-
Notifications
You must be signed in to change notification settings - Fork 2
/
Model.cpp
206 lines (182 loc) · 5.56 KB
/
Model.cpp
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
#include "Model.h"
#include <math.h>
#include <omp.h>
#include <assert.h>
#include "core.h"
#include <iostream>
#include <fstream>
#include <vector>
#include <string>
using namespace std;
cmdline::parser opts;
void Model::Init() {
int L = n_variables();
int K = L-1;
theta.resize(K);
theta_key.resize(K);
grad_theta.resize(K);
expThetaRows.resize(K);
expThetaCols.resize(K);
int n = 0;
M = 0;
for (int l = 0; l < L; l++) {
if (l == 0) continue;
theta[n].resize(n_states(0));
theta_key[n].resize(n_states(0));
grad_theta[n].resize(n_states(0));
expThetaRows[n].resize(n_states(0));
for (int a = 0; a < n_states(0); a++) {
theta[n][a].resize(n_states(l), 0.0);
theta_key[n][a].resize(n_states(l));
grad_theta[n][a].resize(n_states(l), 0.0);
expThetaRows[n][a].resize(n_states(l)+1);
for (int b = 0; b < n_states(l); b++) {
theta[n][a][b] = 0;
// Another param.
theta_key[n][a][b] = M;
M++;
grad_theta[n][a][b] = 0;
}
}
// Transpose of expThetaRows.
expThetaCols[n].resize(n_states(l));
for (int a = 0; a < n_states(l); a++) {
expThetaCols[n][a].resize(n_states(0)+1);
}
n++;
}
}
void Model::BackpropGradient(double *grad) {
// Full rank model.
// Backprop is trivial.
int n = 0;
for (int l = 0; l < n_variables(); l++) {
if (l == 0) continue;
#pragma omp parallel for
for (int s = 0; s < n_states(0); s++) {
for (int t = 0; t < n_states(l); t++) {
int key = theta_key[n][s][t];
grad[key] = grad_theta[n][s][t];
++key;
}
}
++n;
}
}
void Model::SetWeights(const double *const_weight, bool reverse) {
// Full rank model.
// Setweights is trivial.
double *weight = (double *)const_weight;
int n = 0;
for (int l = 0; l < n_variables(); l++) {
if (l == 0) continue;
#pragma omp parallel for
for (int s = 0; s < n_states(0); s++) {
for (int t = 0; t < n_states(l); t++) {
// int key = map_key(n, s, t, n_states(0), n_states(l));
int key = theta_key[n][s][t];
if (!reverse) {
theta[n][s][t] = weight[key];
} else {
weight[key] = theta[n][s][t];
}
++key;
}
}
++n;
}
Exponentiate();
}
void Model::Exponentiate() {
int n = 0;
for (int l = 0; l < n_variables(); l++) {
if (l == 0) continue;
vector<double> temp(n_states(l));
#pragma omp parallel for
for (int s = 0; s < n_states(0); s++) {
for (int t = 0; t < n_states(l); t++) {
expThetaRows[n][s][t] = n_trees() * theta[n][s][t];
}
exptab(expThetaRows[n][s],
expThetaRows[n][s], n_states(l));
}
#pragma omp parallel for
for (int t = 0; t < n_states(l); t++) {
for (int s = 0; s < n_states(0); s++) {
expThetaCols[n][t][s] = n_trees() * theta[n][s][t];
}
exptab(expThetaCols[n][t],
expThetaCols[n][t], n_states(0));
}
++n;
}
}
// Read the model file from disk.
void Model::ReadModel(string mf) {
ifstream myfile(mf);
ReadParams(myfile);
Init();
myfile >> M;
double *weights = new double[M];
for (int m = 0; m < M; ++m) {
myfile >> weights[m];
}
myfile.close();
SetWeights(weights, false);
}
// Write out the global model file to disk.
void Model::WriteModel(string mf) {
ofstream myfile(mf);
WriteParams(myfile);
myfile << M << "\n";
double *weights = new double[M];
SetWeights(weights, true);
for (int m = 0; m < M; ++m) {
myfile << weights[m] << " ";
}
myfile.close();
}
// This function returns the value of the log-likelihood lower bound
// given the current log-partition
double Model::ComputeObjective(const Moments &mom,
double partition) {
double res = 0;
for (int k = 0; k < mom.L - 1; k++) {
for (int pa = 0; pa < mom.nPairs[k]; pa++) {
const vector<int> &p = mom.Pairs[k][pa];
res += theta[k][p[0]][p[1]] * ((double)p[2]) / mom.N;
}
}
printf("---Energy %.4e, Partition %.4e \t \t \t \t \t >>> OBJECTIVE %.4e\n",
res, partition / (n_trees()),
res - partition / (n_trees()));
return res - partition / (n_trees());
}
struct Moments train_moments;
struct Moments valid_moments;
// Read the moments text file.
void ReadMoments(string mf, Moments *moments) {
ifstream myfile(mf);
myfile >> moments->N >> moments->L;
moments->sizes.resize(moments->L);
for (int k = 0; k < moments->L; k++) {
myfile >> moments->sizes[k];
}
moments->nPairs.resize(moments->L-1);
moments->Pairs.resize(moments->L-1);
for (int k = 0; k < moments->L - 1; k++) {
myfile >> moments->nPairs[k];
printf("k=%d, npairs=%d\n", k, moments->nPairs[k]);
moments->Pairs[k].resize(moments->nPairs[k]);
for (int i = 0; i < moments->nPairs[k]; i++) {
moments->Pairs[k][i].resize(3);
for (int j = 0; j < 3; j++) {
int p;
myfile >> p;
moments->Pairs[k][i][j] = p;
assert(p >= 0);
}
}
}
myfile.close();
}