-
Notifications
You must be signed in to change notification settings - Fork 0
/
GeneralizedLinearModel.h
322 lines (252 loc) · 8.2 KB
/
GeneralizedLinearModel.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
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
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
#pragma once
#include <random>
#include <fstream>
#include <iostream>
#include <sstream>
#include "ActivationFunctions.h"
#include "CostFunctions.h"
#include "WeightsInitializer.h"
#include "GradientSolvers.h"
namespace GLM {
template<class DerivedClass, typename InputType = Eigen::VectorXd, typename OutputType = Eigen::VectorXd, typename WeightsType = Eigen::MatrixXd, class Solver = SGD::AdamWSolver<>, class BatchInputType = Eigen::MatrixXd, class BatchOutputType = BatchInputType>
class GeneralizedLinearModelBase
{
public:
GeneralizedLinearModelBase(int szi = 1, int szo = 1)
: inputs(szi), outputs(szo)
{
}
virtual ~GeneralizedLinearModelBase() = default;
virtual OutputType Predict(const InputType& input)
{
return solver.activationFunction(W * input + b);
}
virtual BatchOutputType AddBatchWithParamsAdjusment(const BatchInputType& batchInput, const BatchOutputType& batchOutput)
{
static_cast<DerivedClass*>(this)->AddBatchNoParamsAdjustment(batchInput, batchOutput);
return solver.getWeightsAndBias(W, b);
}
const BatchOutputType& getPrediction() const
{
return solver.getPrediction();
}
void setPrediction(const BatchOutputType& p)
{
solver.setPrediction(p);
}
const BatchInputType& getInput() const
{
return solver.getInput();
}
double getLoss() const
{
return solver.getLoss();
}
double getLoss(const BatchOutputType& prediction, const BatchOutputType& target) const
{
return solver.getLoss(prediction, target);
}
int getNrInputs() const
{
return inputs;
}
int getNrOutputs() const
{
return outputs;
}
Solver& getSolver()
{
return solver;
}
bool saveModel(std::ofstream& os) const
{
try
{
os << solver.activationFunction.getName() << std::endl;
os << solver.lossFunction.getName() << std::endl;
os << inputs << " " << outputs << std::endl;
const static Eigen::IOFormat csv(Eigen::FullPrecision, Eigen::DontAlignCols, ", ", "\n");
os << W.format(csv) << std::endl;
os << b.format(csv) << std::endl;
}
catch (...)
{
return false;
}
return true;
}
bool loadModel(std::ifstream& is)
{
try
{
std::string activationFunctionName;
is >> activationFunctionName;
if (activationFunctionName != solver.activationFunction.getName())
{
std::cout << "Activation function mismatch: " << activationFunctionName << " vs " << solver.activationFunction.getName() << std::endl;
return false;
}
std::string costFunctionName;
is >> costFunctionName;
if (costFunctionName != solver.lossFunction.getName())
{
std::cout << "Cost function mismatch: " << costFunctionName << " vs " << solver.lossFunction.getName() << std::endl;
return false;
}
is >> inputs >> outputs;
is.ignore();
std::cout << "Loading parameters for " << inputs << " inputs and " << outputs << " outputs" << std::endl;
W.resize(outputs, inputs);
b.resize(outputs);
std::string matRow;
int row = 0;
while (getline(is, matRow))
{
std::stringstream matRowStrstr(matRow);
std::string field;
int col = 0;
while (getline(matRowStrstr, field, ','))
{
W(row, col) = stod(field);
++col;
}
++row;
if (row == outputs)
{
//std::cout << "Done reading W" << std::endl;
//std::cout << W << std::endl;
break;
}
}
row = 0;
std::string field;
while (getline(is, field))
{
b(row) = stod(field);
++row;
if (row == outputs)
{
//std::cout << "Done reading b" << std::endl;
//std::cout << b << std::endl;
break;
}
}
}
catch (...)
