merge

src/cnn/backpropagation.c

@@ -82,6 +82,7 @@ void backward_linearisation(Kernel_nn* ker, float*** input, float*** input_z, fl
 
     // Weights
     int cpt = 0;
+    int nb_elem = depth_input*dim_input*dim_input*size_output;
     for (int i=0; i < depth_input; i++) {
         for (int k=0; k < dim_input; k++) {
             for (int l=0; l < dim_input; l++) {

src/cnn/cnn.c

@@ -76,7 +76,7 @@ void write_image_in_network_260(unsigned char* image, int height, int width, flo
 }
 
 void forward_propagation(Network* network) {
-    int activation, input_depth, input_width, output_depth, output_width;
+    int activation, pooling, input_depth, input_width, output_depth, output_width;
     int n = network->size;
     float*** input;
     float*** output;
@@ -93,6 +93,7 @@ void forward_propagation(Network* network) {
         output_depth = network->depth[i+1];
         output_width = network->width[i+1];
         activation = k_i->activation;
+        pooling = k_i->pooling;
 
         if (k_i->nn) {
             drop_neurones(input, 1, 1, input_width, network->dropout);
@@ -119,7 +120,13 @@ void forward_propagation(Network* network) {
                 printf("Le réseau ne peut pas finir par une pooling layer\n");
                 return;
             } else { // Pooling sur une matrice
-                make_average_pooling(input, output, activation, output_depth, output_width);
+                if (pooling==1) {
+                    make_average_pooling(input, output, activation, output_depth, output_width);
+                } else if (pooling==2) {
+                    make_max_pooling(input, output, activation, output_depth, output_width);
+                } else {
+                    printf("Erreur dans la reconnaissance de la couche de pooling: %d,%d \n", pooling, i);
+                }
             }
             copy_input_to_input_z(output, output_a, output_depth, output_width, output_width);
         }

src/cnn/creation.c

@@ -24,6 +24,7 @@ Network* create_network(int max_size, float learning_rate, int dropout, int init
     for (int i=0; i < max_size-1; i++) {
         network->kernel[i] = (Kernel*)malloc(sizeof(Kernel));
     }
+    network->kernel[0]->linearisation = 0;
     network->width[0] = input_dim;
     network->depth[0] = input_depth;
     network->kernel[0]->nn = NULL;
@@ -36,7 +37,6 @@ Network* create_network(int max_size, float learning_rate, int dropout, int init
 Network* create_network_lenet5(float learning_rate, int dropout, int activation, int initialisation, int input_dim, int input_depth) {
     Network* network = create_network(8, learning_rate, dropout, initialisation, input_dim, input_depth);
     network->kernel[0]->activation = activation;
-    network->kernel[0]->linearisation = 0;
     add_convolution(network, 6, 28, activation);
     add_2d_average_pooling(network, 14);
     add_convolution(network, 16, 10, activation);
@@ -50,7 +50,6 @@ Network* create_network_lenet5(float learning_rate, int dropout, int activation,
 Network* create_simple_one(float learning_rate, int dropout, int activation, int initialisation, int input_dim, int input_depth) {
     Network* network = create_network(3, learning_rate, dropout, initialisation, input_dim, input_depth);
     network->kernel[0]->activation = activation;
-    network->kernel[0]->linearisation = 0;
     add_dense_linearisation(network, 80, activation);
     add_dense(network, 10, SOFTMAX);
     return network;
@@ -112,6 +111,29 @@ void add_2d_average_pooling(Network* network, int dim_output) {
     network->kernel[k_pos]->nn = NULL;
     network->kernel[k_pos]->activation = IDENTITY; // Ne contient pas de fonction d'activation
     network->kernel[k_pos]->linearisation = 0;
+    network->kernel[k_pos]->pooling = 1;
+    create_a_cube_input_layer(network, n, network->depth[n-1], network->width[n-1]/2);
+    create_a_cube_input_z_layer(network, n, network->depth[n-1], network->width[n-1]/2); // Will it be used ?
+    network->size++;
+}
+
+void add_2d_max_pooling(Network* network, int dim_output) {
+    int n = network->size;
+    int k_pos = n-1;
+    int dim_input = network->width[k_pos];
+    if (network->max_size == n) {
+        printf("Impossible de rajouter une couche de max pooling, le réseau est déjà plein\n");
+        return;
+    }
+    if (dim_input%dim_output != 0) {
+        printf("Erreur de dimension dans le max pooling\n");
+        return;
+    }
+    network->kernel[k_pos]->cnn = NULL;
+    network->kernel[k_pos]->nn = NULL;
+    network->kernel[k_pos]->activation = IDENTITY; // Ne contient pas de fonction d'activation
+    network->kernel[k_pos]->linearisation = 0;
+    network->kernel[k_pos]->pooling = 2;
     create_a_cube_input_layer(network, n, network->depth[n-1], network->width[n-1]/2);
     create_a_cube_input_z_layer(network, n, network->depth[n-1], network->width[n-1]/2); // Will it be used ?
     network->size++;
@@ -132,6 +154,7 @@ void add_convolution(Network* network, int depth_output, int dim_output, int act
     network->kernel[k_pos]->nn = NULL;
     network->kernel[k_pos]->activation = activation;
     network->kernel[k_pos]->linearisation = 0;
+    network->kernel[k_pos]->pooling = 0;
     network->kernel[k_pos]->cnn = (Kernel_cnn*)malloc(sizeof(Kernel_cnn));
     Kernel_cnn* cnn = network->kernel[k_pos]->cnn;
 
