-
Notifications
You must be signed in to change notification settings - Fork 0
/
fc_net.h
290 lines (260 loc) · 8.92 KB
/
fc_net.h
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
287
288
289
290
#ifndef __FC_NET_H__
#define __FC_NET_H__
#include<string>
#include<map>
#include"matrix.h"
#include<iostream>
#include<sstream>
class fc_net
{
private:
typedef std::map<std::string,matrix<double>> map1str;
typedef std::map<std::string,map1str> map2str;
public:
explicit fc_net(const int layer[],const int num):num_layers(num)
{
for(int i=1;i!=num_layers;++i)
{
std::string strW="W"+num2str(i);
std::string strb="b"+num2str(i);
matrix<double> W_temp(layer[i-1],layer[i]);
W_temp.init_xavier();
matrix<double> b_temp(1,layer[i]);
b_temp.init_zeros();
parameter[strW]=W_temp;
parameter[strb]=b_temp;
}
}
explicit fc_net()=default;
map1str get_parameter()
{
return parameter;
}
int get_num_layers()
{
return num_layers;
}
public:
double train(matrix<double> &dataset,matrix<int> &labels,const double learning_rate,
const double reg);
matrix<int> predict(matrix<double> &dataset);
private:
map1str parameter;
int num_layers;
private:
std::string num2str(const double i)
{
std::stringstream ss;
ss<<i;
return ss.str();
}
private:
map1str softmax(matrix<double> &scores,matrix<int> &labels,const double reg);
map2str affine_forward(matrix<double> &x,matrix<double> &W,matrix<double> &b);
map1str affine_backward(matrix<double> &upstream_grad,map1str &cache);
map2str relu_forward(matrix<double> &lhs);
matrix<double> relu_backward(matrix<double> &upstream_grad,map1str &cache);
map2str batchnorm_forward(matrix<double> &x,map1str ¶,const std::string mode,const double momentum);
map1str batchnorm_backward(matrix<double> &upstream_grad,map1str &cache);
map2str dropout_forward(matrix<double> &x, const double p,const std::string mode);
matrix<double> dropout_backward(matrix<double> &upstream_grad,map1str &cache);
};
fc_net::map1str fc_net::softmax(matrix<double> &scores,matrix<int> &labels,const double reg)
{
assert((labels.axisy()==1));
map1str result;
matrix<double> correct_scores(labels.axisx(),1);
for(int i=0;i!=labels.axisx();++i)
{
correct_scores[i][0]=scores[i][labels[i][0]];
}
matrix<double> exp_=(scores-correct_scores).exp();
matrix<double> exp_sum=exp_.sum(0);
matrix<double> log_=exp_sum.log();
matrix<double> loss=matrix<double>(1,1);
loss[0][0]=log_.sum()/scores.axisx();
for(int i=1;i!=num_layers;++i)
{
loss[0][0]+=reg*(parameter["W"+num2str(i)].square().sum());
}
result["loss"]=loss;
matrix<double> dscores=exp_/exp_sum;
for(int i=0;i!=labels.axisx();++i)
{
dscores[i][labels[i][0]]-=1;
}
dscores/=labels.axisx();
result["dscores"]=dscores;
for(int i=1;i!=num_layers;++i)
{
result["dW"+num2str(i)]=2*reg*parameter["W"+num2str(i)];
}
return result;
}
fc_net::map2str fc_net::affine_forward(matrix<double> &x,matrix<double> &W,matrix<double> &b)
{
map2str result;
matrix<double> affine=x.dot(W)+b;
map1str obj;
map1str cache;
obj["affine"]=affine;
cache["x"]=x;
cache["W"]=W;
cache["b"]=b;
result["affine"]=obj;
result["cache"]=cache;
return result;
}
fc_net::map2str fc_net::relu_forward(matrix<double> &lhs)
{
map2str result;
map1str obj;
map1str cache;
matrix<double> relu=lhs.maximum(0.0);
obj["relu"]=relu;
cache["input"]=lhs;
result["relu"]=obj;
result["cache"]=cache;
return result;
}
fc_net::map1str fc_net::affine_backward(matrix<double> &upstream_grad,map1str &cache)
{
map1str result;
matrix<double> dx;
matrix<double> dW;
matrix<double> db;
dx=upstream_grad.dot(cache["W"].t());
dW=cache["x"].t().dot(upstream_grad);
db=upstream_grad.sum(1);
assert(db.axisx()==1);
result["dx"]=dx;
result["dW"]=dW;
result["db"]=db;
return result;
}
matrix<double> fc_net::relu_backward(matrix<double> &upstream_grad,map1str &cache)
{
matrix<double> result=upstream_grad*(cache["input"]>0);
return result;
}
fc_net::map2str fc_net::batchnorm_forward(matrix<double> &x,map1str ¶,const std::string mode,const double momentum)
{
map2str result;
map1str output;
map1str cache;
assert((mode=="train")||(mode=="test"));
if(mode=="train")
{
matrix<double> sample_mean=x.mean(1);
