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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371
| #include <cmath>
#include <ctime>
#include <cassert>
#include <fstream>
#include <iostream>
#include <algorithm>
#include <boost/array.hpp>
#include <boost/random.hpp>
#include <boost/cstdlib.hpp>
#include <boost/tuple/tuple.hpp>
#include <boost/numeric/ublas/io.hpp>
#include <boost/numeric/ublas/vector.hpp>
#include <boost/numeric/ublas/matrix.hpp>
using namespace std;
using namespace boost;
using namespace boost::numeric;
const double learning_rate = 0.1; // define learning rate
const double lambda = 0.0; // define lambda for weight decay
const unsigned int mnbatch_sz = 1250; // define size of batch
const unsigned int epoc = -1; // define number of epoch
const unsigned int midsz = 200; // define number of hiden neurons
template <typename T = double>
struct sum {
sum(const T & init = T()) : value(init) { }
void operator()(const T & val) {
value += val;
}
T value;
};
struct tanh_nl {
template <typename T>
T operator()(const T & vec) {
T res(vec.size());
for (size_t i = 0, e = vec.size(); i != e; ++i) {
res(i) = tanh(vec(i));
}
return res;
}
};
struct softmax_nl {
template <typename T>
T operator()(const T & vec) {
T tmp(vec.size());
for (size_t i = 0, e = vec.size(); i != e; ++i) {
tmp(i) = exp(vec(i));
}
typename T::value_type exp_sum =
for_each(tmp.begin(), tmp.end(),
sum<typename T::value_type>(0)).value;
T res(vec.size());
for (size_t i = 0, e = vec.size(); i != e; ++i) {
res(i) = tmp(i) / exp_sum;
}
return res;
}
};
template <int Dim, int NbClass>
class MLP_tanh_softmax
{
public:
typedef ublas::vector<double> mlp_vec;
typedef ublas::matrix<double> mlp_mat;
typedef tuple<mlp_mat, mlp_mat, // W1, W2
mlp_vec, mlp_vec> mlp_params; // b1, b2
typedef array<mlp_vec, 4> fprop_vecs;
#define PARAMS_SIZE mlp_mat(midsz, Dim), mlp_mat(NbClass, midsz), \
mlp_vec(midsz), mlp_vec(NbClass)
#define INIT_MATS(x) do { \
get<0>((x)) = ublas::zero_matrix<double>(midsz, Dim); \
get<1>((x)) = ublas::zero_matrix<double>(NbClass, midsz); \
} while(false)
#define INIT_VECS(x) do { \
get<2>((x)) = ublas::zero_vector<double>(midsz); \
get<3>((x)) = ublas::zero_vector<double>(NbClass); \
} while(false)
MLP_tanh_softmax()
{
reset();
}
void reset()
{
params = mlp_params(PARAMS_SIZE);
mt19937 rng(static_cast<boost::uint32_t>(time(0)));
// Initialisation aléatoire des paramêtres de la couche cachée dans ]-1/Dim, 1/Dim[
// Où Dim est la taille des vecteurs d'entrée
double ini = 1.0 / sqrt(double(Dim));
uniform_real<> dist1(-ini, ini);
variate_generator<mt19937&, uniform_real<> > rand1(rng, dist1);
for (mlp_mat::array_type::iterator
it = get<0>(params).data().begin(),
end = get<0>(params).data().end();
it != end; ++it) {
*it = rand1();
}
// Initialisation aléatoire des paramêtres de la couche de sortie dans ]-1/midsz, 1/midsz[
//Où midsz est la taille de la couche cachée
ini = 1.0 / sqrt(double(midsz));
uniform_real<> dist2(-ini, ini);
variate_generator<mt19937&, uniform_real<> > rand2(rng, dist2);
for (mlp_mat::array_type::iterator
it = get<1>(params).data().begin(),
end = get<1>(params).data().end();
it != end; ++it) {
*it = rand2();
}
// Initialisation des biais à 0
INIT_VECS(params);
}
int test(const mlp_vec & vec)
{
fprop_vecs tmp;
forward_prop(vec, tmp);
return distance(tmp[3].begin(),
max_element(tmp[3].begin(), tmp[3].end()));
}
int test(const vector<mlp_vec> & test_set,
const vector<int> & classes,
size_t first, size_t size)
{
int res = 0;
for (size_t i = first, e = first + size; i != e; ++i) {
int pred = test(test_set[i]);
if (pred == classes[i])
++res;
}
return res;
}
// Entraine le réseau de neurones par mini batch jusqu'à ce que l'on trouve un minimum local
// (supposé atteind lorsque l'entrainement stagne pour stop_count epoques)
// L'ensemble de données est coupé en deux partie (chacune composés d'éléments contigues dans l'ensemble)
// La partie d'entrainement, sur laquel sera effectuée la rétropropagation, et la partie validation
// avec laquel on teste les performances (et donc l'évolution itérative).
