-
Notifications
You must be signed in to change notification settings - Fork 0
/
optimizer.h
170 lines (156 loc) · 7.07 KB
/
optimizer.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
#ifndef _OPTIMIZER_H
#define _OPTIMIZER_H
#include "nn.h"
#include "basic_operations.h"
typedef struct Velocity{
Matrix* weights;
Matrix* biases;
}Velocity;
typedef struct Moments{
Matrix* moment1_W;
Matrix* moment1_b;
Matrix* moment2_W;
Matrix* moment2_b;
}Moments;
typedef struct SGD{
float lr;
float momentum;
int nesterov;
float batch_size;
Velocity velocity[100];
NN* model;
}SGD;
typedef struct Adam{
float lr;
float batch_size;
Moments moments[100];
NN* model;
}Adam;
const float beta_1 = 0.9;
const float beta_2 = 0.999;
const float epsilon = 1e-8;
void initSGD(SGD* optimizer,NN* model, float lr ,float batch_size, float momentum)
{
optimizer->lr = lr;
optimizer->momentum = momentum;
optimizer->model = model;
optimizer->batch_size = batch_size;
optimizer->nesterov = 0;
for(int i=0; i<optimizer->model->num_of_layers ; i++)
{
if(!optimizer->model->layers[i].is_trainable)
continue;
optimizer->velocity[i].weights = createMatrix(optimizer->model->layers[i].weight_gradients->rows, optimizer->model->layers[i].weight_gradients->columns , 0.0f);
optimizer->velocity[i].biases = createMatrix(optimizer->model->layers[i].bias_gradients->rows, optimizer->model->layers[i].bias_gradients->columns , 0.0f);
}
}
void optimizeSGD(SGD* optimizer)
{
for(int i=0; i<optimizer->model->num_of_layers ; i++)
{
if(!optimizer->model->layers[i].is_trainable)
continue;
//calculate velocity = momentum*velocity - lr*gradients
scaleMatrix( optimizer->velocity[i].weights , optimizer->momentum);
scaleMatrix(optimizer->velocity[i].biases , optimizer->momentum);
scaleMatrix( optimizer->model->layers[i].weight_gradients , -optimizer->lr/optimizer->batch_size);
scaleMatrix(optimizer->model->layers[i].bias_gradients , -optimizer->lr/optimizer->batch_size);
addMatrix(optimizer->velocity[i].weights, optimizer->model->layers[i].weight_gradients, optimizer->velocity[i].weights);
addMatrix(optimizer->velocity[i].biases, optimizer->model->layers[i].bias_gradients, optimizer->velocity[i].biases);
if(optimizer->nesterov)
{
addMatrix(optimizer->model->layers[i].weights, optimizer->model->layers[i].weight_gradients, optimizer->model->layers[i].weights);
addMatrix(optimizer->model->layers[i].biases, optimizer->model->layers[i].bias_gradients, optimizer->model->layers[i].biases);
scale_and_addMatrix(optimizer->model->layers[i].weights, optimizer->velocity[i].weights, optimizer->model->layers[i].weights , optimizer->momentum);
scale_and_addMatrix(optimizer->model->layers[i].biases, optimizer->velocity[i].biases, optimizer->model->layers[i].biases , optimizer->momentum);
}
else
{
//update weights W = W + velocity
addMatrix(optimizer->model->layers[i].weights, optimizer->velocity[i].weights, optimizer->model->layers[i].weights);
addMatrix(optimizer->model->layers[i].biases, optimizer->velocity[i].biases, optimizer->model->layers[i].biases);
}
//zero gradients
zeroMatrix(optimizer->model->layers[i].weight_gradients);
zeroMatrix(optimizer->model->layers[i].bias_gradients);
}
}
void freeSGD(SGD* optimizer)
{
freeNN(optimizer->model);
