TransE代码理解_github - zulihit/transe-learn

TransE理解

论文地址：https://proceedings.neurips.cc/paper_files/paper/2013/hash/1cecc7a77928ca8133fa24680a88d2f9-Abstract.html
参考代码地址：https://github.com/zulihit/TransE

import codecs
import random
import json
import math
import numpy as np
import pandas as pd
import copy
import time

# 数据集地址
file_FB15k = "FB15k/"
1
2
3
4
5
6
7
8
9
10
11

entity2id = {}
relation2id = {}
1
2

数据加载

# 数据加载
def data_loader(file):
    file1 = file + "train.txt"
    file2 = file + "entity2id.txt"
    file3 = file + "relation2id.txt"

    with codecs.open(file2, 'r') as f1, codecs.open(file3, 'r') as f2:  # entity2id和relation2id
        lines1 = f1.readlines()  # readlines：作为列表返回文件中的所有行，其中每一行都是列表对象中的一项
        lines2 = f2.readlines()  # 也就是将每一行拿出来组成一个列表返回
        for line in lines1:
            line = line.strip().split('\t')  # 这一步是拆分出实体和编号    strip()：删除头尾的空格     \t：水平制表符相当于tab
            if len(line) != 2:  # 如果这一行不是两个元素，就放弃这一行
                continue
            entity2id[line[0]] = int(line[1])  # 转换成字典{实体：id}

        for line in lines2:  # 同理，制作关系和编号的字典
            line = line.strip().split('\t')
            if len(line) != 2:
                continue
            relation2id[line[0]] = int(line[1])

    entity_set = set()  # 用set去重
    relation_set = set()
    triple_list = []

    with codecs.open(file1, 'r') as f:  # 文件读尽量用codecs.open方法，一般不会出现编码的问题
        content = f.readlines()  # 读取训练集
        for line in content:
            triple = line.strip().split("\t")
            if len(triple) != 3:
                continue

            h_ = entity2id[triple[0]]  # 找到训练集中三元组的对应编号
            t_ = entity2id[triple[1]]
            r_ = relation2id[triple[2]]

            triple_list.append([h_, t_, r_])  # 储存三元组的编号

            entity_set.add(h_)  # 储存头实体的编号
            entity_set.add(t_)  # 储存尾实体的编号

            relation_set.add(r_)  # 储存关系的编号

    return entity_set, relation_set, triple_list  # 返回实体的编号集合，关系的编号集合，三元组的编号集合
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44

entity_set, relation_set, triple_list = data_loader(file_FB15k)
print("entity_set长度(里面是这样的entity_set: {0, 1, 2, 3,...14950}):",len(entity_set))
print("relation_set长度(里面是这样的relation_set: {0, 1, 2, 3,...1344}):",len(relation_set))
print("triple_list:",triple_list[2]) #这里需要注意的是[头实体，尾实体，关系]
1
2
3
4

entity_set长度(里面是这样的entity_set: {0, 1, 2, 3,...14950}): 14951
relation_set长度(里面是这样的relation_set: {0, 1, 2, 3,...}): 1345
triple_list: [7374, 13062, 648]
1
2
3

print("Complete load. entity : %d , relation : %d , triple : %d" % (
        len(entity_set), len(relation_set), len(triple_list)))
1
2

Complete load. entity : 14951 , relation : 1345 , triple : 483142
1

L1和L2

def distanceL2(h, r, t):
    # 为方便求梯度，去掉sqrt（sqrt是计算平方根的函数）
    return np.sum(np.square(h + r - t))


def distanceL1(h, r, t):
    return np.sum(np.fabs(h + r - t))  # fabs()方法返回数字的绝对值
1
2
3
4
5
6
7

关系向量&实体向量（嵌入）初始化

# 关系向量&实体向量（嵌入）初始化
def emb_initialize(entity_set,relation_set,embedding_dim):
    relation_dict = {}
    entity_dict = {}

    for relation in relation_set:
        # 初始化关系
        # np.random.uniform从一个均匀分布[low,high)中随机采样，注意定义域是左闭右开，即包含low，不包含high
        r_emb_temp = np.random.uniform(-6 / math.sqrt(embedding_dim),  # math.sqrt正平方根
                                        6 / math.sqrt(embedding_dim),
                                        embedding_dim)
        
        # 伪代码中关系向量初始化的办法（L2范数归一化）：关系向量除以它的L2范数，每个r_emb_temp代表一个关系的向量
        # 关于L2范数归一化：https://www.pianshen.com/article/9664381455/
        relation_dict[relation] = r_emb_temp / np.linalg.norm(r_emb_temp, ord=2)  # np.linalg.norm求范数，ord=2代表二范数
    
    
    for entity in entity_set:
        # 初始化实体
        e_emb_temp = np.random.uniform(-6 / math.sqrt(embedding_dim),
                                        6 / math.sqrt(embedding_dim),
                                        embedding_dim)
        entity_dict[entity] = e_emb_temp / np.linalg.norm(e_emb_temp, ord=2)

    relation = relation_dict  # 初始化后的关系  {关系编号：L2范数归一化后的向量}
    entity = entity_dict  # 初始化后的实体   {实体编号：L2范数归一化后的向量}
    print("关系向量  relation[0]：\n",relation[0] ) # 返回一个关系字典
    print("实体向量  entity[0]：\n",entity[0]) 
    return relation,entity



relation,entity=emb_initialize(entity_set,relation_set,50)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33

