TransE的程序实现——学习，查阅，注释
作者：互联网
代码来源：https://github.com/thunlp/KB2E/blob/master/TransE/
很久没碰C++了，为了实验成功，还是要仔仔细细抠一抠代码才行。
第一部分：头文件的引入以及常量定义
1 #include<iostream> /*输入输出流*/ 2 #include<cstring> /*C字符串操作函数*/ 3 #include<cstdio> /*标准输入输出的C++形式*/ 4 #include<map> /*定义了一种关联容器，数据结构*/ 5 #include<vector> /*顺序容器，常用于表示向量*/ 6 #include<string> /*字符串操作函数*/ 7 #include<ctime> /*日期时间结构体*/ 8 #include<cmath> /*数学操作*/ 9 #include<cstdlib> /*提供一些函数和符号常量，如第二部分出现的RAND_MAX*/ 10 using namespace std; //添加命名空间 11 12 #define pi 3.1415926535897932384626433832795 13 14 bool L1_flag=1; /*标识数为1时，该变量为L1范数；否则，则为L2范数*/
第二部分：简单计算函数的定义
1 //normal distribution 正态分布
2 double rand(double min, double max) /*返回[min,max)之间的随机数*/
3 {
4 return min+(max-min)*rand()/(RAND_MAX+1.0);
5 }
6 double normal(double x, double miu,double sigma) /*返回一个均值为miu标准差为sigma的正态分布函数在x处的函数值*/
7 {
8 return 1.0/sqrt(2*pi)/sigma*exp(-1*(x-miu)*(x-miu)/(2*sigma*sigma));
9 }
10 double randn(double miu,double sigma, double min ,double max) /*通过产生随机数的方式，返回满足某条件的自变量x*/
11 {
12 double x,y,dScope;
13 do{
14 x=rand(min,max); /*调用自定义函数rand(min,max)*/
15 y=normal(x,miu,sigma); /*调用自定义函数normal(x,miu,sigma)*/
16 dScope=rand(0.0,normal(miu,miu,sigma)); /*当x值取miu时，正态函数达到最大值，该变量在0与正态函数最大值取随机数*/
17 }while(dScope>y); /*x处的函数值小于该随机数时，循环停止*/
18 return x;
19 }
20
21 double sqr(double x) /*平方函数*/
22 {
23 return x*x;
24 }
25
26 double vec_len(vector<double> &a) /*返回向量a的模，即L2范数*/
27 {
28 double res=0;
29 for (int i=0; i<a.size(); i++) /*遍历向量长度*/
30 res+=a[i]*a[i];
31 res = sqrt(res);
32 return res;
33 }
rand()返回一个0到最大随机数RAND_MAX（确定）的任意整数，RAND_MAX至少为32767。
第三部分：变量的定义
1 string version; 2 char buf[100000],buf1[100000]; 3 int relation_num,entity_num; /*定义关系数量，实体数量*/ 4 map<string,int> relation2id,entity2id; /*关系和实体使用关联容器，字符串作关键词，整型作关键字的值*/ 5 map<int,string> id2entity,id2relation; /*与上相反*/ 6 7 map<int,map<int,int> > left_entity,right_entity; 8 map<int,double> left_num,right_num;
第7行，left_entity表示在此relation下头实体对应的尾实体的个数，3个int分别表示relation_id,headentity_id,个数。right_entity表示在此relation下尾实体对应的头实体的个数，3个int分别表示relation_id,tailentity_id,个数。主要用于计算采样概率p。
第8行，leftnum表示平均每个头实体对应多少个尾实体。rightnum表示平均每个尾实体对应多少头实体。int仍然表示relation_id。
第四部分：训练类的定义
1 class Train{
2
3 public:
4 map<pair<int,int>, map<int,int> > ok;
5 void add(int x,int y,int z)
6 {
7 fb_h.push_back(x);
8 fb_r.push_back(z);
9 fb_l.push_back(y);
10 ok[make_pair(x,z)][y]=1;
11 }
12 void run(int n_in,double rate_in,double margin_in,int method_in)
13 {
14 n = n_in;
15 rate = rate_in;
16 margin = margin_in;
17 method = method_in;
18 relation_vec.resize(relation_num);
19 for (int i=0; i<relation_vec.size(); i++)
20 relation_vec[i].resize(n);
21 entity_vec.resize(entity_num);
22 for (int i=0; i<entity_vec.size(); i++)
23 entity_vec[i].resize(n);
24 relation_tmp.resize(relation_num);
25 for (int i=0; i<relation_tmp.size(); i++)
26 relation_tmp[i].resize(n);
27 entity_tmp.resize(entity_num);
28 for (int i=0; i<entity_tmp.size(); i++)
29 entity_tmp[i].resize(n);
30 for (int i=0; i<relation_num; i++)
31 {
32 for (int ii=0; ii<n; ii++)
33 relation_vec[i][ii] = randn(0,1.0/n,-6/sqrt(n),6/sqrt(n));
34 }
35 for (int i=0; i<entity_num; i++)
