mirror of
https://github.com/pjreddie/darknet.git
synced 2023-08-10 21:13:14 +03:00
IOU loss function
This commit is contained in:
Joseph Redmon
parent
feabcc31de
commit
989ab8c38a
@ -93,6 +93,7 @@ void visualize(char *cfgfile, char *weightfile)
|
||||
|
||||
int main(int argc, char **argv)
|
||||
{
|
||||
//test_box();
|
||||
//test_convolutional_layer();
|
||||
if(argc < 2){
|
||||
fprintf(stderr, "usage: %s <function>\n", argv[0]);
|
||||
|
||||
20
src/data.c
20
src/data.c
@ -65,22 +65,22 @@ matrix load_image_paths(char **paths, int n, int w, int h)
|
||||
return X;
|
||||
}
|
||||
|
||||
typedef struct box{
|
||||
typedef struct{
|
||||
int id;
|
||||
float x,y,w,h;
|
||||
float left, right, top, bottom;
|
||||
} box;
|
||||
} box_label;
|
||||
|
||||
box *read_boxes(char *filename, int *n)
|
||||
box_label *read_boxes(char *filename, int *n)
|
||||
{
|
||||
box *boxes = calloc(1, sizeof(box));
|
||||
box_label *boxes = calloc(1, sizeof(box_label));
|
||||
FILE *file = fopen(filename, "r");
|
||||
if(!file) file_error(filename);
|
||||
float x, y, h, w;
|
||||
int id;
|
||||
int count = 0;
|
||||
while(fscanf(file, "%d %f %f %f %f", &id, &x, &y, &w, &h) == 5){
|
||||
boxes = realloc(boxes, (count+1)*sizeof(box));
|
||||
boxes = realloc(boxes, (count+1)*sizeof(box_label));
|
||||
boxes[count].id = id;
|
||||
boxes[count].x = x;
|
||||
boxes[count].y = y;
|
||||
@ -97,11 +97,11 @@ box *read_boxes(char *filename, int *n)
|
||||
return boxes;
|
||||
}
|
||||
|
||||
void randomize_boxes(box *b, int n)
|
||||
void randomize_boxes(box_label *b, int n)
|
||||
{
|
||||
int i;
|
||||
for(i = 0; i < n; ++i){
|
||||
box swap = b[i];
|
||||
box_label swap = b[i];
|
||||
int index = rand_r(&data_seed)%n;
|
||||
b[i] = b[index];
|
||||
b[index] = swap;
|
||||
@ -114,7 +114,7 @@ void fill_truth_detection(char *path, float *truth, int classes, int num_boxes,
|
||||
labelpath = find_replace(labelpath, ".jpg", ".txt");
|
||||
labelpath = find_replace(labelpath, ".JPEG", ".txt");
|
||||
int count = 0;
|
||||
box *boxes = read_boxes(labelpath, &count);
|
||||
box_label *boxes = read_boxes(labelpath, &count);
|
||||
randomize_boxes(boxes, count);
|
||||
float x,y,w,h;
|
||||
float left, top, right, bot;
|
||||
@ -174,10 +174,10 @@ void fill_truth_detection(char *path, float *truth, int classes, int num_boxes,
|
||||
if(background) truth[index++] = 0;
|
||||
truth[index+id] = 1;
|
||||
index += classes;
|
||||
truth[index++] = y;
|
||||
truth[index++] = x;
|
||||
truth[index++] = h;
|
||||
truth[index++] = y;
|
||||
truth[index++] = w;
|
||||
truth[index++] = h;
|
||||
}
|
||||
free(boxes);
|
||||
}
|
||||
|
||||
@ -81,9 +81,9 @@ void train_detection(char *cfgfile, char *weightfile)
|
||||
if (imgnet){
|
||||
plist = get_paths("/home/pjreddie/data/imagenet/det.train.list");
|
||||
}else{
|
||||
//plist = get_paths("/home/pjreddie/data/voc/trainall.txt");
|
||||
plist = get_paths("/home/pjreddie/data/voc/trainall.txt");
|
||||
//plist = get_paths("/home/pjreddie/data/coco/trainval.txt");
|
||||
plist = get_paths("/home/pjreddie/data/voc/all2007-2012.txt");
|
||||
//plist = get_paths("/home/pjreddie/data/voc/all2007-2012.txt");
|
||||
}
|
||||
paths = (char **)list_to_array(plist);
|
||||
pthread_t load_thread = load_data_detection_thread(imgs, paths, plist->size, classes, net.w, net.h, side, side, background, &buffer);
|
||||
@ -95,12 +95,12 @@ void train_detection(char *cfgfile, char *weightfile)
|
||||
train = buffer;
|
||||
load_thread = load_data_detection_thread(imgs, paths, plist->size, classes, net.w, net.h, side, side, background, &buffer);
|
||||
|
||||
/*
|
||||
image im = float_to_image(net.w, net.h, 3, train.X.vals[114]);
|
||||
image copy = copy_image(im);
|
||||
draw_detection(copy, train.y.vals[114], 7);