关系向量  relation[0]：
 [-0.06124931  0.05381147 -0.0581685   0.00965697  0.02497375  0.21745009
  0.00957725 -0.11250099  0.0888398   0.00656037  0.16423882  0.09413429
  0.22683678 -0.22823347  0.03280322  0.03559384 -0.04643496  0.25829363
 -0.16438011  0.16169445  0.04193395  0.01949435  0.00759052  0.16294735
 -0.15454095  0.22251406 -0.16183664 -0.0116275  -0.22146209 -0.15569882
 -0.26301086  0.0739365   0.15301796  0.02804698  0.09017532 -0.19006649
  0.15857319 -0.20143701  0.0257449   0.22238414 -0.14402593  0.07735178
 -0.16285992  0.11581199 -0.04704164 -0.0047954  -0.02833891  0.1342455
  0.21679655  0.2492022 ]
实体向量  entity[0]：
 [-0.04303367  0.02916706 -0.02495438 -0.04141798  0.1203754  -0.01869183
 -0.16110016  0.10667474  0.20629609  0.09499036 -0.17413892  0.22841423
  0.14314854 -0.06291098  0.23317245 -0.07646504 -0.09352154  0.22610284
  0.03655919 -0.22014374  0.22399264 -0.17550837 -0.19925698  0.21728113
  0.18104154  0.13808154 -0.11072903 -0.1916764  -0.01511032  0.08146982
  0.22693272  0.09733503  0.09761     0.18111636  0.00903148 -0.20821903
 -0.17265452  0.00196428 -0.11702929  0.0885079   0.16693078  0.14552579
 -0.08605191 -0.06132584 -0.20229443  0.04791079  0.02205383  0.12145553
 -0.02652395  0.1460217 ]
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

由此可以看出，这里是把实体（关系）用一个50维的嵌入向量表示，一个思路就是嵌入维度越大，所能表示的特征越多。

负样本

def Corrupt(entity, triple):  # 随机替换头实体和尾实体制作负样本
    corrupted_triple = copy.deepcopy(triple)  # deepcopy：将被复制对象完全再复制一遍作为独立的新个体单独存在
    seed = random.random()
    if seed > 0.5:
        # 替换head
        head = triple[0]
        rand_head = head
        while (rand_head == head):
            rand_head = random.randint(0, len(entity) - 1)  # 从实体集合随机选择一个实体作为头实体替换
        corrupted_triple[0] = rand_head

    else:
        # 替换tail
        tail = triple[1]
        rand_tail = tail
        while (rand_tail == tail):
            rand_tail = random.randint(0, len(entity) - 1)  # 从实体集合随机选择一个实体作为尾实体替换
        corrupted_triple[1] = rand_tail
    return corrupted_triple

print("三元组：",triple_list[3])
print("负三元组(随机替换头尾实体)：",Corrupt(entity,triple_list[3])) #这里需要注意的是[头实体，尾实体，关系]


1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

三元组： [11436, 7445, 143]
负三元组(随机替换头尾实体)： [11436, 10061, 143]
1
2

更新实体向量（嵌入）

# 损失函数
def hinge_loss(dist_correct, dist_corrupt):  # 损失函数
    return max(0, dist_correct - dist_corrupt + 1) #这里设置margin为1，正确三元组的L1（L2）范数-错误三元组的L1（L2）范数
1
2
3

def update_embeddings(entity,relation, Tbatch,learning_rate):
    # 这里深拷贝了整个字典,会造成训练速度较慢，但是是易于理解的，增加训练速度的版本是transE_speed
    copy_entity = copy.deepcopy(entity)
    copy_relation = copy.deepcopy(relation)
    loss=0
    n=0 #用于提前退出，因为这个只是用于展示
    for triple, corrupted_triple in Tbatch:
        # 取copy里的vector累积更新
        h_correct_update = copy_entity[triple[0]]  # 取正确三元组的头实体的（初始化）向量
        t_correct_update = copy_entity[triple[1]]  # 取正确三元组的尾实体的（初始化）向量
        relation_update = copy_relation[triple[2]]  # 取正确三元组的关系的（初始化）向量
        
        h_corrupt_update = copy_entity[corrupted_triple[0]]  # 取错误三元组的头实体的(初始化)向量
        t_corrupt_update = copy_entity[corrupted_triple[1]]  # 取错误三元组的尾实体的(初始化)向量