static const unsigned stop_count = 5;
void train(const vector<mlp_vec> & data_set,
const vector<int> & classes,
size_t train_sz, size_t valid_sz,
ostream & ostr)
{
bool min_found = false; //< At least local one
double actual_min = 100.0;
unsigned int pos = 0, count = 0, egal_count = 0;
unsigned int i = 0;
while (i < epoc) {
if (pos == 0) { // One more epoc
static double last_error = 100.0;
ostr << "Iteration " << count++ << endl;
int tst = 0;
double error_percent = 0.0;
if (valid_sz != 0) {
tst = test(data_set, classes, 0, train_sz);
error_percent =
100.0 * (train_sz - tst) / double(train_sz);
ostr << "Erreur Entrainement : "
<< error_percent << endl;
tst = test(data_set, classes,
train_sz, valid_sz);
error_percent =
100.0 * (valid_sz - tst) / double(valid_sz);
ostr << "Erreur Validation : "
<< error_percent << endl;
ostr << endl;
}
if (last_error == error_percent) {
if (min_found && ++egal_count == stop_count)
break;
} else {
if (error_percent <= actual_min) {
min_found = false;
actual_min = 100.0;
}
if (error_percent > last_error) {
min_found = true;
actual_min = last_error;
}
egal_count = 0;
last_error = error_percent;
}
}
int mnbatch_size = min(mnbatch_sz, train_sz - pos);
mlp_params grad = calc_grad(data_set, classes,
pos, mnbatch_size);
get<0>(params) -= learning_rate * get<0>(grad);
get<1>(params) -= learning_rate * get<1>(grad);
get<2>(params) -= learning_rate * get<2>(grad);
get<3>(params) -= learning_rate * get<3>(grad);
if ((pos += mnbatch_size) > train_sz) {
++i;
pos = 0;
}
}
}
void dump(ostream & ostr)
{
ostr << "W1:\t" << get<0>(params)
<< "\nW2:\t" << get<1>(params)
<< "\nb1:\t" << get<2>(params)
<< "\nb2:\t" << get<3>(params)
<< endl;
}
private:
mlp_params calc_grad(const vector<mlp_vec> & train_set,
const vector<int> & classes,
size_t first,
unsigned int mnbatch_size)
{
mlp_params grad(PARAMS_SIZE);
INIT_MATS(grad); INIT_VECS(grad);
for (unsigned int i = 0; i < mnbatch_size; ++i) {
int index = first + i;
fprop_vecs tmp;
forward_prop(train_set[index], tmp);
mlp_params tmp_grad;
back_prop(train_set[index],
classes[index],
tmp, tmp_grad);
get<0>(grad) += get<0>(tmp_grad);
get<1>(grad) += get<1>(tmp_grad);
get<2>(grad) += get<2>(tmp_grad);
get<3>(grad) += get<3>(tmp_grad);
}
get<0>(grad) /= double(mnbatch_size);
get<1>(grad) /= double(mnbatch_size);
get<2>(grad) /= double(mnbatch_size);
get<3>(grad) /= double(mnbatch_size);
return grad;
}
void forward_prop(const mlp_vec & vec, fprop_vecs & res)
{
fprop_vecs ret;
ret[0] = prod(get<0>(params), vec) + get<2>(params);
ret[1] = mid_func(ret[0]);
ret[2] = prod(get<1>(params), ret[1]) + get<3>(params);
ret[3] = out_func(ret[2]);
#if !(defined NDEBUG)
cout << "Input vector:\t" << vec << endl;
cout << "Middle activ:\t" << ret[0] << endl;
cout << "Middle w/ nl:\t" << ret[1] << endl;
cout << "Output activ:\t" << ret[2] << endl;
cout << "Output w/ nl:\t" << ret[3] << endl;
#endif
swap(ret, res);
}
void back_prop(const mlp_vec & vec, int clas,
const fprop_vecs & fp_vecs,
mlp_params & res)
{
mlp_params ret(PARAMS_SIZE);
for (size_t i = 0, e = get<3>(ret).size(); i != e; ++i) {
// Caca beurk, mais évite un branchement... :-)
get<3>(ret)[i] = fp_vecs[3][i] - double(i == clas);
}
for (size_t i = 0, e = get<1>(ret).size1(); i != e; ++i) {
for (size_t j = 0, f = get<1>(ret).size2(); j != f; ++j) {
get<1>(ret)(i,j) = get<3>(ret)[i] * fp_vecs[1][j] +
2 * lambda * get<1>(params)(i,j);
}
}
mlp_vec dcdhs(fp_vecs[1].size());
for (size_t i = 0, e = dcdhs.size(); i != e; ++i) {
double sum = 0.0;
for (size_t j = 0, f = get<1>(params).size1(); j != f; ++j) {
sum += get<3>(ret)[j] * get<1>(params)(j,i);
}
dcdhs[i] = sum;
}
for (size_t i = 0, e = get<2>(ret).size(); i != e; ++i) {
get<2>(ret)[i] = dcdhs[i] *
(1 - fp_vecs[1][i] * fp_vecs[1][i]);
}
for (size_t i = 0, e = get<0>(ret).size1(); i != e; ++i) {
for (size_t j = 0, f = get<0>(ret).size2(); j != f; ++j) {
get<0>(ret)(i,j) = get<2>(ret)[i] * vec[j] +
2 * lambda * get<0>(params)(i,j);
}
}
swap(ret, res);
}
#undef PARAMS_SIZE
#undef INIT_MATS
#undef INIT_VECS
mlp_params params; // mlp_mat W1; mlp_mat W2; mlp_vec b1; mlp_vec b2;
tanh_nl mid_func;
softmax_nl out_func;
};
int main()
{
// Vecteurs de dimmension 784, 10 classes
typedef MLP_tanh_softmax<784, 10> MLPts;
MLPts mlp_ts = MLPts();
// Loading MNIST
cout << "Loading... "; cout.flush();
vector<MLPts::mlp_vec> data;
vector<int> classes;
ifstream train("mnist.txt");
size_t i = 0, count = 0;
double tmp;
while (train >> tmp && count < 30000) {
if (i == 784) {
classes.push_back(int(tmp));
i = 0;
++count;
continue;
} else if (i == 0) {
data.push_back(MLPts::mlp_vec(784));
}
data.back()[i++] = tmp;
}
cout << "Loaded. (" << data.size() << " items)" << endl;
ofstream output("res.txt");
mlp_ts.train(data, classes, 25000, 5000, output);
output << endl;
mlp_ts.dump(output);
return exit_success;
} |
Partager