{
std::cout << "Exception thrown while loading model" << std::endl;
return false;
}
return true;
}
protected:
int inputs;
int outputs;
WeightsType W;
OutputType b;
Solver solver;
};
template<typename InputType = Eigen::VectorXd, typename OutputType = Eigen::VectorXd, typename WeightsType = Eigen::MatrixXd, class Solver = SGD::AdamWSolver<>, class BatchInputType = Eigen::MatrixXd, class BatchOutputType = BatchInputType>
class GeneralizedLinearModel : public GeneralizedLinearModelBase<GeneralizedLinearModel<InputType, OutputType, WeightsType, Solver, BatchInputType, BatchOutputType>, InputType, OutputType, WeightsType, Solver, BatchInputType, BatchOutputType>
{
public:
using BaseType = GeneralizedLinearModelBase<GeneralizedLinearModel<InputType, OutputType, WeightsType, Solver, BatchInputType, BatchOutputType>, InputType, OutputType, WeightsType, Solver, BatchInputType, BatchOutputType>;
GeneralizedLinearModel(int szi = 1, int szo = 1) : BaseType(szi, szo)
{
Initialize(szi, szo);
}
void Initialize(Initializers::WeightsInitializerInterface& initializer)
{
for (int j = 0; j < BaseType::W.cols(); ++j)
for (int i = 0; i < BaseType::W.rows(); ++i)
BaseType::W(i, j) = initializer.get(BaseType::getNrInputs(), BaseType::getNrOutputs());
}
void AddBatchNoParamsAdjustment(const BatchInputType& batchInput, const BatchOutputType& batchOutput)
{
BaseType::solver.AddBatch(batchInput, batchOutput);
BatchOutputType pred(batchOutput.rows(), batchOutput.cols());
BatchOutputType linpred(batchOutput.rows(), batchOutput.cols());
for (unsigned int i = 0; i < batchInput.cols(); ++i)
{
linpred.col(i) = BaseType::W * batchInput.col(i) + BaseType::b;
pred.col(i) = BaseType::solver.activationFunction(linpred.col(i));
}
BaseType::solver.setLinearPrediction(linpred);
BaseType::solver.setPrediction(pred);
}
BatchInputType BackpropagateBatch(const BatchOutputType& grad) const
{
InputType firstCol = BackpropagateGradient(grad.col(0));
BatchInputType res(firstCol.size(), grad.cols());
res.col(0) = firstCol;
for (int i = 1; i < grad.cols(); ++i)
res.col(i) = BackpropagateGradient(grad.col(i));
return res;
}
protected:
void Initialize(int szi = 1, int szo = 1)
{
BaseType::solver.Initialize(szi, szo);
//W = WeightsType::Random(szo, szi);
BaseType::W.resize(szo, szi);
// Eigen has a Random generator (random between -1 and 1 by default), but for now I'll stick with this one:
std::random_device rd;
std::mt19937 rde(rd());
const double x = 1. / sqrt(szi);
std::uniform_real_distribution<> dist(-x, x);
for (int i = 0; i < szo; ++i)
for (int j = 0; j < szi; ++j)
BaseType::W(i, j) = dist(rde);
BaseType::b = OutputType::Zero(szo);
}
InputType BackpropagateGradient(const OutputType& grad) const
{
return BaseType::W.transpose() * grad;
}
};
template<class Solver>
class GeneralizedLinearModel<double, double, double, Solver, Eigen::RowVectorXd> : public GeneralizedLinearModelBase<GeneralizedLinearModel<double, double, double, Solver, Eigen::RowVectorXd>, double, double, double, Solver, Eigen::RowVectorXd>
{
public:
using BaseType = GeneralizedLinearModelBase<GeneralizedLinearModel<double, double, double, Solver, Eigen::RowVectorXd>, double, double, double, Solver, Eigen::RowVectorXd>;
GeneralizedLinearModel(int szi = 1, int szo = 1) : BaseType(szi, szo)
{
Initialize(szi, szo);
}
void Initialize(Initializers::WeightsInitializerInterface& initializer)
{
BaseType::W = initializer.get();
}
void AddBatchNoParamsAdjustment(const Eigen::RowVectorXd& batchInput, const Eigen::RowVectorXd& batchOutput)
{
BaseType::solver.AddBatch(batchInput, batchOutput);
Eigen::RowVectorXd pred(batchOutput.cols());
Eigen::RowVectorXd linpred(batchOutput.cols());
for (unsigned int i = 0; i < batchInput.cols(); ++i)
{
linpred(i) = BaseType::W * batchInput(i) + BaseType::b;
pred(i) = BaseType::solver.activationFunction(linpred(i));
}
BaseType::solver.setLinearPrediction(linpred);
BaseType::solver.setPrediction(pred);
}
Eigen::RowVectorXd BackpropagateBatch(const Eigen::RowVectorXd& grad) const
{
Eigen::RowVectorXd res(grad.size());
for (int i = 0; i < grad.size(); ++i)
res(i) = BackpropagateGradient(grad(i));
return res;
}
protected:
void Initialize(int szi = 1, int szo = 1)
{
BaseType::solver.Initialize(szi, szo);
BaseType::W = 0;
BaseType::b = 0;
}
double BackpropagateGradient(const double& grad) const
{
return BaseType::W * grad;
}
};
}