        # 取原始的vector计算梯度
        h_correct = entity[triple[0]]  # 正确三元组的头实体向量
        t_correct = entity[triple[1]]  # 正确三元组的尾实体向量
        relation_temp = relation[triple[2]]  # 正确三元组的关系向量
        h_corrupt = entity[corrupted_triple[0]]  # 错误三元组的头实体向量
        t_corrupt = entity[corrupted_triple[1]]  # 错误三元组的尾实体向量
        print("正确三元组：",triple[0],triple[2],triple[1])
        print("错误三元组：",corrupted_triple[0],triple[2],corrupted_triple[1])
        if True:  # 选择l1还是l2范数计算距离这是一个超参数,假设这里选择L1范数
            dist_correct = distanceL1(h_correct, relation_temp, t_correct)  # 计算正确三元组的L1范数
            dist_corrupt = distanceL1(h_corrupt, relation_temp, t_corrupt)  # 计算错误三元组的L1范数
            print("正确三元组的L1范数:",dist_correct)
            print("错误三元组的L1范数:",dist_corrupt)
        #else:
        #    dist_correct = distanceL2(h_correct, relation, t_correct)  # 计算正确三元组的L2范数
        #    dist_corrupt = distanceL2(h_corrupt, relation, t_corrupt)  # 计算错误三元组的L2范数

        err = hinge_loss(dist_correct, dist_corrupt)  # 计算三元组的距离函数，也是原文中的距离函数，目标是使正确三元组和错误三元组区分开
        print("三元组的距离: ",err)
        if err > 0:  # 误差只可能是大于0和等于0,等于0时不能求导
            loss += err
            print("loss:",loss)
            # 关于L1范数的求导方法：参考了刘知远组实现中的实现，是先对L2范数求导，逐元素判断正负，为正赋值为1，负则为-1。
            grad_pos = 2 * (h_correct + relation_temp - t_correct)  # 正确三元组对L2范数求导的结果
            grad_neg = 2 * (h_corrupt + relation_temp - t_corrupt)  # 错误三元组对L2范数求导的结果
            print("正确三元组的梯度：\n",grad_pos)
            print("错误三元组的梯度：\n",grad_neg)
            if True:
                for i in range(len(grad_pos)):
                    if (grad_pos[i] > 0):
                        grad_pos[i] = 1
                    else:
                        grad_pos[i] = -1

                for i in range(len(grad_neg)):
                    if (grad_neg[i] > 0):
                        grad_neg[i] = 1
                    else:
                        grad_neg[i] = -1
            '''
            update_embeddings函数中，要对correct triplet和corrupted triplet都进行更新。
            虽然写作$(h,l,t)$和$(h',l,t')$，但两个三元组只有一个entity不同（前面说了，不同时替换头尾实体），
            所以在每步更新时重叠实体需要更新两次（和更新relation一样）。
            例如正确的三元组是（1，2，3），错误的是（1，2，4），那么1和2都需要更新两次
            '''
            # 更新正确三元组的向量，更新的过程也就是梯度下降的过程
            # （h+r-t）head系数为正，减梯度；tail系数为负，加梯度
            # 基于正确三元组更新向量
            h_correct_update -= learning_rate * grad_pos
            t_correct_update -= (-1) * learning_rate * grad_pos
            relation_update -= learning_rate * grad_pos  # 正确三元组的时候relation前面的符号是正号
            print("更新正确三元组向量(头实体)：\n",h_correct_update)
            print("更新正确三元组向量(尾实体)：\n",t_correct_update)
            print("更新正确三元组向量(关系)：\n",relation_update)
            # corrupt项整体为负，因此符号与correct相反
            # 基于错误三元组更新向量
            if triple[0] == corrupted_triple[0]:  # 若替换的是尾实体，则头实体更新两次
                h_correct_update -= (-1) * learning_rate * grad_neg
                t_corrupt_update -= learning_rate * grad_neg
                relation_update -= (-1) * learning_rate * grad_neg  # 错误三元组的时候relation前面的符号是负号
            elif triple[1] == corrupted_triple[1]:  # 若替换的是头实体，则尾实体更新两次
                h_corrupt_update -= (-1) * learning_rate * grad_neg
                t_correct_update -= learning_rate * grad_neg
                relation_update -= (-1) * learning_rate * grad_neg  # 错误三元组的时候relation前面的符号是负号
            #print("更新错误三元组向量(头实体)：\n",h_correct_update)
        print("==================================")
        n=n+1
        if n==5: #只循环5次，用于展示
            break
    # batch norm，更新完一个epoch重新归一化一下
    for i in copy_entity.keys():  # copy_entity.keys代表实体的编号
        copy_entity[i] /= np.linalg.norm(copy_entity[i])  # l1范数归一化，这里好像跟原文不符，原文每个epoch前貌似用的l2范数归一化，而且只针对实体
    for i in copy_relation.keys():
        copy_relation[i] /= np.linalg.norm(copy_relation[i])  # l1范数归一化
    # 达到批量更新的目的
    entity = copy_entity
    relation = copy_relation
    return loss
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93

