tipe/src/cnn/cnn.c

332 lines
12 KiB
C
Raw Normal View History

#include <stdbool.h>
2022-06-30 10:27:42 +02:00
#include <stdlib.h>
#include <stdio.h>
2023-05-13 10:30:29 +02:00
#include <float.h>
#include <math.h>
2022-09-28 10:20:08 +02:00
2023-05-12 16:16:34 +02:00
#include "../common/include/memory_management.h"
#include "../common/include/colors.h"
#include "../common/include/utils.h"
2022-11-03 18:45:38 +01:00
#include "include/backpropagation.h"
2022-09-26 18:00:31 +02:00
#include "include/initialisation.h"
2022-10-24 12:54:51 +02:00
#include "include/function.h"
#include "include/creation.h"
2022-11-03 16:28:03 +01:00
#include "include/update.h"
2022-10-24 12:54:51 +02:00
#include "include/make.h"
2022-09-30 15:54:21 +02:00
#include "include/cnn.h"
2022-06-30 10:27:42 +02:00
2022-07-05 08:13:25 +02:00
// Augmente les dimensions de l'image d'entrée
#define PADDING_INPUT 2
2022-06-30 10:27:42 +02:00
2022-11-25 15:17:47 +01:00
int indice_max(float* tab, int n) {
int indice = -1;
2023-03-10 18:19:23 +01:00
float maxi = -FLT_MAX;
2023-01-17 15:34:29 +01:00
2022-11-25 15:17:47 +01:00
for (int i=0; i < n; i++) {
if (tab[i] > maxi) {
maxi = tab[i];
indice = i;
}
}
return indice;
}
2022-06-30 10:27:42 +02:00
int will_be_drop(int dropout_prob) {
return (rand() % 100) < dropout_prob;
2022-06-30 10:27:42 +02:00
}
void write_image_in_network_32(int** image, int height, int width, float** input, bool random_offset) {
int i_offset = 0;
int j_offset = 0;
int min_col = 0;
int min_ligne = 0;
if (random_offset) {
/*
<-- min_ligne
.%%:.
######%%%%%%%%%.
.:.:%##########:
. .... ##:
.##
##.
:##
.##.
:#%
%#.
:#%
.##.
##%
%##
##.
##:
:##.
.###.
:###
:#%
<-- max_ligne
^-- min_col
^-- max_col
*/
int sum_colonne[width];
int sum_ligne[height];
for (int i=0; i < width; i++) {
sum_colonne[i] = 0;
}
for (int j=0; j < height; j++) {
sum_ligne[j] = 0;
}
for (int i=0; i < width; i++) {
for (int j=0; j < height; j++) {
sum_ligne[i] += image[i][j];
sum_colonne[j] += image[i][j];
}
}
min_ligne = -1;
while (sum_ligne[min_ligne+1] == 0 && min_ligne < width+1) {
min_ligne++;
}
int max_ligne = width;
while (sum_ligne[max_ligne-1] == 0 && max_ligne > 0) {
max_ligne--;
}
min_col = -1;
while (sum_colonne[min_col+1] == 0 && min_col < height+1) {
min_col++;
}
int max_col = height;
while (sum_colonne[max_col-1] == 0 && max_col > 0) {
max_col--;
}
i_offset = rand()%(27-max_ligne+min_ligne);
j_offset = rand()%(27-max_col+min_col);
}
2022-10-07 14:26:36 +02:00
int padding = (32 - height)/2;
for (int i=0; i < padding; i++) {
for (int j=0; j < 32; j++) {
input[i][j] = 0.;
input[31-i][j] = 0.;
input[j][i] = 0.;
input[j][31-i] = 0.;
}
}
for (int i=0; i < width; i++) {
for (int j=0; j < height; j++) {
int adjusted_i = i + min_ligne - i_offset;
int adjusted_j = j + min_col - j_offset;
// Make sure not to be out of the image
input[i+2][j+2] = adjusted_i < height && adjusted_j < width && adjusted_i >= 0 && adjusted_j >= 0 ? (float)image[adjusted_i][adjusted_j] / 255.0f : 0.;
2022-06-30 10:27:42 +02:00
}
}
}
2022-11-19 16:09:07 +01:00
void write_image_in_network_260(unsigned char* image, int height, int width, float*** input) {
2023-02-19 13:43:09 +01:00
int size_input = 260;
int padding = (size_input - height)/2;
2022-11-19 16:09:07 +01:00
for (int i=0; i < padding; i++) {
2023-02-19 13:43:09 +01:00
for (int j=0; j < size_input; j++) {
2022-11-19 16:09:07 +01:00
for (int composante=0; composante < 3; composante++) {
input[composante][i][j] = 0.;
2023-02-19 13:43:09 +01:00
input[composante][size_input-1-i][j] = 0.;
2022-11-19 16:09:07 +01:00
input[composante][j][i] = 0.;
2023-02-19 13:43:09 +01:00
input[composante][j][size_input-1-i] = 0.;
2022-11-19 16:09:07 +01:00
}
}
}
for (int i=0; i < width; i++) {
for (int j=0; j < height; j++) {
for (int composante=0; composante < 3; composante++) {
input[composante][i+2][j+2] = (float)image[(i*height+j)*3 + composante] / 255.0f;
}
}
}
}
2022-06-30 10:27:42 +02:00
void forward_propagation(Network* network) {
2023-03-03 21:58:05 +01:00
int n = network->size; // Nombre de couches du réseau, il contient n-1 kernels
2022-09-19 18:39:49 +02:00
