我想应用一个函数fn
,它基本上cosine distance
计算两个大的numpy形状数组(10000,100)和(5000,100)行,即我计算一个这些数组中每个行组合的值。
我的实施:
import math
def fn(v1,v2):
sumxx, sumxy, sumyy = 0, 0, 0
for i in range(len(v1)):
x = v1[i]; y = v2[i]
sumxx += x*x
sumyy += y*y
sumxy += x*y
return sumxy/math.sqrt(sumxx*sumyy)
val = []
for i in range(array1.shape[0]):
for j in range(array2.shape[0]):
val.append(fn(array1[i, :], array2[j, :]))
该功能非常快,只需几毫秒:
CPU times: user 4 ms, sys: 0 ns, total: 4 ms
Wall time: 1.24 ms
有没有有效的方法可以做到这一点?
答案 0 :(得分:2)
方法#1:我们可以简单地使用Scipy's cdist
及其cosine
距离功能 -
from scipy.spatial.distance import cdist
val_out = 1 - cdist(array1, array2, 'cosine')
方法#2:使用matrix-multiplication
-
def cosine_vectorized(array1, array2):
sumyy = (array2**2).sum(1)
sumxx = (array1**2).sum(1, keepdims=1)
sumxy = array1.dot(array2.T)
return (sumxy/np.sqrt(sumxx))/np.sqrt(sumyy)
方法#3:使用np.einsum
计算另一个的自平方和 -
def cosine_vectorized_v2(array1, array2):
sumyy = np.einsum('ij,ij->i',array2,array2)
sumxx = np.einsum('ij,ij->i',array1,array1)[:,None]
sumxy = array1.dot(array2.T)
return (sumxy/np.sqrt(sumxx))/np.sqrt(sumyy)
方法#4:引入numexpr
module卸载另一种方法的square-root
次计算 -
import numexpr as ne
def cosine_vectorized_v3(array1, array2):
sumyy = np.einsum('ij,ij->i',array2,array2)
sumxx = np.einsum('ij,ij->i',array1,array1)[:,None]
sumxy = array1.dot(array2.T)
sqrt_sumxx = ne.evaluate('sqrt(sumxx)')
sqrt_sumyy = ne.evaluate('sqrt(sumyy)')
return ne.evaluate('(sumxy/sqrt_sumxx)/sqrt_sumyy')
运行时测试
# Using same sizes as stated in the question
In [185]: array1 = np.random.rand(10000,100)
...: array2 = np.random.rand(5000,100)
...:
In [194]: %timeit 1 - cdist(array1, array2, 'cosine')
1 loops, best of 3: 366 ms per loop
In [195]: %timeit cosine_vectorized(array1, array2)
1 loops, best of 3: 287 ms per loop
In [196]: %timeit cosine_vectorized_v2(array1, array2)
1 loops, best of 3: 283 ms per loop
In [197]: %timeit cosine_vectorized_v3(array1, array2)
1 loops, best of 3: 217 ms per loop