# 测试一下
nbatches = 400  # 根据batchsize计算的，看有多少个batch
batch_size = len(triple_list) // nbatches  # 根据三元组数量来判断batch_size
print("batch size: ", batch_size)
Sbatch = random.sample(triple_list, batch_size)  # 取一个batch的三元组样本
Tbatch = []  # 负样本
# 每个triple选3个负样例
for triple in Sbatch:  # 创造负样本
    corrupted_triple = Corrupt(entity,triple)
    Tbatch.append((triple, corrupted_triple))  # 包含正样本和负样本的列表
selfloss=update_embeddings(entity,relation,Tbatch,0.01)  # 更新向量
1
2
3
4
5
6
7
8
9
10
11

batch size:  1207
正确三元组： 3559 520 13032
错误三元组： 11237 520 13032
正确三元组的L1范数: 9.160141048180064
错误三元组的L1范数: 9.834198022289906
三元组的距离:  0.3259430258901581
loss: 0.3259430258901581
正确三元组的梯度：
 [-0.37901602  0.35999408  0.25189135 -0.48776021 -0.34353277  0.13417585
  0.12309854  0.55943398 -0.40032831 -0.6257741  -0.82395044 -0.64565296
 -0.13032958 -0.42676876  0.06002297 -0.11228136 -0.39317566  0.08805352
  0.13574403  0.37024427  0.39853769 -0.83249792  0.34527112  0.31937902
  0.03344285  0.68976952 -0.41013865  0.21129952 -0.24349778  0.83943312
  0.01960911 -0.72258477  0.17864673  0.82505802 -0.21363834  0.3877319
 -0.18607471 -0.38540827  1.21557675  0.88342087 -0.25167498  0.54851887
  0.01164001 -0.69430878  0.30424088 -0.04292894  0.12082423  0.10352803
  0.04737885 -0.00299312]
错误三元组的梯度：
 [-1.21288997  0.0426023   0.07324305 -0.45636903 -0.11113273  0.17646589
  0.61759574  0.83078712  0.2223841  -0.64670274 -1.24342105 -0.76046952
 -0.21324672 -0.13913355 -0.19443816  0.31559439 -0.55318034 -0.28676876
 -0.1359051  -0.21112282  0.04126087 -0.75024583 -0.26448353  0.63174885
  0.36455667  0.119203    0.06355581  0.17073111 -0.08695611  0.35151922
 -0.84224382 -0.43388996  0.07136001  0.53999022 -0.45768278  0.51899856
 -0.48669738 -0.24570011  1.10199662  0.45988928 -0.54091843  0.40703835
  0.03968196 -0.22040216 -0.46371491 -0.01390717  0.61497178  0.22873204
 -0.36592812  0.32693831]
更新正确三元组向量(头实体)：
 [ 0.2297027   0.06617272 -0.02236433  0.03971393 -0.0183397   0.11760258
 -0.14633861 -0.06802978 -0.06738895 -0.14294786 -0.00737368 -0.16590576
  0.14661019 -0.12888502  0.20840432 -0.22658516  0.19490931 -0.07120115
  0.03070406  0.19483447  0.12137166 -0.23538472  0.12832474 -0.1298464
 -0.17738364  0.10871538 -0.03781839  0.11963854 -0.11801649  0.19612825
  0.13535835 -0.10416392 -0.19844118  0.23489312  0.09573639  0.12432242
  0.06108367 -0.13226323  0.1536877   0.11574505  0.01557258  0.17376242
  0.07849308 -0.09953866  0.17191859 -0.20786826  0.0094432  -0.07974221
  0.13303983 -0.0930276 ]
更新正确三元组向量(尾实体)：
 [ 0.21671441  0.01611362  0.00714604  0.04116654  0.19156749  0.05765291
 -0.18083127 -0.15081217 -0.1026865   0.07931343  0.14264843  0.20150768
  0.0923356   0.12387304  0.18558506 -0.17221731  0.15485868 -0.18112329
 -0.07792114 -0.05432951 -0.16490448  0.18528051  0.04996853 -0.05508658
 -0.03808002 -0.19579378 -0.00197225 -0.12294335  0.12599166 -0.20157153
 -0.0945553   0.09187188 -0.06727104 -0.16573768 -0.04166313 -0.03553562
 -0.11778666 -0.13913391 -0.20343586 -0.17693545  0.1756986   0.04243341
  0.21554448  0.18772311 -0.17024577 -0.03645544  0.21043763 -0.15602006
  0.05864182 -0.07796291]
更新正确三元组向量(关系)：
 [-0.1724963   0.09993794  0.12545604 -0.21242749  0.06814081 -0.02286175
 -0.00294339  0.16693461 -0.20546171 -0.06062576 -0.23195311  0.07458696
 -0.08943938  0.06937367 -0.02280778  0.02822718 -0.20663845 -0.09589538