matrix<double> sample_var=x.var(1);
matrix<double> x_gauss=(x-sample_mean)/((sample_var+1e-5).sqrt());
matrix<double> out=para["gamma"]*x_gauss+para["beta"];
output["result"]=out;
cache["x"]=x;
cache["sample_mean"]=sample_mean;
cache["sample_var"]=sample_var;
cache["x_gauss"]=x_gauss;
cache["gamma"]=para["gamma"];
cache["beta"]=para["beta"];
para["running_mean"] = momentum * para["running_mean"] + (1 - momentum) * para["running_mean"];
para["running_var"] = momentum * para["running_var"] + (1 - momentum) * para["running_var"];
}else if(mode=="test")
{
matrix<double> out=(x-para["running_mean"])*para["gamma"]/(para["running_var"]+1e-5)+para["beta"];
output["result"]=out;
}
result["result"]=output;
result["cache"]=cache;
return result;
}
fc_net::map1str fc_net::batchnorm_backward(matrix<double> &upstream_grad,map1str &cache)
{
map1str result;
int N=upstream_grad.axisx();
matrix<double> dx_gauss=upstream_grad*cache["gamma"];
matrix<double> dx=(N*dx_gauss-((dx_gauss*cache["x_gauss"]).sum(1))*cache["x_gauss"]-dx_gauss.sum(1))/(cache["sample_var"]+1e-5).sqrt()/N;
matrix<double> dgamma=(upstream_grad*cache["x_gauss"]).sum(1);
matrix<double> dbeta=upstream_grad.sum(1);
result["dx"]=dx;
result["dgamma"]=dgamma;
result["dbeta"]=dbeta;
return result;
}
fc_net::map2str fc_net::dropout_forward(matrix<double> &x, const double p,const std::string mode)
{
assert((mode=="train")||(mode=="test"));
assert((p>=0)&&(p<=1));
map2str result;
map1str out;
map1str cache;
if(mode=="train")
{
matrix<double> temp(x.axisx(),x.axisy());
temp.init_uniform();
matrix<double> mask=(temp<p)*temp/p;
out["result"]=x*mask;
cache["mask"]=mask;
}else if(mode=="test")
{
out["result"]=x;
}
result["result"]=out;
result["cache"]=cache;
return result;
}
matrix<double> fc_net::dropout_backward(matrix<double> &upstream_grad,map1str &cache)
{
return upstream_grad*cache["mask"];
}
double fc_net::train(matrix<double> &dataset,matrix<int> &labels,const double learning_rate,
const double reg)
{
std::map<std::string,map2str> cache;
map2str affine_result;
map2str relu_result;
map1str relu_result1;
relu_result1["relu"]=dataset;
relu_result["relu"]=relu_result1;
for(int i=1;i!=num_layers-1;++i)
{
affine_result=affine_forward(relu_result["relu"]["relu"],parameter["W"+num2str(i)],parameter["b"+num2str(i)]);
relu_result=relu_forward(affine_result["affine"]["affine"]);
cache["affine"+num2str(i)]=affine_result;
cache["relu"+num2str(i)]=relu_result;
}
map2str scores=affine_forward(relu_result["relu"]["relu"],parameter["W"+num2str(num_layers-1)],parameter["b"+num2str(num_layers-1)]);
map1str softmax_=softmax(scores["affine"]["affine"],labels,reg);
double loss=softmax_["loss"][0][0];
map1str affine_back_result=affine_backward(softmax_["dscores"],scores["cache"]);
matrix<double> relu_back_result;
parameter["W"+num2str(num_layers-1)]-=learning_rate*affine_back_result["dW"];
parameter["W"+num2str(num_layers-1)]-=learning_rate*softmax_["dW"+num2str(num_layers-1)];
parameter["b"+num2str(num_layers-1)]-=learning_rate*affine_back_result["db"];
for(int i=num_layers-2;i!=0;--i)
{
relu_back_result=relu_backward(affine_back_result["dx"],cache["relu"+num2str(i)]["cache"]);
affine_back_result=affine_backward(relu_back_result,cache["affine"+num2str(i)]["cache"]);
parameter["W"+num2str(i)]-=learning_rate*affine_back_result["dW"];
parameter["W"+num2str(i)]-=learning_rate*softmax_["dW"+num2str(i)];
parameter["b"+num2str(i)]-=learning_rate*affine_back_result["db"];
}
return loss;
}
matrix<int> fc_net::predict(matrix<double> &dataset)
{
matrix<int> result;
map2str affine_result;
map2str relu_result;
map1str relu_result1;
relu_result1["relu"]=dataset;
relu_result["relu"]=relu_result1;
for(int i=1;i!=num_layers-1;++i)
{
affine_result=affine_forward(relu_result["relu"]["relu"],parameter["W"+num2str(i)],parameter["b"+num2str(i)]);
relu_result=relu_forward(affine_result["affine"]["affine"]);
}
map2str scores=affine_forward(relu_result["relu"]["relu"],parameter["W"+num2str(num_layers-1)],parameter["b"+num2str(num_layers-1)]);
result=scores["affine"]["affine"].argmax(0);
return result;
}
#endif // __FC_NET_H__