-
Notifications
You must be signed in to change notification settings - Fork 22
Expand file tree
/
Copy pathOptimizers.cpp
More file actions
211 lines (175 loc) · 7.29 KB
/
Optimizers.cpp
File metadata and controls
211 lines (175 loc) · 7.29 KB
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
/*******************************************************
* Copyright (c) 2017, ArrayFire
* All rights reserved.
*
* This file is distributed under 3-clause BSD license.
* The complete license agreement can be obtained at:
* http://arrayfire.com/licenses/BSD-3-Clause
********************************************************/
#include <af/optim/Optimizers.hpp>
#include <cmath>
using af::autograd::Variable;
using std::vector;
// References:
// SGD and Momentum: http://cs231n.github.io/neural-networks-3/#sgd
// Adam: https://arxiv.org/pdf/1412.6980.pdf
// RMSProp: https://arxiv.org/pdf/1308.0850v5.pdf
// Comparision between various update rules:
// https://www.quora.com/What-are-differences-between-update-rules-like-AdaDelta-RMSProp-AdaGrad-and-AdaM
namespace af
{
namespace optim
{
Optimizer::Optimizer(const vector<Variable> ¶meters)
: m_parameters(parameters.begin(), parameters.end())
{
}
void Optimizer::zeroGrad()
{
for (auto ¶meter : m_parameters) {
parameter.zeroGrad();
}
}
SGDOptimizer::SGDOptimizer(const vector<Variable> ¶meters,
double learning_rate, double momentum,
double weight_decay, bool use_nesterov)
: Optimizer(parameters),
m_use_nesterov(use_nesterov),
m_lr(learning_rate),
m_mu(momentum),
m_wd(weight_decay),
m_velocities()
{
if (momentum != 0) {
m_velocities.reserve(parameters.size());
for (const auto ¶meter : m_parameters) {
m_velocities.push_back(af::constant(0, parameter.dims(), parameter.type()));
m_velocities.back().eval();
}
}
}
void SGDOptimizer::update()
{
for (size_t i = 0; i < m_parameters.size(); i++) {
const af::array &grad = m_parameters[i].grad().array();
af::array &data = m_parameters[i].array();
if (m_wd != 0) {
// Weight decay term
data = data - m_wd * data;
}
if (m_mu != 0) {
af::array &velocity = m_velocities[i];
// Regular momentum
velocity = m_mu * velocity - m_lr * grad;
if (m_use_nesterov) {
// Update for nesterov momentum
data = data + velocity * m_mu - m_lr * grad;
} else {
data = data + velocity;
}
af::eval(velocity, data);
} else {
data = data - m_lr * grad;
af::eval(data);
}
}
}
AdamOptimizer::AdamOptimizer(const vector<Variable> ¶meters,
double learning_rate,
double beta1, double beta2,
double epsilon, double weight_decay)
: Optimizer(parameters),
m_lr(learning_rate),
m_beta1(beta1),
m_beta2(beta2),
m_eps(epsilon),
m_wd(weight_decay),
m_count(0),
m_biased_first(),
m_biased_second()
{
m_biased_first.reserve(parameters.size());
m_biased_second.reserve(parameters.size());
for (const auto ¶meter : m_parameters) {
m_biased_first.push_back(af::constant(0, parameter.dims(), parameter.type()));
m_biased_second.push_back(af::constant(0, parameter.dims(), parameter.type()));
m_biased_first.back().eval();
m_biased_second.back().eval();
}
}
void AdamOptimizer::update()
{
for (size_t i = 0; i < m_parameters.size(); i++) {
const af::array &grad = m_parameters[i].grad().array();
af::array &data = m_parameters[i].array();
if (m_wd != 0) {
// Weight decay term
data = data - m_wd * data;
}
af::array &biased_first = m_biased_first[i];
af::array &biased_second = m_biased_second[i];
biased_first = m_beta1 * biased_first + (1 - m_beta1) * grad;
biased_second = m_beta2 * biased_second + (1 - m_beta2) * grad * grad;
m_count++;
double corrected_bias1 = 1 - std::pow(m_beta1, m_count);
double corrected_bias2 = 1 - std::pow(m_beta2, m_count);
double corrected_lr = m_lr * std::sqrt(corrected_bias2) / corrected_bias1;
data = data - (corrected_lr * biased_first) / (af::sqrt(biased_second) + m_eps);
af::eval(data, biased_first, biased_second);
}
}
RMSPropOptimizer::RMSPropOptimizer(const vector<Variable> ¶meters,
double learning_rate,
double rho,
double epsilon,
double weight_decay,
bool use_first)
: Optimizer(parameters),
m_use_first(use_first),
m_lr(learning_rate),
m_rho(rho),
m_eps(epsilon),
m_wd(weight_decay),
m_first(),
m_second()
{
if (m_use_first) m_first.reserve(parameters.size());
m_second.reserve(parameters.size());
for (const auto ¶meter : m_parameters) {
if (m_use_first) {
m_first.push_back(af::constant(0, parameter.dims(), parameter.type()));
m_first.back().eval();
}
m_second.push_back(af::constant(0, parameter.dims(), parameter.type()));
m_second.back().eval();
}
}
void RMSPropOptimizer::update()
{
for (size_t i = 0; i < m_parameters.size(); i++) {
const af::array &grad = m_parameters[i].grad().array();
af::array &data = m_parameters[i].array();
if (m_wd != 0) {
// Weight decay term
data = data - m_wd * data;
}
af::array &second = m_second[i];
second = m_rho * second + (1 - m_rho) * grad * grad;
// Create shallow copy of second so that we don't update "second" below
af::array moments = second;
if (m_use_first) {
af::array &first = m_first[i];
first = m_rho * first + (1 - m_rho) * grad;
moments = moments - first * first;
}
data = data - (m_lr * grad) / (af::sqrt(moments) + m_eps);
if (m_use_first) {
af::array &first = m_first[i];
af::eval(data, first, second);
} else {
af::eval(data, second);
}
}
}
}
}