2 * mlp-mnist is an implementation of a multi-layer neural network.
4 * Copyright (c) 2008 Idiap Research Institute, http://www.idiap.ch/
5 * Written by Francois Fleuret <francois.fleuret@idiap.ch>
7 * This file is part of mlp-mnist.
9 * mlp-mnist is free software: you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License version 3 as
11 * published by the Free Software Foundation.
13 * mlp-mnist is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
18 * You should have received a copy of the GNU General Public License
19 * along with mlp-mnist. If not, see <http://www.gnu.org/licenses/>.
25 // 2-layer NN, 300 hidden units, mean square error 4.70%
26 // 2-layer NN, 1000 hidden units 4.50%
27 // 3-layer NN, 300+100 hidden units 3.05%
28 // 3-layer NN, 500+150 hidden units 2.95%
30 /*********************************************************************
32 This program, trained on 20,000 (+ 20,000 for the stopping
33 criterion), tested on the 10,000 of the MNIST test set 100 hidden
34 neurons, basic network, 3.48%
38 ./ann --nb-training-examples 20000 --nb-validation-examples 20000 \
39 --mlp-structure 784,200,10 \
40 --data-files ${DATA_DIR}/train-images-idx3-ubyte ${DATA_DIR}/train-labels-idx1-ubyte \
45 ./ann --load-mlp simple.mlp \
46 --data-files ${DATA_DIR}/t10k-images-idx3-ubyte ${DATA_DIR}/t10k-labels-idx1-ubyte \
47 --nb-test-examples 10000
49 *********************************************************************/
63 #define SMALL_BUFFER_SIZE 1024
65 //////////////////////////////////////////////////////////////////////
67 //////////////////////////////////////////////////////////////////////
69 int nb_experiment = 0;
70 int nb_training_examples = 0;
71 int nb_validation_examples = 0;
72 int nb_test_examples = 0;
73 bool save_data = false;
75 char images_filename[SMALL_BUFFER_SIZE] = "\0";
76 char labels_filename[SMALL_BUFFER_SIZE] = "\0";
77 char opt_load_filename[SMALL_BUFFER_SIZE] = "\0";
78 char opt_save_filename[SMALL_BUFFER_SIZE] = "\0";
79 char opt_layer_sizes[SMALL_BUFFER_SIZE] = "\0";
81 char *next_word(char *buffer, char *r, int buffer_size) {
88 while((*r != '"') && (*r != '\0') &&
89 (s<buffer+buffer_size-1))
93 while((*r != '\r') && (*r != '\n') && (*r != '\0') &&
94 (*r != '\t') && (*r != ' ') && (*r != ',') &&
95 (s<buffer+buffer_size-1))
99 while((*r == ' ') || (*r == '\t') || (*r == ',')) r++;
100 if((*r == '\0') || (*r=='\r') || (*r=='\n')) r = NULL;
106 //////////////////////////////////////////////////////////////////////
107 // Simple routine to check we have enough parameters
108 //////////////////////////////////////////////////////////////////////
110 void check_opt(int argc, char **argv, int n_opt, int n, const char *help) {
111 if(n_opt + n >= argc) {
112 cerr << "Missing argument for " << argv[n_opt] << ".\n";
113 cerr << "Expecting " << help << ".\n";
118 void print_help_and_exit(int e) {
119 cout << "ANN. Written by François Fleuret.\n";
120 cout << "$Id: ann.cc,v 1.1 2005-12-13 17:19:11 fleuret Exp $\n";
125 int main(int argc, char **argv) {
127 if(argc == 1) print_help_and_exit(1);
131 // Parsing the command line parameters ///////////////////////////////
137 if(argc == 1 || strcmp(argv[i], "--help") == 0) print_help_and_exit(0);
139 else if(strcmp(argv[i], "--data-files") == 0) {
140 check_opt(argc, argv, i, 2, "<string: pixel filename> <string: label filename>");
141 strncpy(images_filename, argv[i+1], SMALL_BUFFER_SIZE);
142 strncpy(labels_filename, argv[i+2], SMALL_BUFFER_SIZE);
146 else if(strcmp(argv[i], "--load-mlp") == 0) {
147 check_opt(argc, argv, i, 1, "<string: mlp filename>");
148 strncpy(opt_load_filename, argv[i+1], SMALL_BUFFER_SIZE);
152 else if(strcmp(argv[i], "--mlp-structure") == 0) {
153 check_opt(argc, argv, i, 1, "<int: input layer size>,<int: first hidden layer size>,[...,]<int: output layer size>");
154 strncpy(opt_layer_sizes, argv[i+1], SMALL_BUFFER_SIZE);
158 else if(strcmp(argv[i], "--save-mlp") == 0) {
159 check_opt(argc, argv, i, 1, "<string: mlp filename>");
160 strncpy(opt_save_filename, argv[i+1], SMALL_BUFFER_SIZE);
164 else if(strcmp(argv[i], "--nb-experiment") == 0) {
165 check_opt(argc, argv, i, 1, "<int: number of the experiment>");
166 nb_experiment = atoi(argv[i+1]);
170 else if(strcmp(argv[i], "--nb-training-examples") == 0) {
