1, 1, 1, 0]])
>>> new_a # I want to get to
array([[1, 1, 1, 0, 0, 0],
       [0, 1, 1, 1, 0, 0],
       [1, 1, 1, 1, 1, 0]])

我知道我可以创建一个集和循环数组,但我正在寻找一个有效的纯numpy解决方案。我相信有一种方法可以将数据类型设置为void,然后我可以使用numpy。唯一的,但我不知道如何让它工作。


当前回答

如果希望避免转换为一系列元组或其他类似数据结构的内存开销,可以利用numpy的结构化数组。

诀窍是将原始数组视为结构化数组,其中每个项对应于原始数组中的一行。这样就不需要复制,而且效率很高。

举个简单的例子:

import numpy as np

data = np.array([[1, 1, 1, 0, 0, 0],
                 [0, 1, 1, 1, 0, 0],
                 [0, 1, 1, 1, 0, 0],
                 [1, 1, 1, 0, 0, 0],
                 [1, 1, 1, 1, 1, 0]])

ncols = data.shape[1]
dtype = data.dtype.descr * ncols
struct = data.view(dtype)

uniq = np.unique(struct)
uniq = uniq.view(data.dtype).reshape(-1, ncols)
print uniq

要了解发生了什么,可以看看中间结果。

一旦我们将事物视为结构化数组,数组中的每个元素都是原始数组中的一行。(基本上,它是一个类似于元组列表的数据结构。)

In [71]: struct
Out[71]:
array([[(1, 1, 1, 0, 0, 0)],
       [(0, 1, 1, 1, 0, 0)],
       [(0, 1, 1, 1, 0, 0)],
       [(1, 1, 1, 0, 0, 0)],
       [(1, 1, 1, 1, 1, 0)]],
      dtype=[('f0', '<i8'), ('f1', '<i8'), ('f2', '<i8'), ('f3', '<i8'), ('f4', '<i8'), ('f5', '<i8')])

In [72]: struct[0]
Out[72]:
array([(1, 1, 1, 0, 0, 0)],
      dtype=[('f0', '<i8'), ('f1', '<i8'), ('f2', '<i8'), ('f3', '<i8'), ('f4', '<i8'), ('f5', '<i8')])

一旦我们运行numpy。唯一时,我们将得到一个结构化数组:

In [73]: np.unique(struct)
Out[73]:
array([(0, 1, 1, 1, 0, 0), (1, 1, 1, 0, 0, 0), (1, 1, 1, 1, 1, 0)],
      dtype=[('f0', '<i8'), ('f1', '<i8'), ('f2', '<i8'), ('f3', '<i8'), ('f4', '<i8'), ('f5', '<i8')])

然后我们需要将其视为“普通”数组(_存储在ipython中最后一次计算的结果,这就是为什么你会看到_.view…):

In [74]: _.view(data.dtype)
Out[74]: array([0, 1, 1, 1, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 0])

然后重新塑造成一个2D数组(-1是一个占位符,告诉numpy计算正确的行数,给出列数):

In [75]: _.reshape(-1, ncols)
Out[75]:
array([[0, 1, 1, 1, 0, 0],
       [1, 1, 1, 0, 0, 0],
       [1, 1, 1, 1, 1, 0]])

显然,如果你想更简洁,你可以这样写:

import numpy as np

def unique_rows(data):
    uniq = np.unique(data.view(data.dtype.descr * data.shape[1]))
    return uniq.view(data.dtype).reshape(-1, data.shape[1])

data = np.array([[1, 1, 1, 0, 0, 0],
                 [0, 1, 1, 1, 0, 0],
                 [0, 1, 1, 1, 0, 0],
                 [1, 1, 1, 0, 0, 0],
                 [1, 1, 1, 1, 1, 0]])
print unique_rows(data)

结果是:

[[0 1 1 1 0 0]
 [1 1 1 0 0 0]
 [1 1 1 1 1 0]]

其他回答

除了@Jaime出色的回答,另一种折叠一行的方法是使用a.strides[0](假设a是c -毗连的),它等于a.dtype.itemsize* a.c nshape[0]。此外,void(n)是dtype((void,n))的快捷方式。我们最终得到了这个最短的版本:

a[unique(a.view(void(a.strides[0])),1)[1]]

For

[[0 1 1 1 0 0]
 [1 1 1 0 0 0]
 [1 1 1 1 1 0]]

np。给出一个元组列表的唯一作品:

>>> np.unique([(1, 1), (2, 2), (3, 3), (4, 4), (2, 2)])
Out[9]: 
array([[1, 1],
       [2, 2],
       [3, 3],
       [4, 4]])

对于列表列表,它会引发TypeError: unhashable type: 'list'

如果希望避免转换为一系列元组或其他类似数据结构的内存开销,可以利用numpy的结构化数组。

诀窍是将原始数组视为结构化数组,其中每个项对应于原始数组中的一行。这样就不需要复制,而且效率很高。

举个简单的例子:

import numpy as np

data = np.array([[1, 1, 1, 0, 0, 0],
                 [0, 1, 1, 1, 0, 0],
                 [0, 1, 1, 1, 0, 0],
                 [1, 1, 1, 0, 0, 0],
                 [1, 1, 1, 1, 1, 0]])

ncols = data.shape[1]
dtype = data.dtype.descr * ncols
struct = data.view(dtype)

uniq = np.unique(struct)
uniq = uniq.view(data.dtype).reshape(-1, ncols)
print uniq

