Python 加速Kronecker产品Numpy

所以我试图计算两个任意维矩阵的kronecker积。（我仅在示例中使用相同维度的方阵）

最初，我尝试使用

kron

：

a = np.random.random((60,60))
b = np.random.random((60,60))

start = time.time()
a = np.kron(a,b)
end = time.time()

Output: 0.160096406936645

为了提高速度，我使用了

tensordot

：

a = np.random.random((60,60))
b = np.random.random((60,60))

start = time.time()
a = np.tensordot(a,b,axes=0)
a = np.transpose(a,(0,2,1,3))
a = np.reshape(a,(3600,3600))
end = time.time()

Output: 0.11808371543884277

在搜索了一下网络之后，我发现（或者至少在我的理解中）numpy在需要重塑已被转置的张量时会产生一个额外的副本

于是我尝试了以下方法（这段代码显然没有给出a和b的kronecker乘积，但我只是做了一个测试）：

我的问题是：如何计算kronecker乘积而不遇到与转置相关的问题

我只是在寻找一个快速的加速，因此解决方案不必使用

tensordot

编辑

我刚刚在这篇文章中发现，还有另一种方法：

a = np.random.random((60,60))
b = np.random.random((60,60))

c = a

start = time.time()
a = a[:,np.newaxis,:,np.newaxis]
a = a[:,np.newaxis,:,np.newaxis]*b[np.newaxis,:,np.newaxis,:]
a.shape = (3600,3600)
end = time.time()

test = np.kron(c,b)
print(np.array_equal(a,test))
print(end-start)


Output: True
0.05503702163696289

---

我仍然感兴趣的问题是，您是否可以进一步加快计算速度？

einsum

似乎有效：

>>> a = np.random.random((60,60))
>>> b = np.random.random((60,60))
>>> ab = np.kron(a,b)
>>> abe = np.einsum('ik,jl', a, b).reshape(3600,3600)
>>> (abe==ab).all()
True
>>> timeit(lambda: np.kron(a, b), number=10)
1.0697475590277463
>>> timeit(lambda: np.einsum('ik,jl', a, b).reshape(3600,3600), number=10)
0.42500176999601535

简单的广播速度更快：

>>> abb = (a[:, None, :, None]*b[None, :, None, :]).reshape(3600,3600)
>>> (abb==ab).all()
True
>>> timeit(lambda:  (a[:, None, :, None]*b[None, :, None, :]).reshape(3600,3600), number=10)
0.28011218502069823

更新：使用blas和cython，我们可以获得另一个适度（30%）的加速。你自己决定是否值得费心

[setup.py]

from distutils.core import setup
from Cython.Build import cythonize

setup(name='kronecker',
      ext_modules=cythonize("cythkrn.pyx"))

[cythrn.pyx]

import cython
cimport scipy.linalg.cython_blas as blas
import numpy as np

@cython.boundscheck(False)
@cython.wraparound(False)
def kron(double[:, ::1] a, double[:, ::1] b):
    cdef int i = a.shape[0]
    cdef int j = a.shape[1]
    cdef int k = b.shape[0]
    cdef int l = b.shape[1]
    cdef int onei = 1
    cdef double oned = 1
    cdef int m, n
    result = np.zeros((i*k, j*l), float)
    cdef double[:, ::1] result_v = result
    for n in range(i):
        for m in range(k):
            blas.dger(&l, &j, &oned, &b[m, 0], &onei, &a[n, 0], &onei, &result_v[m+k*n, 0], &l)
    return result

要构建，请运行

cython cythrn.pyx

，然后运行

python3 setup.py build

>>> from timeit import timeit
>>> import cythkrn
>>> import numpy as np
>>> 
>>> a = np.random.random((60,60))
>>> b = np.random.random((60,60))
>>>
>>> np.all(cythkrn.kron(a, b)==np.kron(a, b))
True
>>> 
>>> timeit(lambda: cythkrn.kron(a, b), number=10)
0.18925874299020506

加速内存限制的计算

• 完全避免，是可能的（例如克朗和苏姆的例子）
• ，与其他计算相结合时
• 也许float32加上float64就足够了
• 如果此计算在循环中，则只分配一次内存

我在这段代码和@Paul Panzers实现中得到了完全相同的计时，但在这两个实现中我得到了相同的奇怪行为。使用预分配内存时，如果计算是并行的（这是预期的），则绝对不会有速度提升，但如果没有预分配内存，则会有相当大的速度提升

代码

import numba as nb
import numpy as np


@nb.njit(fastmath=True,parallel=True)
def kron(A,B):
    out=np.empty((A.shape[0],B.shape[0],A.shape[1],B.shape[1]),dtype=A.dtype)
    for i in nb.prange(A.shape[0]):
        for j in range(B.shape[0]):
            for k in range(A.shape[1]):
                for l in range(B.shape[1]):
                    out[i,j,k,l]=A[i,k]*B[j,l]
    return out