for (int i=0; i < n-1; i++) {
2023-03-03 21:58:05 +01:00
/*
* On procède kernel par kernel:
* On considère à chaque fois une couche d'entrée, une couche de sortie et le kernel qui contient les informations
* pour passer d'une couche à l'autre
*/
Kernel* k_i = network->kernel[i];
float*** input = network->input[i]; // Couche d'entrée
int input_depth = network->depth[i]; // Dimensions de la couche d'entrée
int input_width = network->width[i];
float*** output_z = network->input_z[i+1]; // Couche de sortie avant que la fonction d'activation ne lui soit appliquée
float*** output = network->input[i+1]; // Couche de sortie
int output_depth = network->depth[i+1]; // Dimensions de la couche de sortie
int output_width = network->width[i+1];
int activation = k_i->activation;
int pooling = k_i->pooling;
int stride = k_i->stride;
int padding = k_i->padding;
2022-09-19 18:39:49 +02:00
2022-11-04 12:02:00 +01:00
if (k_i->nn) {
drop_neurones(input, 1, 1, input_width, network->dropout);
} else {
drop_neurones(input, input_depth, input_width, input_width, network->dropout);
}
2023-03-03 21:58:05 +01:00
/*
* Pour chaque couche excepté le pooling, on propage les valeurs de la couche précédente,
* On copie les valeurs de output dans output_z, puis on applique la fonction d'activation à output_z
*/
2022-10-07 15:32:54 +02:00
if (k_i->cnn) { // Convolution
make_convolution(k_i->cnn, input, output, output_width, stride, padding);
2023-03-28 12:38:06 +02:00
copy_3d_array(output, output_z, output_depth, output_width, output_width);
2023-03-03 21:58:05 +01:00
apply_function_to_matrix(activation, output, output_depth, output_width);
2022-06-30 10:27:42 +02:00
}
2022-10-07 15:32:54 +02:00
else if (k_i->nn) { // Full connection
2023-03-08 20:48:34 +01:00
if (k_i->linearisation == DOESNT_LINEARISE) { // Vecteur -> Vecteur
2022-10-02 20:31:20 +02:00
make_dense(k_i->nn, input[0][0], output[0][0], input_width, output_width);
}
else { // Matrice -> Vecteur
2023-02-28 11:47:57 +01:00
make_dense_linearized(k_i->nn, input, output[0][0], input_depth, input_width, output_width);
2022-10-02 20:31:20 +02:00
}
2023-03-28 12:38:06 +02:00
copy_3d_array(output, output_z, 1, 1, output_width);
2023-03-03 21:58:05 +01:00
apply_function_to_vector(activation, output, output_width);
2022-06-30 10:27:42 +02:00
}
2022-10-02 20:31:20 +02:00
else { // Pooling
int kernel_size = 2*padding + input_width + stride - output_width*stride;
2023-03-03 21:58:05 +01:00
if (i == n-2) {
printf_error("Le réseau ne peut pas finir par un pooling layer\n");
2022-06-30 10:27:42 +02:00
return;
2022-10-02 20:31:20 +02:00
} else { // Pooling sur une matrice
2023-03-08 20:48:34 +01:00
if (pooling == AVG_POOLING) {
make_average_pooling(input, output, kernel_size, output_depth, output_width, stride, padding);
}
else if (pooling == MAX_POOLING) {
make_max_pooling(input, output, kernel_size, output_depth, output_width, stride, padding);
}
else {
2023-03-03 21:58:05 +01:00
printf_error("Impossible de reconnaître le type de couche de pooling: ");
printf("identifiant: %d, position: %d\n", pooling, i);
2023-01-30 09:39:45 +01:00
}
2022-06-30 10:27:42 +02:00
}
2023-03-28 12:38:06 +02:00
copy_3d_array(output, output_z, output_depth, output_width, output_width);
2022-05-13 15:28:45 +02:00
}
}
}
2022-06-30 10:27:42 +02:00
void backward_propagation(Network* network, int wanted_number) {
2023-03-03 21:58:05 +01:00
int n = network->size; // Nombre de couches du réseau
// Backward sur la dernière couche qui utilise toujours SOFTMAX
float* wanted_output = generate_wanted_output(wanted_number, network->width[network->size -1]); // Sortie désirée, permet d'initialiser une erreur
softmax_backward_cross_entropy(network->input[n-1][0][0], wanted_output, network->width[n-1]);
gree(wanted_output);
2022-10-26 18:27:46 +02:00
2023-03-03 21:58:05 +01:00
/*
* On propage à chaque étape:
* - les dérivées de l'erreur par rapport aux poids et biais, que l'on ajoute à ceux existants dans kernel->_->d_bias/d_weights
* - les dérivées de l'erreur par rapport à chaque case de input, qui servent uniquement à la propagation des informations.