要了解发生了什么,可以看看中间结果。

一旦我们将事物视为结构化数组,数组中的每个元素都是原始数组中的一行。(基本上,它是一个类似于元组列表的数据结构。)

In [71]: struct
Out[71]:
array([[(1, 1, 1, 0, 0, 0)],
       [(0, 1, 1, 1, 0, 0)],
       [(0, 1, 1, 1, 0, 0)],
       [(1, 1, 1, 0, 0, 0)],
       [(1, 1, 1, 1, 1, 0)]],
      dtype=[('f0', '<i8'), ('f1', '<i8'), ('f2', '<i8'), ('f3', '<i8'), ('f4', '<i8'), ('f5', '<i8')])

In [72]: struct[0]
Out[72]:
array([(1, 1, 1, 0, 0, 0)],
      dtype=[('f0', '<i8'), ('f1', '<i8'), ('f2', '<i8'), ('f3', '<i8'), ('f4', '<i8'), ('f5', '<i8')])

一旦我们运行numpy。唯一时,我们将得到一个结构化数组:

In [73]: np.unique(struct)
Out[73]:
array([(0, 1, 1, 1, 0, 0), (1, 1, 1, 0, 0, 0), (1, 1, 1, 1, 1, 0)],
      dtype=[('f0', '<i8'), ('f1', '<i8'), ('f2', '<i8'), ('f3', '<i8'), ('f4', '<i8'), ('f5', '<i8')])

然后我们需要将其视为“普通”数组(_存储在ipython中最后一次计算的结果,这就是为什么你会看到_.view…):

In [74]: _.view(data.dtype)
Out[74]: array([0, 1, 1, 1, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, 0])

然后重新塑造成一个2D数组(-1是一个占位符,告诉numpy计算正确的行数,给出列数):

In [75]: _.reshape(-1, ncols)
Out[75]:
array([[0, 1, 1, 1, 0, 0],
       [1, 1, 1, 0, 0, 0],
       [1, 1, 1, 1, 1, 0]])

显然,如果你想更简洁,你可以这样写:

import numpy as np

def unique_rows(data):
    uniq = np.unique(data.view(data.dtype.descr * data.shape[1]))
    return uniq.view(data.dtype).reshape(-1, data.shape[1])

data = np.array([[1, 1, 1, 0, 0, 0],
                 [0, 1, 1, 1, 0, 0],
                 [0, 1, 1, 1, 0, 0],
                 [1, 1, 1, 0, 0, 0],
                 [1, 1, 1, 1, 1, 0]])
print unique_rows(data)

结果是:

[[0 1 1 1 0 0]
 [1 1 1 0 0 0]
 [1 1 1 1 1 0]]

让我们以列表的形式获取整个numpy矩阵,然后从这个列表中删除重复项,最后将我们唯一的列表返回到numpy矩阵中:

matrix_as_list=data.tolist() 
matrix_as_list:
[[1, 1, 1, 0, 0, 0], [0, 1, 1, 1, 0, 0], [0, 1, 1, 1, 0, 0], [1, 1, 1, 0, 0, 0], [1, 1, 1, 1, 1, 0]]

uniq_list=list()
uniq_list.append(matrix_as_list[0])

[uniq_list.append(item) for item in matrix_as_list if item not in uniq_list]

unique_matrix=np.array(uniq_list)
unique_matrix:
array([[1, 1, 1, 0, 0, 0],
       [0, 1, 1, 1, 0, 0],
       [1, 1, 1, 1, 1, 0]])

使用结构化数组的另一种选择是使用void类型的视图,将整行连接成单个项:

a = np.array([[1, 1, 1, 0, 0, 0],
              [0, 1, 1, 1, 0, 0],
              [0, 1, 1, 1, 0, 0],
              [1, 1, 1, 0, 0, 0],
              [1, 1, 1, 1, 1, 0]])

b = np.ascontiguousarray(a).view(np.dtype((np.void, a.dtype.itemsize * a.shape[1])))
_, idx = np.unique(b, return_index=True)

unique_a = a[idx]

>>> unique_a
array([[0, 1, 1, 1, 0, 0],
       [1, 1, 1, 0, 0, 0],
       [1, 1, 1, 1, 1, 0]])

编辑 添加np。as毗连数组遵循@seberg的建议。如果数组不是连续的,这将降低方法的速度。

编辑 以上内容可以稍微加快,但代价可能是不清楚,可以这样做:

unique_a = np.unique(b).view(a.dtype).reshape(-1, a.shape[1])

此外,至少在我的系统上,性能方面它是相同的,甚至更好,比lexsort方法:

a = np.random.randint(2, size=(10000, 6))

%timeit np.unique(a.view(np.dtype((np.void, a.dtype.itemsize*a.shape[1])))).view(a.dtype).reshape(-1, a.shape[1])
100 loops, best of 3: 3.17 ms per loop

%timeit ind = np.lexsort(a.T); a[np.concatenate(([True],np.any(a[ind[1:]]!=a[ind[:-1]],axis=1)))]
100 loops, best of 3: 5.93 ms per loop

a = np.random.randint(2, size=(10000, 100))

%timeit np.unique(a.view(np.dtype((np.void, a.dtype.itemsize*a.shape[1])))).view(a.dtype).reshape(-1, a.shape[1])
10 loops, best of 3: 29.9 ms per loop

%timeit ind = np.lexsort(a.T); a[np.concatenate(([True],np.any(a[ind[1:]]!=a[ind[:-1]],axis=1)))]
10 loops, best of 3: 116 ms per loop