-
Notifications
You must be signed in to change notification settings - Fork 0
/
wrapper.cc
187 lines (138 loc) · 4.96 KB
/
wrapper.cc
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
#include <fstream>
#include <iostream>
#include "party.h"
#include "privacy.h"
#include "wrapper.h"
#include "train.h"
configuration_t* GetConfiguration(char* filename) {
std::ifstream config_file(filename);
auto c = new Configuration(config_file);
return new configuration_t{c};
}
party_t* GetParty(configuration_t* config, char* data_filename, int is_training, int skip_rows=0) {
std::ifstream data_stream(data_filename);
for (int i = 0; i < skip_rows; i++) {
std::string temp;
std::getline(data_stream, temp);
}
auto p = new Party((Configuration*)config->data, data_stream, is_training);
return new party_t{p};
}
model_t* InitialModel(configuration_t* config) {
auto c = (Configuration*)config->data;
auto p = new Eigen::VectorXd(c->d+1);
p->setZero();
return new model_t{p, c};
}
model_t* UnquantizeModel(configuration_t* config, int* quantized, int precision) {
auto c = (Configuration*)config->data;
model_t* m = InitialModel(config);
Eigen::VectorXd* params = (Eigen::VectorXd*)m->params;
for (int i = 0; i < c->d+1; i++) {
(*params)(i) = quantized[i];
}
(*params) /= (1 << precision);
return m;
}
model_t* UnquantizeLongModel(configuration_t* config, long* quantized, int precision) {
auto c = (Configuration*)config->data;
model_t* m = InitialModel(config);
Eigen::VectorXd* params = (Eigen::VectorXd*)m->params;
for (int i = 0; i < c->d+1; i++) {
(*params)(i) = quantized[i];
}
(*params) /= (1L << precision);
return m;
}
int* quantizeVector(Eigen::VectorXd vec, int fractional_bits) {
int shift = 1 << fractional_bits;
auto v = new int[vec.size()];
for (int i = 0; i < vec.size(); i++) {
v[i] = (int)(vec(i) * shift);
}
return v;
}
Eigen::VectorXd unquantizeVector(int* vec, int length, int fractional_bits) {
int shift = 1 << fractional_bits;
Eigen::VectorXd v(length);
for (int i = 0; i < length; i++) {
v(i) = ((double)vec[i]) / shift;
}
return v;
}
int* ComputeGradient(party_t* party, model_t* model) {
auto p = (Party*)party->data;
auto c = (Configuration*)model->config;
auto dgrad = p->ComputeGradient(c, *(Eigen::VectorXd*)model->params);
return quantizeVector(dgrad, c->fractional_bits);
}
void UpdateModel(model_t* model, int step_num, int* gradient) {
auto c = (Configuration*)model->config;
auto dgrad = unquantizeVector(gradient, c->d+1, c->fractional_bits);
int batches_per_epoch = c->m / c->batch_size;
int epoch_num = step_num / batches_per_epoch;
double learning_rate = c->initial_learning_rate / (1 + c->learning_rate_decay * epoch_num);
dgrad /= c->n;
Eigen::VectorXd* params = (Eigen::VectorXd*)model->params;
(*params) -= learning_rate * dgrad;
}
int* ComputeNoisyGradient(party_t* party, model_t* model) {
auto p = (Party*)party->data;
auto c = (Configuration*)model->config;
auto dgrad = p->ComputeGradient(c, *(Eigen::VectorXd*)model->params);
// sample noise distribution
dgrad += c->privacy.generateLogisticRegressionNoise(c->clipping, c->batch_size, c->m, c->epochs, c->d);
return quantizeVector(dgrad, c->fractional_bits);
}
int GetIterationCount(configuration_t* config) {
auto c = (Configuration*)config->data;
return c->epochs * c->m / c->batch_size;
}
double EvaluateModel(party_t* party, model_t* model) {
auto p = (Party*)party->data;
return p->Accuracy(*(Eigen::VectorXd*)model->params);
}
int GetDataFeatureCount(party_t* party) {
auto p = (Party*)party->data;
return p->features.cols();
}
int GetDataRowCount(party_t* party) {
auto p = (Party*)party->data;
return p->features.rows();
}
int GetBatchSize(configuration_t* config) {
auto c = (Configuration*)config->data;
return c->batch_size;
}
int GetQuantizeBitsPrecision(configuration_t* config) {
auto c = (Configuration*)config->data;
return c->fractional_bits;
}
// assume that features and labels have already been allocated and are of the correct size
void QuantizePartyData(party_t* party, long** features, long* labels, int precision) {
auto p = (Party*)party->data;
Eigen::Matrix<long, Eigen::Dynamic, Eigen::Dynamic> long_mat = (p->features * (1 << precision)).cast<long>();
for (int i = 0; i < long_mat.rows(); i++) {
for (int j = 0; j < long_mat.cols(); j++) {
features[i][j] = long_mat(i,j);
}
labels[i] = (long)p->labels(i);
}
}
double GetLearningRate(configuration_t* config, int iteration) {
auto c = (Configuration*)config->data;
return getLearningRate(c, iteration);
}
double GetNoiseStdDev(configuration_t* config) {
auto c = (Configuration*)config->data;
return c->privacy.getMomentsAccountStandardDev(c->clipping, c->batch_size, c->m, c->epochs);
}
double GetRegularizedRegressionNoise(configuration_t* config) {
auto c = (Configuration*)config->data;
// we will only use this wrapper if there are 2 parties
return c->privacy.getRegularizedRegressionStandardDev(c->regularization, 2*c->m);
}
double GetRegularization(configuration_t* config) {
auto c = (Configuration*)config->data;
return c->regularization;
}