 -0.07075319 -0.09404184 -0.11700729  0.03441627  0.06427935  0.20444933
  0.12602504  0.0103756  -0.13922319 -0.16693213  0.15225926 -0.00798322
 -0.25010909 -0.13525658  0.19049351 -0.01810178 -0.21421869  0.0040079
 -0.24190768 -0.16957481  0.22066482  0.11902993  0.06428854  0.11293043
  0.11287141 -0.02989262 -0.22004393  0.17994835  0.23140655 -0.05451384
 -0.08070859  0.04356814]
==================================
正确三元组： 12489 591 3764
错误三元组： 3432 591 3764
正确三元组的L1范数: 10.072297497304145
错误三元组的L1范数: 8.895223316435523
三元组的距离:  2.1770741808686225
loss: 2.5030172067587806
正确三元组的梯度：
 [-0.20675598 -0.83125079  0.40518409 -0.27664591 -0.23839436 -0.04849558
  0.10270012  0.03573358 -0.28901653 -0.54806713  0.52595571  0.19228698
 -0.28891565 -0.34848291  0.60373705 -0.38225062  0.18428005 -0.09909495
  0.48397984  0.33237687  0.39685326  0.26258934  0.20324456 -0.03127339
 -0.08698592 -1.06777599 -0.28804742  0.9433875   0.85814169 -0.38462137
  0.52426569 -1.07268463  0.06561395 -0.1351267   0.85691044  0.80275474
  0.32775847  0.8533333  -0.60398975 -0.62201439  0.3411744   0.08062966
  0.43420765 -0.52652602 -0.06032774 -0.01874879 -0.38629008 -0.37470968
 -0.3143244   0.79667938]
错误三元组的梯度：
 [ 0.57900001 -0.69060333 -0.27781468  0.29042609 -0.41242099 -0.33514097
 -0.49171434 -0.438654    0.08553278 -0.34020408 -0.03004907  0.23082889
  0.44902808 -0.5231981   0.70284469 -0.44376323  0.00236497  0.1680438
  0.03400182  0.96799707 -0.11136612  0.12276734 -0.14667088 -0.05740552
 -0.53705433 -0.39941882 -0.06872949  0.39747502  0.91435588 -0.12735192
  0.01999781 -0.39984922  0.11006548 -0.59624287  0.52324563  0.17826109
  0.54646694  0.85043908 -0.26478771 -0.60848199  0.55856062  0.24023241
  0.27502472  0.38001486  0.38453466  0.05923525 -0.23232777  0.3212445
  0.46356412  0.40164359]
更新正确三元组向量(头实体)：
 [-0.12991586 -0.16440062  0.16125645 -0.06963659  0.00588732 -0.06145422
  0.15302673  0.14754389 -0.17820692 -0.17379509  0.15935649  0.06423061
 -0.19406064  0.09327949 -0.13462955  0.08170153 -0.02968334 -0.2178587
  0.01878297 -0.08934636  0.05930044  0.16612617  0.18147629 -0.05369676
  0.02258476 -0.09889161 -0.04378755  0.21309609 -0.0055314   0.10504572
  0.02104224 -0.10539113  0.00688008  0.19415205  0.09957171  0.14861912
 -0.0955123  -0.02709809  0.08867102 -0.20362468 -0.08499755  0.00741165
  0.18733285 -0.22504347 -0.22273776 -0.18459085 -0.13170815 -0.21553465
 -0.16820369  0.17852849]
更新正确三元组向量(尾实体)：
 [ 0.0050869   0.08520743 -0.0782068   0.02720928  0.01639467  0.1189452
  0.12811563  0.0370657   0.01536301  0.11157175 -0.125298    0.24521036
 -0.02412924  0.11570441 -0.17468522  0.08811775 -0.10033296 -0.19931802
 -0.14839989 -0.00535774 -0.16605157  0.0905454   0.20020356  0.00116181
  0.11966507  0.23274371  0.19135325 -0.10567917 -0.18644574  0.09486631
 -0.19564498  0.21459862 -0.17647544  0.15895447 -0.16585566 -0.08354602
 -0.09783547 -0.17778807  0.14655853  0.03630079 -0.09895671 -0.09075587
  0.07575139  0.17568436 -0.01532985 -0.23376763  0.10097763  0.17765354
  0.20537207  0.02594474]
更新正确三元组向量(关系)：
 [ 0.06162476 -0.13601735 -0.0668712  -0.01147708 -0.07868983  0.18615163
 -0.00356104 -0.1226114   0.07906167  0.04133327 -0.05167663  0.24712324
  0.05547357 -0.12181653  0.23181285 -0.15470908 -0.00850959 -0.00100679
  0.04480705  0.22017705 -0.05692538  0.0257139   0.09034955  0.06922188
  0.08358735 -0.17225267  0.12111709  0.12291848  0.2181565  -0.1724901