for(int i=0; i<optimizer->model->num_of_layers ; i++)
{
freeMatrix(&optimizer->velocity[i].weights);
freeMatrix(&optimizer->velocity[i].biases);
}
}
void calculate_moment1(Matrix* moment, Matrix* gradient)
{
for(int i=0; i < (moment->rows * moment->columns) ; i++)
{
moment->data[i] = ( beta_1 * moment->data[i] + (1-beta_1) * gradient->data[i] );
}
}
void calculate_moment2(Matrix* moment, Matrix* gradients)
{
for(int i=0; i< moment->rows * moment->columns ; i++)
{
moment->data[i] = (beta_2 * moment->data[i] + (1-beta_2) * gradients->data[i] * gradients->data[i]) ;
}
}
void update_parameters_Adam(Matrix* moment1, Matrix* moment2, Matrix* parameters,float lr)
{
for(int i=0; i<moment1->rows * moment1->columns ; i++)
{
parameters->data[i] -= lr*moment1->data[i] / ( sqrtf(moment2->data[i]) + epsilon) ;
}
}
void initAdam(Adam* optimizer,NN* model, float lr ,float batch_size)
{
optimizer->model = model;
optimizer->lr = lr;
optimizer->batch_size = batch_size;
for(int i=0; i< optimizer->model->num_of_layers ; i++)
{
if(!optimizer->model->layers[i].is_trainable)
continue;
optimizer->moments[i].moment1_W = createMatrix(optimizer->model->layers[i].weight_gradients->rows, optimizer->model->layers[i].weight_gradients->columns , 0.0f);
optimizer->moments[i].moment1_b = createMatrix(optimizer->model->layers[i].bias_gradients->rows, optimizer->model->layers[i].bias_gradients->columns , 0.0f);
optimizer->moments[i].moment2_W = createMatrix(optimizer->model->layers[i].weight_gradients->rows, optimizer->model->layers[i].weight_gradients->columns , 0.0f);
optimizer->moments[i].moment2_b = createMatrix(optimizer->model->layers[i].bias_gradients->rows, optimizer->model->layers[i].bias_gradients->columns , 0.0f);
}
}
void optimizeAdam(Adam* optimizer)
{
for(int i=0; i< optimizer->model->num_of_layers ; i++)
{
if(!optimizer->model->layers[i].is_trainable)
continue;
//average the gradients
scaleMatrix( optimizer->model->layers[i].weight_gradients , 1.0/optimizer->batch_size);
scaleMatrix(optimizer->model->layers[i].bias_gradients , 1.0/optimizer->batch_size);
// calculate moments
calculate_moment1(optimizer->moments[i].moment1_W, optimizer->model->layers[i].weight_gradients );
calculate_moment1(optimizer->moments[i].moment1_b, optimizer->model->layers[i].bias_gradients);
calculate_moment2(optimizer->moments[i].moment2_W, optimizer->model->layers[i].weight_gradients );
calculate_moment2(optimizer->moments[i].moment2_b, optimizer->model->layers[i].bias_gradients );
// update parameters
update_parameters_Adam(optimizer->moments[i].moment1_W, optimizer->moments[i].moment2_W , optimizer->model->layers[i].weights, optimizer->lr);
update_parameters_Adam(optimizer->moments[i].moment1_b, optimizer->moments[i].moment2_b, optimizer->model->layers[i].biases, optimizer->lr);
//
clipMatrix(optimizer->model->layers[i].weights, -1.9,1.9);
clipMatrix(optimizer->model->layers[i].biases, -1.9,1.9);
}
}
void freeAdam(Adam* optimizer)
{
freeNN(optimizer->model);
for(int i=0; i<optimizer->model->num_of_layers ; i++)
{
freeMatrix(&optimizer->moments[i].moment1_W);
freeMatrix(&optimizer->moments[i].moment1_b);
freeMatrix(&optimizer->moments[i].moment2_W);
freeMatrix(&optimizer->moments[i].moment2_b);
}
}
#endif