Python计算余弦相似度的方法及时间比较

形如有一个矩阵ashape形如（4,99999）和bshape形如(32,99999)，a的每一行去和b的每一行做余弦相似度计算，output一个shape形如(4，32)的余弦相似度矩阵。

a = np.random.random(size = (4,99999)).tolist()
b = np.random.random(size = (32,99999)).tolist()

from scipy import spatial
t1 = time.time()
cos_mat_sci = list(map(lambda x:list(map(lambda y:1-spatial.distance.cosine(x,y), b)),a))
t2 = time.time()
print('scipy time cost',t2-t1,'s')
#time cost 1.325453519821167 s

time_start=time.time()
cos_mat_sk = list(map(lambda x:list(map(lambda y:cosine_similarity([x,y])[0][1], b)),a))
time_end=time.time()
print('sklearn time cost',time_end-time_start,'s') 
#time cost 1.7672739028930664 s

def cos_sim(a, b):
    a_norm = np.linalg.norm(a)
    b_norm = np.linalg.norm(b)
    cos = np.dot(a,b)/(a_norm * b_norm)
    return cos

time_start=time.time()
cos_mat_np = list(map(lambda x:list(map(lambda y:cos_sim(x,y), b)),a))
time_end=time.time()
print('numpy time cost',time_end-time_start,'s') 
#time cost 3.1022114753723145 s

import torch
import torch.nn.functional as F
a_tf = torch.FloatTensor(a)
b_tf = torch.FloatTensor(b)
time_start=time.time()
cos_mat_torch = list(map(lambda x:list(map(lambda y:F.cosine_similarity(x,y, dim=0), b_tf)),a_tf))
time_end=time.time()
print('torch time cost',time_end-time_start,'s') 
# time cost 0.028922557830810547 s

scipy time cost 1.325453519821167 s
sklearn time cost 1.7672739028930664 s
numpy time cost 3.1022114753723145 s
torch time cost 0.028922557830810547 s
>>> cos_mat_sci[0][:3]
[0.7521347867015734, 0.7507867569813598, 0.7505274811256897]
>>> cos_mat_sk[0][:3]
[0.7521347867015696, 0.7507867569813551, 0.7505274811256867]
>>> cos_mat_np[0][:3]
[0.7521347867015734, 0.7507867569813594, 0.7505274811256898]
>>> cos_mat_torch[0][:3]
[tensor(0.7521), tensor(0.7508), tensor(0.7505)]

综合而言，torch用时最短，但在默认情况下其内部计算精度较低；综合而言scipy在相对要求精度的前提下是比较优秀的。np.linalg根据定义计算相对是最耗时的。