36 {
37 for (int ii=0; ii<n; ii++)
38 entity_vec[i][ii] = randn(0,1.0/n,-6/sqrt(n),6/sqrt(n));
39 norm(entity_vec[i]);
40 }
41
42
43 bfgs();
44 }
45
46 private:
47 int n,method;
48 double res;//loss function value
49 double count,count1;//loss function gradient
50 double rate,margin;
51 double belta;
52 vector<int> fb_h,fb_l,fb_r;
53 vector<vector<int> > feature;
54 vector<vector<double> > relation_vec,entity_vec;
55 vector<vector<double> > relation_tmp,entity_tmp;
56 double norm(vector<double> &a)
57 {
58 double x = vec_len(a);
59 if (x>1)
60 for (int ii=0; ii<a.size(); ii++)
61 a[ii]/=x;
62 return 0;
63 }
64 int rand_max(int x)
65 {
66 int res = (rand()*rand())%x;
67 while (res<0)
68 res+=x;
69 return res;
70 }
71
72 void bfgs()
73 {
74 res=0;
75 int nbatches=100;
76 int nepoch = 1000;
77 int batchsize = fb_h.size()/nbatches;
78 for (int epoch=0; epoch<nepoch; epoch++)
79 {
80
81 res=0;
82 for (int batch = 0; batch<nbatches; batch++)
83 {
84 relation_tmp=relation_vec;
85 entity_tmp = entity_vec;
86 for (int k=0; k<batchsize; k++)
87 {
88 int i=rand_max(fb_h.size());
89 int j=rand_max(entity_num);
90 double pr = 1000*right_num[fb_r[i]]/(right_num[fb_r[i]]+left_num[fb_r[i]]);
91 if (method ==0)
92 pr = 500;
93 if (rand()%1000<pr)
94 {
95 while (ok[make_pair(fb_h[i],fb_r[i])].count(j)>0)
96 j=rand_max(entity_num);
97 train_kb(fb_h[i],fb_l[i],fb_r[i],fb_h[i],j,fb_r[i]);
98 }
99 else
100 {
101 while (ok[make_pair(j,fb_r[i])].count(fb_l[i])>0)
102 j=rand_max(entity_num);
103 train_kb(fb_h[i],fb_l[i],fb_r[i],j,fb_l[i],fb_r[i]);
104 }
105 norm(relation_tmp[fb_r[i]]);
106 norm(entity_tmp[fb_h[i]]);
107 norm(entity_tmp[fb_l[i]]);
108 norm(entity_tmp[j]);
109 }
110 relation_vec = relation_tmp;
111 entity_vec = entity_tmp;
112 }
113 cout<<"epoch:"<<epoch<<' '<<res<<endl;
114 FILE* f2 = fopen(("relation2vec."+version).c_str(),"w");
115 FILE* f3 = fopen(("entity2vec."+version).c_str(),"w");
116 for (int i=0; i<relation_num; i++)
117 {
118 for (int ii=0; ii<n; ii++)
119 fprintf(f2,"%.6lf\t",relation_vec[i][ii]);
120 fprintf(f2,"\n");
121 }
122 for (int i=0; i<entity_num; i++)
123 {
124 for (int ii=0; ii<n; ii++)
125 fprintf(f3,"%.6lf\t",entity_vec[i][ii]);
126 fprintf(f3,"\n");
127 }
128 fclose(f2);
129 fclose(f3);
130 }
131 }
132 double res1;
133 double calc_sum(int e1,int e2,int rel)
134 {
135 double sum=0;
136 if (L1_flag)
137 for (int ii=0; ii<n; ii++)
138 sum+=fabs(entity_vec[e2][ii]-entity_vec[e1][ii]-relation_vec[rel][ii]);
139 else
140 for (int ii=0; ii<n; ii++)
141 sum+=sqr(entity_vec[e2][ii]-entity_vec[e1][ii]-relation_vec[rel][ii]);
142 return sum;
143 }
144 void gradient(int e1_a,int e2_a,int rel_a,int e1_b,int e2_b,int rel_b)
145 {
146 for (int ii=0; ii<n; ii++)
147 {
148
149 double x = 2*(entity_vec[e2_a][ii]-entity_vec[e1_a][ii]-relation_vec[rel_a][ii]);
150 if (L1_flag)
151 if (x>0)
152 x=1;
153 else
154 x=-1;
155 relation_tmp[rel_a][ii]-=-1*rate*x;
156 entity_tmp[e1_a][ii]-=-1*rate*x;
157 entity_tmp[e2_a][ii]+=-1*rate*x;
158 x = 2*(entity_vec[e2_b][ii]-entity_vec[e1_b][ii]-relation_vec[rel_b][ii]);
159 if (L1_flag)
160 if (x>0)
161 x=1;
162 else
163 x=-1;