@@ -190,6 +213,7 @@ void add_dense(Network* network, int output_units, int activation) {
     Kernel_nn* nn = network->kernel[k_pos]->nn;
     network->kernel[k_pos]->activation = activation;
     network->kernel[k_pos]->linearisation = 0;
+    network->kernel[k_pos]->pooling = 0;
     nn->input_units = input_units;
     nn->output_units = output_units;
     nn->bias = (float*)malloc(sizeof(float)*output_units);
@@ -228,6 +252,7 @@ void add_dense_linearisation(Network* network, int output_units, int activation)
     Kernel_nn* nn = network->kernel[k_pos]->nn;
     network->kernel[k_pos]->activation = activation;
     network->kernel[k_pos]->linearisation = 1;
+    network->kernel[k_pos]->pooling = 0;
     nn->input_units = input_units;
     nn->output_units = output_units;
 

src/cnn/free.c

@@ -25,7 +25,7 @@ void free_a_line_input_layer(Network* network, int pos) {
     free(network->input_z[pos]);
 }
 
-void free_2d_average_pooling(Network* network, int pos) {
+void free_2d_pooling(Network* network, int pos) {
     free_a_cube_input_layer(network, pos+1, network->depth[pos+1], network->width[pos+1]);
 }
 
@@ -123,7 +123,7 @@ void free_network(Network* network) {
                 free_dense_linearisation(network, i);
             }
         } else { // Pooling
-            free_2d_average_pooling(network, i);
+            free_2d_pooling(network, i);
         }
     }
     free_network_creation(network);

src/cnn/include/creation.h

@@ -39,6 +39,11 @@ void create_a_line_input_layer(Network* network, int pos, int dim);
 */
 void add_2d_average_pooling(Network* network, int dim_output);
 
+/*
+* Ajoute au réseau une couche de max pooling valide de dimension dim*dim
+*/
+void add_2d_max_pooling(Network* network, int dim_output);
+
 /*
 * Ajoute au réseau une couche de convolution dim*dim et initialise les kernels
 */

src/cnn/include/free.h

@@ -16,9 +16,9 @@ void free_a_cube_input_layer(Network* network, int pos, int depth, int dim);
 void free_a_line_input_layer(Network* network, int pos);
 
 /*
-* Libère l'espace mémoire alloué dans 'add_2d_average_pooling' (creation.c)
+* Libère l'espace mémoire alloué dans 'add_2d_average_pooling' ou 'add_2d_max_pooling' (creation.c)
 */
-void free_2d_average_pooling(Network* network, int pos);
+void free_2d_pooling(Network* network, int pos);
 
 /*
 * Libère l'espace mémoire dans 'add_convolution' (creation.c)

src/cnn/include/make.h

@@ -12,11 +12,17 @@ void make_convolution_cpu(Kernel_cnn* kernel, float*** input, float*** output, i
 * Effectue la convolution sur le CPU ou GPU
 */
 void make_convolution(Kernel_cnn* kernel, float*** input, float*** output, int output_dim);
+
 /*
 * Effectue un average pooling avec stride=size
 */
 void make_average_pooling(float*** input, float*** output, int size, int output_depth, int output_dim);
 
+/*
+* Effectue un max pooling avec stride=size
+*/
+void make_max_pooling(float*** input, float*** output, int size, int output_depth, int output_dim);
+
 /*
 * Effectue une full connection
 */

src/cnn/include/struct.h

@@ -25,6 +25,7 @@ typedef struct Kernel {
     Kernel_nn* nn; // NULL si ce n'est pas un nn
     int activation; // Vaut l'identifiant de la fonction d'activation
     int linearisation; // Vaut 1 si c'est la linéarisation d'une couche, 0 sinon
+    int pooling; // 0 si pas pooling, 1 si average_pooling, 2 si max_pooling
 } Kernel;
 
 

src/cnn/include/train.h

@@ -6,8 +6,8 @@
 
 #define EPOCHS 10
 #define BATCHES 500
-#define USE_MULTITHREADING
-#define LEARNING_RATE 0.01
+//#define USE_MULTITHREADING
+#define LEARNING_RATE 0.05
 
 
 /*

src/cnn/make.c

@@ -1,9 +1,17 @@
 #include <stdio.h>
+#include <float.h>
 
 #include "../include/colors.h"
 #include "include/convolution.h"
 #include "include/make.h"
 
+float max_flt(float a, float b) {
+    // Return the max between the two floats
+    if (a>b) {
+        return a;
+    }
+    return b;
+}
 
 void make_average_pooling(float*** input, float*** output, int size, int output_depth, int output_dim) {
     // input[output_depth][output_dim+size-1][output_dim+size-1]
@@ -25,6 +33,25 @@ void make_average_pooling(float*** input, float*** output, int size, int output_
     }
 }
 
+void make_max_pooling(float*** input, float*** output, int size, int output_depth, int output_dim) {
+    // input[output_depth][output_dim+size-1][output_dim+size-1]
+    // output[output_depth][output_dim][output_dim]
+    float m;
+    for (int i=0; i < output_depth; i++) {
+        for (int j=0; j < output_dim; j++) {
+            for (int k=0; k < output_dim; k++) {
+                m = FLT_MIN;
+                for (int a=0; a < size; a++) {
+                    for (int b=0; b < size; b++) {
+                        m = max_flt(m, input[i][size*j +a][size*k +b]);
+                    }
+                }
+                output[i][j][k] = m;
+            }
+        }
+    }
+}
+
 void make_dense(Kernel_nn* kernel, float* input, float* output, int size_input, int size_output) {
     // input[size_input]
     // output[size_output]