forked from larryniven/nn
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathcnn-frame.cc
198 lines (159 loc) · 7.85 KB
/
cnn-frame.cc
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
#include "nn/cnn-frame.h"
namespace cnn {
std::shared_ptr<tensor_tree::vertex> make_cnn_tensor_tree(int layer)
{
tensor_tree::vertex root { "nil" };
for (int i = 0; i < layer; ++i) {
tensor_tree::vertex conv { "nil" };
conv.children.push_back(tensor_tree::make_tensor("conv weight"));
conv.children.push_back(tensor_tree::make_tensor("conv bias"));
root.children.push_back(std::make_shared<tensor_tree::vertex>(conv));
}
return std::make_shared<tensor_tree::vertex>(root);
}
std::shared_ptr<tensor_tree::vertex> make_densenet_tensor_tree(int layer)
{
tensor_tree::vertex root { "nil" };
tensor_tree::vertex conv { "nil" };
for (int i = 0; i < layer; ++i) {
for (int j = 0; j < i + 1; ++j) {
conv.children.push_back(tensor_tree::make_tensor("conv"));
}
}
root.children.push_back(std::make_shared<tensor_tree::vertex>(conv));
return std::make_shared<tensor_tree::vertex>(root);
}
std::shared_ptr<tensor_tree::vertex> make_tensor_tree(cnn_t const& config)
{
tensor_tree::vertex root { "nil" };
for (int i = 0; i < config.layers.size(); ++i) {
if (config.layers[i].type == "conv") {
tensor_tree::vertex conv { "nil" };
conv.children.push_back(tensor_tree::make_tensor("conv weight"));
conv.children.push_back(tensor_tree::make_tensor("conv bias"));
root.children.push_back(std::make_shared<tensor_tree::vertex>(conv));
} else if (config.layers[i].type == "fc") {
tensor_tree::vertex fc { "nil" };
fc.children.push_back(tensor_tree::make_tensor("weight"));
fc.children.push_back(tensor_tree::make_tensor("bias"));
root.children.push_back(std::make_shared<tensor_tree::vertex>(fc));
} else if (config.layers[i].type == "framewise-fc") {
tensor_tree::vertex fc { "nil" };
fc.children.push_back(tensor_tree::make_tensor("weight"));
fc.children.push_back(tensor_tree::make_tensor("bias"));
root.children.push_back(std::make_shared<tensor_tree::vertex>(fc));
}
}
return std::make_shared<tensor_tree::vertex>(root);
}
cnn_t load_param(std::istream& is)
{
std::string line;
cnn_t result;
while (std::getline(is, line) && line != "#") {
if (ebt::startswith(line, "conv")) {
auto parts = ebt::split(line);
layer_t ell { "conv" };
assert(parts.size() == 5);
ell.data = std::make_shared<std::tuple<int, int, int, int>>(
std::make_tuple(std::stoi(parts[1]), std::stoi(parts[2]),
std::stoi(parts[4]), std::stoi(parts[4])));
result.layers.push_back(ell);
} else if (ebt::startswith(line, "max-pooling")) {
auto parts = ebt::split(line);
layer_t ell { "max-pooling" };
assert(parts.size() == 5);
ell.data = std::make_shared<std::tuple<int, int, int, int>>(
std::make_tuple(std::stoi(parts[1]), std::stoi(parts[2]),
std::stoi(parts[3]), std::stoi(parts[4])));
result.layers.push_back(ell);
} else if (ebt::startswith(line, "fc")) {
result.layers.push_back(layer_t { "fc" });
} else if (ebt::startswith(line, "framewise-fc")) {
result.layers.push_back(layer_t { "framewise-fc" });
} else if (ebt::startswith(line, "relu")) {
result.layers.push_back(layer_t { "relu" });
} else if (ebt::startswith(line, "logsoftmax")) {
result.layers.push_back(layer_t { "logsoftmax" });
} else {
throw std::logic_error("unable to parse: " + line);
}
}
result.param = make_tensor_tree(result);
tensor_tree::load_tensor(result.param, is);
return result;
}
void save_param(cnn_t& config, std::ostream& os)
{
for (int i = 0; i < config.layers.size(); ++i) {
auto& ell = config.layers[i];
if (ell.type == "conv") {
auto& t = *std::static_pointer_cast<std::tuple<int, int, int, int>>(ell.data);
os << "conv " << std::get<0>(t) << " " << std::get<1>(t)
<< " " << std::get<2>(t) << " " << std::get<3>(t) << std::endl;
} else if (ell.type == "max-pooling") {
std::tuple<int, int, int, int> t = *std::static_pointer_cast<
std::tuple<int, int, int, int>>(ell.data);
os << "max-pooling " << std::get<0>(t)
<< " " << std::get<1>(t) << " " << std::get<2>(t)
<< " " << std::get<3>(t) << std::endl;
} else if (ell.type == "fc") {
os << "fc" << std::endl;
} else if (ell.type == "framewise-fc") {
os << "framewise-fc" << std::endl;
} else if (ell.type == "relu") {
os << "relu" << std::endl;
} else if (ell.type == "logsoftmax") {
os << "logsoftmax" << std::endl;
} else {
throw std::logic_error("unable to parse: " + ell.type);
}
}
os << "#" << std::endl;
tensor_tree::save_tensor(config.param, os);
}
std::shared_ptr<transcriber>
make_transcriber(cnn_t const& config, double dropout, std::default_random_engine *gen)
{
cnn::multilayer_transcriber multi_trans;
for (int i = 0; i < config.layers.size(); ++i) {
auto& ell = config.layers[i];
if (ell.type == "conv") {
auto& d = *std::static_pointer_cast<std::tuple<int, int, int, int>>(ell.data);
auto t = std::make_shared<conv_transcriber>(
conv_transcriber { std::get<0>(d), std::get<1>(d),
std::get<2>(d), std::get<3>(d) });
multi_trans.layers.push_back(t);
} else if (ell.type == "max-pooling") {
auto& d = *std::static_pointer_cast<std::tuple<int, int, int, int>>(ell.data);
auto t = std::make_shared<max_pooling_transcriber>(
max_pooling_transcriber { std::get<0>(d), std::get<1>(d),
std::get<2>(d), std::get<3>(d) });
multi_trans.layers.push_back(t);
} else if (ell.type == "fc") {
auto t = std::make_shared<fc_transcriber>(fc_transcriber{});
multi_trans.layers.push_back(t);
if (dropout != 0.0) {
multi_trans.layers.push_back(std::make_shared<dropout_transcriber>(
dropout_transcriber {dropout, *gen}));
}
} else if (ell.type == "framewise-fc") {
auto t = std::make_shared<framewise_fc_transcriber>(framewise_fc_transcriber{});
multi_trans.layers.push_back(t);
if (dropout != 0.0) {
multi_trans.layers.push_back(std::make_shared<dropout_transcriber>(
dropout_transcriber {dropout, *gen}));
}
} else if (ell.type == "relu") {
auto t = std::make_shared<relu_transcriber>(relu_transcriber {});
multi_trans.layers.push_back(t);
} else if (ell.type == "logsoftmax") {
auto t = std::make_shared<logsoftmax_transcriber>(logsoftmax_transcriber {});
multi_trans.layers.push_back(t);
} else {
throw std::logic_error("unknown layer type: " + ell.type);
}
}
return std::make_shared<multilayer_transcriber>(multi_trans);
}
}