  0.01544562 -0.18635256 -0.18054854 -0.07276093  0.13302785  0.13921223
  0.13155607  0.24597667 -0.21410737 -0.04108172  0.12662804 -0.08785269
  0.07552237  0.16746481  0.20724404 -0.02855117  0.06954074  0.23583335
  0.24641356  0.21575595]
==================================
正确三元组： 5971 242 1937
错误三元组： 5971 242 12379
正确三元组的L1范数: 10.838409158700102
错误三元组的L1范数: 10.324248809616375
三元组的距离:  1.5141603490837277
loss: 4.017177555842508
正确三元组的梯度：
 [-0.2583954   0.75117153 -0.70019031  0.37722074 -0.22390412 -0.37128223
 -0.37664392 -0.01448126 -0.36556971  0.67554285  0.36036681  0.58679804
 -0.56599424  0.79970803  0.75649744  0.77286221 -0.26326082  0.14947996
 -0.67023404  0.02380669 -0.76563816  0.7129515   0.94371575  0.15692248
 -0.05732602  0.04239009 -0.1545963  -0.17958389  0.50802604 -0.48672663
 -0.55946516 -0.60719591 -0.48154064 -0.06867678  0.07234063  0.15862991
 -0.35610727 -0.00660994  0.26682611  0.14069412 -0.46288159  0.37630276
 -0.9899526   0.45630982 -1.08016621  0.20687556 -0.2692745  -0.53531404
 -0.43873147 -1.07163607]
错误三元组的梯度：
 [-0.73952717  0.90640584 -0.59234038  0.13835401 -0.02106316 -0.23818818
 -0.20512517  0.41712066 -1.08411331  0.53046497  0.46595905 -0.04045682
 -0.51185347  0.89064319  0.97088536  0.09680112 -0.213252    0.43408414
 -0.47290427 -0.2345759   0.02207074  0.43654382  0.07963893 -0.64194125
 -0.36320662 -0.08258446 -0.7346493   0.04089253 -0.08650893 -0.21946392
 -0.97173921 -0.02592454 -0.2536815   0.14912094  0.63431087  0.63769143
 -0.58913114  0.5826085   0.78669324 -0.33504835 -0.10750919  0.73398359
 -0.32363269  0.28295547 -0.98100014  0.24760741 -0.36813614 -0.25353888
 -0.30583286 -0.1667329 ]
更新正确三元组向量(头实体)：
 [-0.08934616 -0.0456189  -0.04346564  0.06966536 -0.03918103 -0.11917562
 -0.19253371  0.05937545 -0.22708118  0.0902964   0.23607142 -0.0409323
  0.05726446  0.10340646  0.23561154  0.0353324   0.080485    0.16431726
 -0.177941   -0.14737261 -0.16632585  0.04386661  0.09845879 -0.17417872
 -0.24173488  0.0009411  -0.14134846 -0.21125681  0.1364854  -0.21538004
 -0.19120584  0.08871361  0.04766718 -0.00025555 -0.02630896  0.04873789
 -0.09395941  0.20110266  0.17984689  0.08842095  0.06703177  0.21846395
 -0.18744731  0.20221225 -0.15945802 -0.12049582  0.06079068 -0.07816313
  0.00929007 -0.0691089 ]
更新正确三元组向量(尾实体)：
 [-0.12878045 -0.13286137  0.1605667   0.05689098  0.18539835  0.19010479
 -0.08320368  0.05279641 -0.18501312 -0.09568796  0.11891911 -0.19168768
  0.17982601 -0.04070374  0.03453008 -0.10578747 -0.00246913  0.18461503
  0.19290356 -0.17485211  0.17971106 -0.11025855 -0.20289303 -0.18443148
  0.01375137 -0.10042602 -0.21482027 -0.05401662 -0.12962625 -0.04619295
 -0.07648262  0.17444274  0.20365959  0.1535966   0.19770323  0.00588395
  0.08088917  0.19095797  0.09442052 -0.08938368  0.16257912  0.0060258
  0.15515716  0.12779967  0.15341584 -0.07033769  0.00725963  0.14077621
 -0.02993194  0.19856707]
更新正确三元组向量(关系)：
 [-0.13863199  0.25834329 -0.11606282  0.14583599  0.14262733  0.15363929
 -0.04899193  0.01618034 -0.1107168   0.12178707  0.03303109  0.11264364
 -0.13043557  0.22574382  0.14716726  0.21531124 -0.18458455  0.06503775
  0.06572753 -0.04557615 -0.00678217  0.17235059  0.14050606  0.03820848
  0.25682324 -0.11017207 -0.12076996  0.09744824 -0.04209863 -0.04417623
 -0.13500936 -0.18786882 -0.05477791  0.14951375  0.2301825   0.00646102
  0.02679495  0.01655034  0.01798669 -0.13745758 -0.10589344 -0.05428676