171 check_opt(argc, argv, i, 1, "<int: number of examples for the training>");
172 nb_training_examples = atoi(argv[i+1]);
176 else if(strcmp(argv[i], "--nb-validation-examples") == 0) {
177 check_opt(argc, argv, i, 1, "<int: number of examples for the validation>");
178 nb_validation_examples = atoi(argv[i+1]);
182 else if(strcmp(argv[i], "--nb-test-examples") == 0) {
183 check_opt(argc, argv, i, 1, "<int: number of examples for the test>");
184 nb_test_examples = atoi(argv[i+1]);
188 else if(strcmp(argv[i], "--save-data") == 0) {
194 cerr << "Unknown option " << argv[i] << "\n";
195 print_help_and_exit(1);
200 cout << "Loading the data file ..."; cout.flush();
201 image_set.load_mnist_format(images_filename, labels_filename);
202 cout << " done.\n"; cout.flush();
204 cout << "Database contains " << image_set.nb_pics()
205 << " images of resolution " << image_set.width() << "x" << image_set.height()
206 << " divided into " << image_set.nb_obj() << " objects.\n";
208 srand48(nb_experiment);
211 int *layer_sizes = 0;
213 if(opt_layer_sizes[0]) {
214 char *s = opt_layer_sizes;
215 char token[SMALL_BUFFER_SIZE];
216 while(s) { s = next_word(token, s, SMALL_BUFFER_SIZE); nb_layers++; }
219 cerr << "Need at least two layers.\n";
223 layer_sizes = new int[nb_layers];
226 while(s) { s = next_word(token, s, SMALL_BUFFER_SIZE); layer_sizes[n++] = atoi(token); }
229 // Loading or creating a perceptron from scratch /////////////////////
231 MultiLayerPerceptron *mlp = 0;
233 if(opt_load_filename[0]) {
235 ifstream stream(opt_load_filename);
237 cerr << "Can not read " << opt_load_filename << ".\n";
241 cout << "Loading network " << opt_load_filename << " ... "; cout.flush();
242 mlp = new MultiLayerPerceptron(stream);
243 cout << "done (layers of sizes";
244 for(int l = 0; l < mlp->nb_layers(); l++) cout << " " << mlp->layer_size(l);
245 cout << ")\n"; cout.flush();
247 } else if(nb_layers > 0) {
249 if(layer_sizes[0] != image_set.width() * image_set.height() ||
250 layer_sizes[nb_layers-1] != image_set.nb_obj()) {
251 cerr << "For this data set, the input layer has to be of size " << image_set.width() * image_set.height() << ",\n";
252 cerr << "and the output has to be of size " << image_set.nb_obj() << ".\n";
256 cout << "Creating a new network (layers of sizes";
257 for(int i = 0; i < nb_layers; i++) cout << " " << layer_sizes[i];
260 mlp = new MultiLayerPerceptron(nb_layers, layer_sizes);
261 mlp->init_random_weights(1e-1);
264 // Training the perceptron ///////////////////////////////////////////
266 ImageSet training_set, validation_set, test_set;
268 if(nb_training_examples > 0)
269 training_set.extract_unused_pictures(image_set, nb_training_examples);
271 if(nb_validation_examples > 0)
272 validation_set.extract_unused_pictures(image_set, nb_validation_examples);
274 if(save_data && mlp) mlp->save_data();
276 if(nb_training_examples > 0) {
277 if(validation_set.nb_pics() == 0) {
278 cerr << "We need validation pictures for training.\n";
281 cout << "Training the network with " << nb_training_examples << " training and " << nb_validation_examples << " validation examples.\n"; cout.flush();
282 mlp->train(&training_set, &validation_set);
285 // Saving the perceptron /////////////////////////////////////////////
287 if(opt_save_filename[0]) {
289 cerr << "No perceptron to save.\n";
293 ofstream stream(opt_save_filename);
295 cerr << "Can not write " << opt_save_filename << ".\n";
299 cout << "Saving network " << opt_save_filename << " ... "; cout.flush();
301 cout << "done.\n"; cout.flush();
304 // Testing the perceptron ////////////////////////////////////////////
306 if(nb_test_examples > 0) {
307 test_set.extract_unused_pictures(image_set, nb_test_examples);
308 cout << "Error rate " << mlp->error(&test_set) << " (" << mlp->classification_error(&test_set)*100 << "%)\n";
310 // This is to test the analytical gradient
311 // scalar_t gradient[mlp->nb_weights()], numerical_gradient[mlp->nb_weights()];
312 // mlp->compute_gradient(&test_set, gradient);
313 // mlp->compute_numerical_gradient(&test_set, numerical_gradient);
314 // for(int i = 0; i < mlp->nb_weights(); i++) cout << "TEST " << gradient[i] << " " << numerical_gradient[i] << "\n";
317 // Flushing the log //////////////////////////////////////////////////
319 delete[] layer_sizes;