假设我有9个大小为N的数组(A,B,C,.. J).我想创建一个新的N 3×3矩阵数组,例如

matrices[i] = [[A[i], B[i], C[i]],
               [D[i], E[i], F[i]],
               [G[i], H[i], J[i]]]

一个简单的解决方案是将每个条目添加到for循环中的数组矩阵中,如下所示：

for i in range(len(matrices)):
    matrices[i] = [[A[i], B[i], C[i]],
            [D[i], E[i], F[i]],
            [G[i], H[i], J[i]]]

任何人都知道如何以更快的矢量化方式避免for循环的方法的提示？是否存在一些智能索引操作或其他操作.

解决方法:

一种方法是将那些堆积在具有np.column_stack的列中,并使用np.reshape进行整形-

np.column_stack((A,B,C,D,E,F,G,H,J)).reshape(-1,3,3)

众所周知,与np.concatenate的连接速度更快,因此将其与2D转置和重塑一起使用-

np.concatenate((A,B,C,D,E,F,G,H,J)).reshape(9,-1).T.reshape(-1,3,3)

另一个使用np.concatenate,3D转置和重塑-

np.concatenate((A,B,C,D,E,F,G,H,J)).reshape(3,3,-1).transpose(2,0,1)

运行时测试-

In [59]: # Setup input arrays
    ...: N = 1000
    ...: A = np.random.randint(0,9,(N,))
    ...: B = np.random.randint(0,9,(N,))
    ...: C = np.random.randint(0,9,(N,))
    ...: D = np.random.randint(0,9,(N,))
    ...: E = np.random.randint(0,9,(N,))
    ...: F = np.random.randint(0,9,(N,))
    ...: G = np.random.randint(0,9,(N,))
    ...: H = np.random.randint(0,9,(N,))
    ...: J = np.random.randint(0,9,(N,))
    ...: 

In [60]: %timeit np.column_stack((A,B,C,D,E,F,G,H,J)).reshape(-1,3,3)
10000 loops, best of 3: 84.4 µs per loop

In [61]: %timeit np.concatenate((A,B,C,D,E,F,G,H,J)).reshape(9,-1).T.reshape(-1,3,3)
100000 loops, best of 3: 15.8 µs per loop

In [62]: %timeit np.concatenate((A,B,C,D,E,F,G,H,J)).reshape(3,3,-1).transpose(2,0,1)
100000 loops, best of 3: 14.8 µs per loop