@nb.njit(fastmath=True,parallel=False)
def kron_preallocated(A,B,out):
    for i in nb.prange(A.shape[0]):
        for j in range(B.shape[0]):
            for k in range(A.shape[1]):
                for l in range(B.shape[1]):
                    out[i,j,k,l]=A[i,k]*B[j,l]
    return out

@nb.njit(fastmath=True,parallel=True)
def kron_and_sum(A,B):
    out=0.
    for i in nb.prange(A.shape[0]):
        TMP=np.float32(0.)
        for j in range(B.shape[0]):
            for k in range(A.shape[1]):
                for l in range(B.shape[1]):
                    out+=A[i,k]*B[j,l]
    return out

计时

#Create some data
out_float64=np.empty((a.shape[0],b.shape[0],a.shape[1],b.shape[1]),dtype=np.float64)
out_float32=np.empty((a.shape[0],b.shape[0],a.shape[1],b.shape[1]),dtype=np.float32)
a_float64 = np.random.random((60,60))
b_float64 = np.random.random((60,60))
a_float32 = a_float64.astype(np.float32)
b_float32 = b_float64.astype(np.float32)


#Reference
%timeit np.kron(a_float64,b_float64)
147 ms ± 1.22 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

#If you have to allocate memory for every calculation (float64)
%timeit B=kron(a_float64,b_float64).reshape(3600,3600)
17.6 ms ± 244 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

#If you don't have to allocate memory for every calculation (float64)
%timeit B=kron_preallocated(a_float64,b_float64,out_float64).reshape(3600,3600)
8.08 ms ± 269 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

#If you have to allocate memory for every calculation (float32)
%timeit B=kron(a_float32,b_float32).reshape(3600,3600)
9.27 ms ± 185 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

#If you don't have to allocate memory for every calculation (float32)
%timeit B=kron_preallocated(a_float32,b_float32,out_float32).reshape(3600,3600)
3.95 ms ± 155 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

#Example for a joined operation (sum of kroncker product)
#which isn't memory bottlenecked
%timeit B=kron_and_sum(a_float64,b_float64)
881 µs ± 104 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

---

谢谢你的解决方案。对于高维张量，我也看到了与tensordot收缩指数的转置相同的问题，但我看到einsum和广播没有比tensordot有所改进。这是意料之中的还是仍然有优势？@user1058860我不知道tensordot的内部工作原理，所以我无法回答这个问题。但我很好奇一个人能把它推多远，见更新的帖子。我想，blas/cython方法在不投入大量开发时间的情况下，速度是最快的，您是在哪个平台上获得这种时间安排的？在Windows（Core i5第8代）上，使用Cython实现，我得到预分配内存8ms单线程/8ms多线程，没有预分配内存，我得到大约40ms单线程/20ms多线程（几乎与我的Numba implementation完全相同）@max9111这是一台相当旧的x86_64 linux机器，我不认为它有超高速blas。你的观察结果令人费解，例如，我没有收到任何提前分配给家里写信的加速。也就是说，这可能是我计时的一个缺陷，因为我无法控制操作系统如何回收刚刚释放的内存。

#Create some data
out_float64=np.empty((a.shape[0],b.shape[0],a.shape[1],b.shape[1]),dtype=np.float64)
out_float32=np.empty((a.shape[0],b.shape[0],a.shape[1],b.shape[1]),dtype=np.float32)
a_float64 = np.random.random((60,60))
b_float64 = np.random.random((60,60))
a_float32 = a_float64.astype(np.float32)
b_float32 = b_float64.astype(np.float32)


#Reference
%timeit np.kron(a_float64,b_float64)
147 ms ± 1.22 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

#If you have to allocate memory for every calculation (float64)
%timeit B=kron(a_float64,b_float64).reshape(3600,3600)
17.6 ms ± 244 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

#If you don't have to allocate memory for every calculation (float64)
%timeit B=kron_preallocated(a_float64,b_float64,out_float64).reshape(3600,3600)
8.08 ms ± 269 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

#If you have to allocate memory for every calculation (float32)
%timeit B=kron(a_float32,b_float32).reshape(3600,3600)
9.27 ms ± 185 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

#If you don't have to allocate memory for every calculation (float32)
%timeit B=kron_preallocated(a_float32,b_float32,out_float32).reshape(3600,3600)
3.95 ms ± 155 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

#Example for a joined operation (sum of kroncker product)
#which isn't memory bottlenecked
%timeit B=kron_and_sum(a_float64,b_float64)
881 µs ± 104 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)