* Ainsi, on écrase les valeurs contenues dans input, mais on utilise celles restantes dans input_z qui indiquent les valeurs avant
* la composition par la fonction d'activation pour pouvoir continuer à remonter.
*/
2022-11-03 17:50:11 +01:00
for (int i=n-2; i >= 0; i--) {
2022-10-02 20:31:20 +02:00
// Modifie 'k_i' à partir d'une comparaison d'informations entre 'input' et 'output'
2023-03-03 21:58:05 +01:00
Kernel* k_i = network->kernel[i];
float*** input = network->input[i];
float*** input_z = network->input_z[i];
int input_depth = network->depth[i];
int input_width = network->width[i];
float*** output = network->input[i+1];
int output_depth = network->depth[i+1];
int output_width = network->width[i+1];
int is_last_layer = i==0;
int activation = is_last_layer?SIGMOID:network->kernel[i-1]->activation;
2022-06-30 10:27:42 +02:00
2023-01-17 15:34:29 +01:00
2022-10-26 18:27:46 +02:00
if (k_i->cnn) { // Convolution
backward_convolution(k_i->cnn, input, input_z, output, input_depth, input_width, output_depth, output_width, -activation, is_last_layer);
2022-10-26 18:27:46 +02:00
} else if (k_i->nn) { // Full connection
2023-03-08 20:48:34 +01:00
if (k_i->linearisation == DOESNT_LINEARISE) { // Vecteur -> Vecteur
backward_dense(k_i->nn, input[0][0], input_z[0][0], output[0][0], input_width, output_width, -activation, is_last_layer);
2022-10-26 18:27:46 +02:00
} else { // Matrice -> vecteur
backward_linearisation(k_i->nn, input, input_z, output[0][0], input_depth, input_width, output_width, -activation);
2022-10-26 18:27:46 +02:00
}
} else { // Pooling
2023-03-08 20:48:34 +01:00
if (k_i->pooling == AVG_POOLING) {
backward_average_pooling(input, output, input_width, output_width, input_depth); // Depth pour input et output a la même valeur
} else {
2023-03-10 18:19:23 +01:00
backward_max_pooling(input, output, input_width, output_width, input_depth); // Depth pour input et output a la même valeur
2023-03-08 20:48:34 +01:00
}
2022-10-26 18:27:46 +02:00
}
2022-06-30 10:27:42 +02:00
}
}
2022-11-04 12:02:00 +01:00
void drop_neurones(float*** input, int depth, int dim1, int dim2, int dropout) {
2023-03-03 21:58:05 +01:00
for (int i=0; i < depth; i++) {
for (int j=0; j < dim1; j++) {
for (int k=0; k < dim2; k++) {
2022-11-04 12:02:00 +01:00
if (will_be_drop(dropout))
input[i][j][k] = 0;
}
}
}
}
2022-10-07 15:32:54 +02:00
float compute_mean_squared_error(float* output, float* wanted_output, int len) {
2023-03-03 21:58:05 +01:00
/*
* $E = \frac{ \sum_{i=0}^n (output_i - desired output_i)^2 }{n}$
*/
if (len == 0) {
printf_error("MSE: division par 0\n");
2022-10-07 15:32:54 +02:00
return 0.;
}
float loss=0.;
for (int i=0; i < len ; i++) {
loss += (output[i]-wanted_output[i])*(output[i]-wanted_output[i]);
}
return loss/len;
}
2022-06-30 10:27:42 +02:00
float compute_cross_entropy_loss(float* output, float* wanted_output, int len) {
float loss=0.;
for (int i=0; i < len ; i++) {
2022-06-30 10:27:42 +02:00
if (wanted_output[i]==1) {
if (output[i]==0.) {
loss -= log(FLT_EPSILON);
}
else {
loss -= log(output[i]);
}
}
}
2023-02-24 11:01:59 +01:00
return loss/len;
2022-06-30 10:27:42 +02:00
}
2023-01-17 15:34:29 +01:00
float* generate_wanted_output(int wanted_number, int size_output) {
float* wanted_output = (float*)nalloc(size_output, sizeof(float));
for (int i=0; i < size_output; i++) {
2022-06-30 10:27:42 +02:00
if (i==wanted_number) {
wanted_output[i]=1;
}
else {
wanted_output[i]=0;
}
}
return wanted_output;
}