 -0.12237182  0.12374234 -0.19720924  0.12359591 -0.1581683  -0.01871768
 -0.22858774 -0.23814207]
==================================
正确三元组： 13591 1291 12733
错误三元组： 13053 1291 12733
正确三元组的L1范数: 10.301110197065395
错误三元组的L1范数: 10.956140962774905
三元组的距离:  0.34496923429049
loss: 4.362146790132998
正确三元组的梯度：
 [-0.16795658 -0.12949096  0.86120369 -0.63356445 -0.43294176  0.1614915
  0.72136151  0.09835598 -0.02436859  0.65419805  0.89448053  0.27997895
 -0.38054446  0.00558313 -0.07562275 -0.42429659  0.14469542 -0.23622529
  0.50427403  0.65682846  0.18051775 -0.23145447 -0.30247826 -0.16453314
  0.45272069 -0.25732558  0.35740638 -1.05463174 -0.51151366 -0.51385635
 -0.24108557  0.10228883 -0.70497518  0.5301379  -0.26027818  0.42738776
  0.01525057  0.31995543 -0.57657569  0.27153534  0.54535743  0.87142385
 -0.89743388  0.52178093 -0.39267808  1.03495809  0.10653143  0.09360188
 -0.47958875 -0.69549495]
错误三元组的梯度：
 [-0.67771651  0.12805868  0.72408947 -0.19640587 -0.74853482 -0.34909758
  0.44424789 -0.12377116  0.1842216   0.376691    0.37375837  0.54991684
 -0.03252688 -0.01503632  0.63432943 -0.60377107 -0.15897948 -0.04480671
  0.45616743  0.53679793  0.964548    0.2923565   0.26962191 -0.25070664
  0.13142219 -0.5687559   0.66908898 -1.3728184  -0.31446713 -0.92413308
 -0.21311364 -0.29992207 -0.5525467   1.15602521  0.06037807  0.24726618
 -0.07502866 -0.29430864 -0.34061504  0.46896628 -0.15371691  0.77467971
 -1.09272412  0.19261579 -0.56736933  0.3018252   0.67899331  0.47809996
 -0.43818598 -0.40905736]
更新正确三元组向量(头实体)：
 [ 0.03719875 -0.18102125  0.18655618  0.00234794 -0.00593733  0.20994527
  0.2321054  -0.07844085 -0.08978608  0.19390504  0.22651861  0.08342132
  0.04610836  0.02628527 -0.21922707 -0.01643649  0.13454788 -0.01742153
  0.00986113 -0.04198968 -0.22021667 -0.20016039 -0.05358228 -0.17935688
  0.17247146  0.19220583 -0.20978991 -0.01921698  0.08702367 -0.02535365
 -0.12990253 -0.03767222 -0.06934149 -0.17999268 -0.14629433 -0.15465156
 -0.08023339  0.08533757 -0.16837369 -0.17871502  0.21934804  0.17723644
  0.03689182  0.13320018  0.01651606  0.14489723 -0.11404736  0.03530656
 -0.0057997  -0.22038112]
更新正确三元组向量(尾实体)：
 [-0.0589535  -0.20761707 -0.07000709  0.24815392  0.17894299  0.10897056
  0.13024331 -0.05340003 -0.08946519 -0.13465577 -0.07007775 -0.05664604
  0.0386865   0.06207172 -0.13455015  0.03412457  0.03162785 -0.10439232
 -0.20029494 -0.12670348 -0.04617016 -0.22046366  0.09273154 -0.0229746
  0.13404191  0.18611716 -0.23198154  0.23538903  0.15334151  0.04696953
  0.09957064 -0.02955164  0.18214594 -0.17534037 -0.23096015 -0.102417
 -0.2346907  -0.1462859   0.06890641 -0.17964636  0.05445803 -0.07502705
  0.2170002  -0.15103151  0.12880815 -0.11985806 -0.11893992  0.0426667
  0.07104367 -0.09966913]
更新正确三元组向量(关系)：
 [-0.15013055 -0.0613413   0.14403858 -0.04097624 -0.00159057 -0.05022896
  0.22881867  0.04421881  0.0181366  -0.03146179  0.1206439  -0.03007788
 -0.16769409  0.00857801  0.07686555 -0.13158723 -0.06057233 -0.17508344
  0.01198095  0.21370042  0.23430539 -0.10603051  0.02507469  0.1041157
  0.1579308  -0.10475145  0.12651156 -0.24270986 -0.15943899 -0.154605
  0.13893038  0.029265   -0.07100017  0.23972126 -0.18480491  0.23592844
 -0.17683202 -0.10164575 -0.02100775  0.10483633  0.07778871  0.15344844
 -0.23860857 -0.05334123 -0.05404695  0.22272376  0.01837316  0.02416109
