Newer
Older
/**
* DESCRIPTION OF THE FILE
*
* @author Michal Kravčenko
* @date 30.7.18 -
*/
#include <random.hpp>
#include "NormalizedGradientDescent.h"
#include "message.h"
//#include "../mpi_wrapper.h"
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
namespace lib4neuro {
NormalizedGradientDescent::NormalizedGradientDescent(
double epsilon,
int max_iters,
size_t batch
) {
this->tolerance = epsilon;
this->maximum_niters = max_iters;
this->batch = batch;
}
NormalizedGradientDescent::~NormalizedGradientDescent() {
}
void NormalizedGradientDescent::optimize(lib4neuro::ErrorFunction& ef,
std::ofstream* ofs) {
double err_ = ef.eval();
COUT_INFO("Finding a solution via Normalized Gradient Descent method ..." << std::endl);
COUT_INFO("Initial error: " << err_ << std::endl);
if (ofs && ofs->is_open() && lib4neuro::mpi_rank == 0) {
*ofs << "Finding a solution via Normalized Gradient Descent method ..." << std::endl;
*ofs << "Initial error: " << err_ << std::endl;
}
double grad_norm = this->tolerance * 10.0;
double grad_norm_prev;
size_t i;
long long int iter_idx = this->maximum_niters;
size_t iter_counter = 0;
double prev_val, val = 0.0, c = 1.25;
size_t n_parameters = ef.get_dimension();
std::vector<double>* gradient_current(new std::vector<double>(n_parameters));
std::vector<double>* gradient_prev(new std::vector<double>(n_parameters));
std::vector<double>* params_current = new std::vector<double>(ef.get_parameters());
std::vector<double>* params_prev(new std::vector<double>(n_parameters));
std::vector<double>* ptr_mem;
std::fill(gradient_current->begin(),
gradient_current->end(),
0.0);
std::fill(gradient_prev->begin(),
gradient_prev->end(),
0.0);
val = err_;
prev_val = err_;
double total_time = -MPI_Wtime( );
double cooling = 1;
while (val > this->tolerance && (iter_idx != 0)) {
iter_idx--;
iter_counter++;
grad_norm_prev = grad_norm;
/* reset of the current gradient */
std::fill(gradient_current->begin(),
gradient_current->end(),
0.0);
ef.calculate_error_gradient_normalized(*params_current,
*gradient_current,
this->batch
);
/* Update of the parameters */
double scaling = 0.001*val/err_;
scaling = 0.00001 * cooling;
for (i = 0; i < gradient_current->size(); ++i) {
(*params_prev)[i] = (*params_current)[i] - scaling*(*gradient_current)[i];
}
prev_val = val;
val = ef.eval(params_prev);
/* switcheroo */
ptr_mem = gradient_prev;
gradient_prev = gradient_current;
gradient_current = ptr_mem;
ptr_mem = params_prev;
params_prev = params_current;
params_current = ptr_mem;
// grad_norm *= scaling;
COUT_INFO( " iteration " << iter_counter << ", direction norm " << scaling << ", error " << val );
// if( prev_val < val ){
// cooling *= 0.7;
// }
// else{
// cooling *= 1.05;
// }
//
// if( iter_counter % 500 == 0 ){
// cooling = 1;
// }
}
if (iter_idx == 0) {
COUT_INFO(std::endl << "Maximum number of iterations (" << this->maximum_niters
<< ") was reached! Final error: " << val << std::endl);
if (ofs && ofs->is_open() && lib4neuro::mpi_rank == 0) {
*ofs << "Maximum number of iterations (" << this->maximum_niters << ") was reached! Final error: "
<< val << std::endl;
}
} else {
COUT_INFO(std::endl << "Gradient Descent method converged after "
<< this->maximum_niters - iter_idx
<< " iterations. Final error:" << val
<< std::endl);
#ifdef L4N_DEBUG
if (ofs && ofs->is_open() && lib4neuro::mpi_rank == 0) {
*ofs << "Gradient Descent method converged after "
<< this->maximum_niters - iter_idx
<< " iterations."
<< std::endl;
}
#endif
}
total_time += MPI_Wtime( );
this->optimal_parameters = *params_current;
ef.set_parameters(this->optimal_parameters);
delete gradient_current;
delete gradient_prev;
delete params_current;
delete params_prev;
COUT_INFO( "Finished in " << total_time << " [s], in " << iter_counter << " iterations" );
COUT_INFO( " " << total_time / iter_counter << " [s] per iteration" );
COUT_INFO( " error " << val );
}
}