-
Notifications
You must be signed in to change notification settings - Fork 0
/
Network.hpp
196 lines (164 loc) · 5.49 KB
/
Network.hpp
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
#pragma once
#include <Layer.hpp>
#include <fstream>
#include <memory>
#include <cmath>
#include <random>
#include <vector>
namespace yolo {
class Policy {
public:
virtual float get_current_rate_ratio(size_t current_batch) const = 0;
virtual ~Policy() {}
};
class ConstantPolicy : public Policy {
public:
float get_current_rate_ratio(size_t /* current_batch */) const override {
return 1; //legacy value
}
};
class StepPolicy : public Policy {
public:
float get_current_rate_ratio(size_t current_batch) const override {
return std::pow(_scale, current_batch / _step);
}
private:
int _step = 1;
float _scale = 1.;
};
class StepsPolicy : public Policy {
public:
struct Step {
size_t rank;
float scale;
Step(size_t aRank, float aScale) : rank(aRank), scale(aScale) {}
};
StepsPolicy(const std::vector<Step> &steps) : _steps(steps) {}
float get_current_rate_ratio(size_t current_batch) const override {
float ratio = 1;
for (const auto &step : _steps) {
if (step.rank > current_batch)
return ratio;
ratio *= step.scale;
}
return ratio;
}
private:
std::vector<Step> _steps;
};
class ExpPolicy : public Policy {
public:
float get_current_rate_ratio(size_t current_batch) const override {
return std::pow(_gamma, current_batch);
}
private:
float _gamma = 1;
};
class PolyPolicy : public Policy {
public:
float get_current_rate_ratio(size_t /*current_batch*/) const override {
throw "Don't know how to implement this...";
}
};
class RandomPolicy : public Policy {
public:
float get_current_rate_ratio(size_t /* current_batch */) const override {
static std::random_device rd;
static std::mt19937 gen(rd());
throw "Not correclty implemented";
return std::pow(std::uniform_real_distribution<>(0., 1.)(gen), 1 /* FIXME what to do of this? net.power */);
}
};
class SigPolicy : public Policy {
public:
float get_current_rate_ratio(size_t current_batch) const override {
return 1. / (1. + std::exp(_gamma * (current_batch - _step)));
}
private:
int _step = 1;
float _gamma = 1;
};
class Network {
public:
int _batch = 1;
size_t _subdivisions = 1;
Size _input_size; // FIXME name not really clear...
size_t _channels = 0;
float _momentum = .9;
float _decay = .0001;
float _angle = 0;
float _saturation = 1;
float _exposure = 1;
float _hue = 0;
float _learning_rate = .001;
size_t _burn_in = 0;
size_t _max_batches = 0;
void setPolicy(std::unique_ptr<Policy> policy) {
_policy = std::move(policy);
}
void addRoute(const std::vector<int> &layers_idx) {
std::vector<Layer *> route_layers; // FIXME use weak_ptr
for (const auto &idx : layers_idx) {
Layer *l;
if (idx >= 0) {
l =_layers[idx].get();
} else { // negative indexes reference layers starting from back
l =_layers[_layers.size() + idx].get();
}
route_layers.push_back(l);
}
addLayer(std::make_unique<RouteLayer>(route_layers));
}
void addLayer(std::unique_ptr<Layer> layer) {
/* For the first layer we give the network input size; for next ones,
* each layer N gets the output size of the layer N - 1 */
const auto &format = _layers.empty() ? Format(_input_size.width, _input_size.height, _channels, _batch) : _layers.back()->getOutputFormat();
layer->setInputFormat(format);
std::cout << std::setw(4) << _layers.size() << std::setw(15) << layer->getName()
<< std::setw(4) << format.width << " x " << std::setw(4) << format.height << " x " << std::setw(4) << format.channels
<< " -> "
<< std::setw(4) << layer->getOutputSize().width << " x " << std::setw(4) << layer->getOutputSize().height << " x " << std::setw(4)
<< layer->getOutputChannels() << std::endl;
_layers.push_back(std::move(layer));
}
void loadWeights(std::istream &in) {
int major, minor, revision;
in.read((char *)&major, sizeof(major))
.read((char *)&minor, sizeof(minor))
.read((char *)&revision, sizeof(revision));
if ((major * 10 + minor) >= 2) {
in.read((char *)&_seen, sizeof(_seen));
} else {
int iseen = 0;
in.read((char *)&iseen, sizeof(iseen));
_seen = iseen;
}
for (auto &layer : _layers) {
layer->loadWeights(in);
}
}
auto predict(const std::vector<float> &_input, float thresh) {
if (_input.size() != _input_size.height * _input_size.width * _channels) {
throw std::invalid_argument("Actual input size does not match netork size");
}
auto *input = &_input;
size_t idx = 0;
for (auto &layer : _layers) {
std::cout << "Processing layer " << idx++ << " ..." << std::flush;
input = &layer->forward(*input);
std::cout << " done" << std::endl;
}
const auto *regionlayer = dynamic_cast<RegionLayer *>(_layers.back().get());
if (regionlayer != nullptr)
return regionlayer->get_region_boxes(thresh);
const auto *detectionlayer = dynamic_cast<DetectionLayer *>(_layers.back().get());
if (detectionlayer != nullptr)
return detectionlayer->get_boxes(thresh);
throw std::invalid_argument("Dont know which detection type you are using.");
}
private:
std::vector<std::unique_ptr<Layer>> _layers;
std::unique_ptr<Policy> _policy;
size_t _seen;
};
}