numpy广播机制加速距离矩阵计算（numpy永远的神！）

由于楼主最近在打比赛，要对目标检测后的结果做分类，下游的分类任务需要与类别标签库中的图片做相似度计算，需要求距离矩阵，目标检测框出的图像特征reshape后是一个479432048的矩阵，总共有近1亿个元素，带查询库中的图像特征是310692048的矩阵，要求这两个超级大矩阵的距离矩阵。

首先想到的最朴素的思想就是for循环；尝试了之后发现需要耗费1200秒时间，但是比赛限制了inference时间要在30min以内，于是我想要加速

经过查询后得知，numpy的广播机制对于矩阵运算的效率提升十分明显，于是将for循环转换成numpy运算如下：

# -*-coding: utf-8-*-

'''
*Author     :Jianfeng Zhang 
*e-mail     :13052931019@163.com
*Blog       :https://me.csdn.net/qq_39004111
*Github     :https://github.com/JianfengZhang112358
*Data       :
*Description:
'''

import numpy as np
import time


def EuclideanDistances(A, B):
    BT = B.T
    vecProd = A * BT
    SqA =  A.getA()**2
    sumSqA = np.matrix(np.sum(SqA, axis=1))
    sumSqAEx = np.tile(sumSqA.transpose(), (1, vecProd.shape[1]))
    SqB = B.getA()**2
    sumSqB = np.sum(SqB, axis=1)
    sumSqBEx = np.tile(sumSqB, (vecProd.shape[0], 1))
    SqED = sumSqBEx + sumSqAEx - 2*vecProd
    ED = (SqED.getA())**0.5
    return np.matrix(ED)


a = np.array(
[[ 4, 9,  9, 12],
 [11,  8,  1, 3],
 [ 2,  6,  4, 5],
 [ 4,  7, 11, 96],
 [ 6,  7,  9, 58]])

b = np.array(
[[ 5,  8,  8, 78],
 [10,  5,  1, 12],
 [ 6,  6,  9, 54],
 [ 2,  1,  2, 8],
 [ 9,  1,  3, 12],
 [ 6,  1,  6, 45],
 [ 9, 10,  3, 33],
 [10,  4,  8, 7]]
)

#最low的for循环
results = np.zeros((len(a),len(b)))
s = time.time()
for i in range(len(a)):
    for j in range(len(b)):
        results[i,j] = np.sum((a[i]-b[j])**2,axis=-1)**0.5
e = time.time()
print('Time: %.6f s' % (e - s))


s = time.time()
c = np.sum((a[:,None] - b)**2, axis=-1)**.5  #这个是最快的(但是需要大内存)
e = time.time()
print('Time: %.6f s'%(e-s))


a = np.matrix(a)
b = np.matrix(b)
s = time.time()
c = EuclideanDistances(a,b)
e = time.time()
print('Time: %.6f s'%(e-s))

采用numpy的广播机制后，速度提升了两个数量级，只需要4.8秒，效果十分Nice。