 -0.13295101 -0.19703548]
==================================
正确三元组： 9399 1226 14152
错误三元组： 9399 1226 14262
正确三元组的L1范数: 9.992409340840295
错误三元组的L1范数: 9.942922552785314
三元组的距离:  1.0494867880549812
loss: 5.411633578187979
正确三元组的梯度：
 [ 0.3680397   0.32089285  0.62452467 -0.11492201 -0.09262002 -0.43011459
  0.51431885 -0.30068848 -0.30219077  0.75059382 -0.82481915 -0.51594505
  0.79295137 -0.07095999 -0.3475567  -0.03309919  0.27992241 -1.04000586
 -0.95227105  0.25276741  0.05232533  0.88260994 -0.07266404  0.29301016
 -0.15009508 -0.26230577  0.14339828 -0.37980872  1.19761505  0.07059943
  0.33072642  0.19519713  0.53772988  0.2357335   0.17952196 -0.01149659
  0.46860705 -0.44929227  0.04870082 -0.6611804  -0.47435454  0.55607143
 -0.33234647 -0.67720789 -0.47300547 -0.27851468  0.76242954  0.57561381
  0.01060576 -0.29284736]
错误三元组的梯度：
 [ 0.46310257  0.21061891  1.24805701  0.13339741 -0.00294819 -0.22907679
  0.07070678 -0.78669113  0.00839461  0.00849556 -0.81720606 -0.35734205
  0.28473943  0.14521598 -0.14422891 -0.399389    0.1475983  -1.13260506
 -0.46865487  0.04638166  0.08336715  0.75340647 -0.13428276  0.6772844
 -0.57690078 -0.26076736  0.30900861  0.29524747  0.84295118 -0.43183746
  0.05549771 -0.49581024 -0.14328479  0.09092644  0.24354843 -0.23388881
  0.66421034 -0.91978058 -0.01725474 -0.52507413 -0.46899945  0.55209168
 -0.1651277  -0.68579278  0.23637797 -0.39954224  0.56613526  0.90607213
 -0.2003177  -0.84620808]
更新正确三元组向量(头实体)：
 [-0.01831672 -0.18055599  0.20603467  0.22398648 -0.15489488  0.01648978
  0.00368893 -0.0251046  -0.12057599  0.00527216 -0.00037307 -0.13813036
  0.0430363   0.11004196  0.1058476   0.1641077   0.0257299  -0.19869179
 -0.18634728  0.22703158  0.17503868  0.09201114  0.10029876  0.17451867
 -0.20132297  0.06525971  0.1391338   0.13339879  0.2087574  -0.22631443
  0.0393066   0.11737633  0.08829377  0.10877147 -0.01418208 -0.21159191
  0.20547908 -0.19480133 -0.15336722 -0.03435928 -0.06201322  0.1951056
 -0.22891573  0.06897806  0.05911209  0.12833171 -0.04070961  0.057704
  0.11166147 -0.09678127]
更新正确三元组向量(尾实体)：
 [-0.05657985 -0.16712552  0.15457465  0.2142093  -0.06569875  0.11085332
 -0.11512669 -0.06287761 -0.01930346 -0.12491337  0.15426069 -0.06612348
 -0.20488343  0.18162021  0.21864213 -0.00233526  0.05588688  0.11340089
  0.168908    0.02132967  0.13220786 -0.21963586  0.18884161  0.19709918
 -0.13572044  0.07838115  0.03376749  0.18850162 -0.15418359 -0.16783863
  0.01981965 -0.1195906  -0.10097688 -0.13422315  0.15743777 -0.22389771
  0.11677102  0.00318792  0.08573639  0.12540468 -0.00666724 -0.2050859
 -0.07770921  0.2196374   0.2013742   0.06186475 -0.17044875 -0.07208446
 -0.07473142 -0.08943234]
更新正确三元组向量(关系)：
 [ 0.11575673  0.14387689  0.23080231 -0.03723818  0.07288612 -0.09069376
  0.1083438  -0.15811724 -0.01982286  0.21511138 -0.22777581 -0.15596564
  0.11855595  0.06609826 -0.03098382 -0.15299256  0.14011818 -0.17791025
 -0.09088025 -0.10931821 -0.04666815  0.09965796  0.08221084  0.13908558
  0.02055499 -0.08803145 -0.06366717 -0.10480152  0.20586653  0.06377551
  0.11587626 -0.16936836  0.0495943  -0.15512788  0.23138083  0.01194591
  0.11559546  0.00334312  0.23345403 -0.14082623 -0.15183128 -0.15215579
  0.01503328 -0.15794461 -0.06424062 -0.1757243   0.22147562  0.12801845
 -0.21109001 -0.10907475]
==================================
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286