物理转置一个大的非方形 numpy 矩阵

Physically transposing a large non-square numpy matrix

有没有比 array.transpose.copy() 更快的物理转置大型 2D numpy 矩阵的方法?是否有任何例程可以有效地使用内存?

无论您做什么,都需要 O(n^2) 时间和记忆。我认为 .transpose.copy(用 C 语言编写)将是您应用程序最有效的选择。

编辑:假设您确实需要复制矩阵

可能值得看看转置的作用,这样我们就清楚您所说的 'physically tranposing' 是什么意思。

从一个小的 (4,3) 数组开始:

In [51]: arr = np.array([[1,2,3],[10,11,12],[22,23,24],[30,32,34]])             
In [52]: arr                                                                    
Out[52]: 
array([[ 1,  2,  3],
       [10, 11, 12],
       [22, 23, 24],
       [30, 32, 34]])

这是用一维数据缓冲区存储的,我们可以用ravel显示:

In [53]: arr.ravel()                                                            
Out[53]: array([ 1,  2,  3, 10, 11, 12, 22, 23, 24, 30, 32, 34])

strides 告诉它按 8 个字节步进列,按 24 (3*8) 步进行:

In [54]: arr.strides                                                            
Out[54]: (24, 8)

我们可以弄清楚“F”顺序 - 顺着顺序排列:

In [55]: arr.ravel(order='F')                                                   
Out[55]: array([ 1, 10, 22, 30,  2, 11, 23, 32,  3, 12, 24, 34])

虽然 [53] 是 view,但 [55] 是副本。

现在转置:

In [57]: arrt=arr.T                                                             
In [58]: arrt                                                                   
Out[58]: 
array([[ 1, 10, 22, 30],
       [ 2, 11, 23, 32],
       [ 3, 12, 24, 34]])

这个view;我们可以遍历 [53] 数据缓冲区,以 8 字节的步长向下移动行。使用 arrt 进行计算基本上与使用 arr 一样快。使用 strided 迭代,订单 'F' 与订单 'C'.

一样快
In [59]: arrt.strides                                                           
Out[59]: (8, 24)

原订单:

In [60]: arrt.ravel(order='F')                                                  
Out[60]: array([ 1,  2,  3, 10, 11, 12, 22, 23, 24, 30, 32, 34])

但是做 'C' ravel 会创建一个副本,与 [55]

相同
In [61]: arrt.ravel(order='C')                                                  
Out[61]: array([ 1, 10, 22, 30,  2, 11, 23, 32,  3, 12, 24, 34])

复制转置会生成一个按 'C' 顺序转置的数组。这是你的 'physical transpose':

In [62]: arrc = arrt.copy()                                                     
In [63]: arrc.strides                                                           
Out[63]: (32, 8)

像 [61] 那样重塑转置确实会制作一个副本,但通常我们不需要明确制作副本。我认为这样做的唯一原因是为了避免在以后的计算中出现多次冗余。

我假设如果行在内存中是连续的,并且您没有足够的可用内存来制作副本,那么您需要执行更有效地使用 CPU 缓存的逐行操作。

维基百科在 in-place matrix transposition 上有一篇文章。事实证明,这样的换位是不平凡的。这是那里描述的循环算法:

import numpy as np
from numba import njit

@njit # comment this line for debugging
def transpose_in_place(a):
    """In-place matrix transposition for a rectangular matrix.
    
     
    
    Parameter:
    
    - a: 2D array. Unless it's a square matrix, it will be scrambled
      in the process.
    
    Return:
        
    - transposed array, using the same in-memory data storage as the 
      input array.
      
    This algorithm is typically 10x slower than a.T.copy().
    Only use it if you are short on memory.
    """
    if a.shape == (1, 1):
        return a # special case

    n, m = a.shape
    
    # find max length L of permutation cycle by starting at a[0,1].
    # k is the index in the flat buffer; i, j are the indices in
    # a.
    L = 0
    k = 1
    while True:
        j = k % m
        i = k // m
        k = n*j + i
        L += 1
        if k == 1:
            break
    permut = np.zeros(L, dtype=np.int32)

    # Now do the permutations, one cycle at a time    
    seen = np.full(n*m, False)
    aflat = a.reshape(-1) # flat view
    
    for k0 in range(1, n*m-1):
        if seen[k0]:
            continue
        # construct cycle
        k = k0
        permut[0] = k0
        q = 1 # size of permutation array
        while True:
            seen[k] = True
            # note that this is slightly faster than the formula
            # on Wikipedia, k = n*k % (n*m-1)      
            i = k // m
            j = k - i*m            
            k = n*j + i
            if k == k0:
                break
            permut[q] = k
            q += 1
            
        # apply cyclic permutation
        tmp = aflat[permut[q-1]]
        aflat[permut[1:q]] = aflat[permut[:q-1]]
        aflat[permut[0]] = tmp            
    
    aT = aflat.reshape(m, n)
    return aT

def test_transpose(n, m):
    a = np.arange(n*m).reshape(n, m)
    aT = a.T.copy()
    assert np.all(transpose_in_place(a) == aT)

def roundtrip_inplace(a):
    a = transpose_in_place(a)
    a = transpose_in_place(a)
    
def roundtrip_copy(a):
    a = a.T.copy()
    a = a.T.copy()

if __name__ == '__main__':
    test_transpose(1, 1)
    test_transpose(3, 4)
    test_transpose(5, 5)
    test_transpose(1, 5)
    test_transpose(5, 1)
    test_transpose(19, 29)

即使我在这里使用 numba.njit 以便编译转置函数中的循环,它仍然比复制转置慢很多。

n, m = 1000, 10000
a_big = np.arange(n*m, dtype=np.float64).reshape(n, m)

%timeit -r2 -n10 roundtrip_copy(a_big)
54.5 ms ± 153 µs per loop (mean ± std. dev. of 2 runs, 10 loops each)

%timeit -r2 -n1 roundtrip_inplace(a_big)
614 ms ± 141 ms per loop (mean ± std. dev. of 2 runs, 1 loop each)