164 relation_tmp[rel_b][ii]-=rate*x;
165 entity_tmp[e1_b][ii]-=rate*x;
166 entity_tmp[e2_b][ii]+=rate*x;
167 }
168 }
169 void train_kb(int e1_a,int e2_a,int rel_a,int e1_b,int e2_b,int rel_b)
170 {
171 double sum1 = calc_sum(e1_a,e2_a,rel_a);
172 double sum2 = calc_sum(e1_b,e2_b,rel_b);
173 if (sum1+margin>sum2)
174 {
175 res+=margin+sum1-sum2;
176 gradient( e1_a, e2_a, rel_a, e1_b, e2_b, rel_b);
177 }
178 }
179 };
第五部分：类变量的定义和数据集准备
1 Train train;
2 void prepare()
3 {
4 FILE* f1 = fopen("../data/entity2id.txt","r");
5 FILE* f2 = fopen("../data/relation2id.txt","r");
6 int x;
7 while (fscanf(f1,"%s%d",buf,&x)==2)
8 {
9 string st=buf;
10 entity2id[st]=x;
11 id2entity[x]=st;
12 entity_num++;
13 }
14 while (fscanf(f2,"%s%d",buf,&x)==2)
15 {
16 string st=buf;
17 relation2id[st]=x;
18 id2relation[x]=st;
19 relation_num++;
20 }
21 FILE* f_kb = fopen("../data/train.txt","r");
22 while (fscanf(f_kb,"%s",buf)==1)
23 {
24 string s1=buf;
25 fscanf(f_kb,"%s",buf);
26 string s2=buf;
27 fscanf(f_kb,"%s",buf);
28 string s3=buf;
29 if (entity2id.count(s1)==0)
30 {
31 cout<<"miss entity:"<<s1<<endl;
32 }
33 if (entity2id.count(s2)==0)
34 {
35 cout<<"miss entity:"<<s2<<endl;
36 }
37 if (relation2id.count(s3)==0)
38 {
39 relation2id[s3] = relation_num;
40 relation_num++;
41 }
42 left_entity[relation2id[s3]][entity2id[s1]]++;
43 right_entity[relation2id[s3]][entity2id[s2]]++;
44 train.add(entity2id[s1],entity2id[s2],relation2id[s3]);
45 }
46 for (int i=0; i<relation_num; i++)
47 {
48 double sum1=0,sum2=0;
49 for (map<int,int>::iterator it = left_entity[i].begin(); it!=left_entity[i].end(); it++)
50 {
51 sum1++;
52 sum2+=it->second;
53 }
54 left_num[i]=sum2/sum1;
55 }
56 for (int i=0; i<relation_num; i++)
57 {
58 double sum1=0,sum2=0;
59 for (map<int,int>::iterator it = right_entity[i].begin(); it!=right_entity[i].end(); it++)
60 {
61 sum1++;
62 sum2+=it->second;
63 }
64 right_num[i]=sum2/sum1;
65 }
66 cout<<"relation_num="<<relation_num<<endl;
67 cout<<"entity_num="<<entity_num<<endl;
68 fclose(f_kb);
69 }
第六部分：未知还没搞懂功能的函数
1 int ArgPos(char *str, int argc, char **argv) {
2 int a;
3 for (a = 1; a < argc; a++) if (!strcmp(str, argv[a])) {
4 if (a == argc - 1) {
5 printf("Argument missing for %s\n", str);
6 exit(1);
7 }
8 return a;
9 }
10 return -1;
11 }
第七部分：主函数流程
1 int main(int argc,char**argv)
2 {
3 srand((unsigned) time(NULL));
4 int method = 1;
5 int n = 100;
6 double rate = 0.001;
7 double margin = 1;
8 int i;
9 if ((i = ArgPos((char *)"-size", argc, argv)) > 0) n = atoi(argv[i + 1]);
10 if ((i = ArgPos((char *)"-margin", argc, argv)) > 0) margin = atoi(argv[i + 1]);
11 if ((i = ArgPos((char *)"-method", argc, argv)) > 0) method = atoi(argv[i + 1]);
12 cout<<"size = "<<n<<endl;
13 cout<<"learing rate = "<<rate<<endl;
14 cout<<"margin = "<<margin<<endl;
15 if (method)
16 version = "bern";
17 else
18 version = "unif";
19 cout<<"method = "<<version<<endl;
20 prepare();
21 train.run(n,rate,margin,method);
22 }
left_entity：在此relation下头实体对应的尾实体的个数，3个int分别表示relation_id,headentity_id,个数标签：tmp,程序实现,double,entity,int,fb,relation,TransE,查阅 来源： https://www.cnblogs.com/real-zz-11/p/real_zz.html