|
||||
free_image(copy);
|
||||
*/
|
||||
/*
|
||||
image im = float_to_image(net.w, net.h, 3, train.X.vals[114]);
|
||||
image copy = copy_image(im);
|
||||
draw_detection(copy, train.y.vals[114], 7);
|
||||
free_image(copy);
|
||||
*/
|
||||
|
||||
printf("Loaded: %lf seconds\n", sec(clock()-time));
|
||||
time=clock();
|
||||
@ -120,30 +120,30 @@ void train_detection(char *cfgfile, char *weightfile)
|
||||
|
||||
void predict_detections(network net, data d, float threshold, int offset, int classes, int nuisance, int background, int num_boxes, int per_box)
|
||||
{
|
||||
matrix pred = network_predict_data(net, d);
|
||||
int j, k, class;
|
||||
for(j = 0; j < pred.rows; ++j){
|
||||
for(k = 0; k < pred.cols; k += per_box){
|
||||
float scale = 1.;
|
||||
int index = k/per_box;
|
||||
int row = index / num_boxes;
|
||||
int col = index % num_boxes;
|
||||
if (nuisance) scale = 1.-pred.vals[j][k];
|
||||
for (class = 0; class < classes; ++class){
|
||||
int ci = k+classes+background+nuisance;
|
||||
float y = (pred.vals[j][ci + 0] + row)/num_boxes;
|
||||
float x = (pred.vals[j][ci + 1] + col)/num_boxes;
|
||||
float h = pred.vals[j][ci + 2]; //* distance_from_edge(row, num_boxes);
|
||||
h = h*h;
|
||||
float w = pred.vals[j][ci + 3]; //* distance_from_edge(col, num_boxes);
|
||||
w = w*w;
|
||||
float prob = scale*pred.vals[j][k+class+background+nuisance];
|
||||
if(prob < threshold) continue;
|
||||
printf("%d %d %f %f %f %f %f\n", offset + j, class, prob, y, x, h, w);
|
||||
}
|
||||
matrix pred = network_predict_data(net, d);
|
||||
int j, k, class;
|
||||
for(j = 0; j < pred.rows; ++j){
|
||||
for(k = 0; k < pred.cols; k += per_box){
|
||||
float scale = 1.;
|
||||
int index = k/per_box;
|
||||
int row = index / num_boxes;
|
||||
int col = index % num_boxes;
|
||||
if (nuisance) scale = 1.-pred.vals[j][k];
|
||||
for (class = 0; class < classes; ++class){
|
||||
int ci = k+classes+background+nuisance;
|
||||
float y = (pred.vals[j][ci + 0] + row)/num_boxes;
|
||||
float x = (pred.vals[j][ci + 1] + col)/num_boxes;
|
||||
float h = pred.vals[j][ci + 2]; //* distance_from_edge(row, num_boxes);
|
||||
h = h*h;
|
||||
float w = pred.vals[j][ci + 3]; //* distance_from_edge(col, num_boxes);
|
||||
w = w*w;
|
||||
float prob = scale*pred.vals[j][k+class+background+nuisance];
|
||||
if(prob < threshold) continue;
|
||||
printf("%d %d %f %f %f %f %f\n", offset + j, class, prob, y, x, h, w);
|
||||
}
|
||||
}
|
||||
free_matrix(pred);
|
||||
}
|
||||
free_matrix(pred);
|
||||
}
|
||||
|
||||
void validate_detection(char *cfgfile, char *weightfile)
|
||||
|
||||
@ -3,7 +3,9 @@
|
||||
#include "softmax_layer.h"
|
||||
#include "blas.h"
|
||||
#include "cuda.h"
|
||||
#include "utils.h"
|
||||
#include <stdio.h>
|
||||
#include <string.h>
|
||||
#include <stdlib.h>
|
||||
|
||||
int get_detection_layer_locations(detection_layer layer)
|
||||
@ -26,6 +28,8 @@ detection_layer *make_detection_layer(int batch, int inputs, int classes, int co
|
||||
layer->coords = coords;
|
||||
layer->rescore = rescore;
|
||||
layer->nuisance = nuisance;
|
||||
layer->cost = calloc(1, sizeof(float));
|
||||
layer->does_cost=1;
|
||||
layer->background = background;
|
||||
int outputs = get_detection_layer_output_size(*layer);
|
||||
layer->output = calloc(batch*outputs, sizeof(float));
|
||||
@ -63,6 +67,169 @@ void dark_zone(detection_layer layer, int class, int start, network_state state)
|
||||
}
|
||||
}
|
||||
|
||||
typedef struct{
|
||||
float dx, dy, dw, dh;
|
||||
} dbox;
|
||||
|
||||
dbox derivative(box a, box b)
|
||||
{
|
||||
dbox d;
|
||||
d.dx = 0;
|
||||
d.dw = 0;
|
||||
float l1 = a.x - a.w/2;
|
||||
float l2 = b.x - b.w/2;
|
||||
if (l1 > l2){
|
||||
d.dx -= 1;
|
||||
d.dw += .5;
|
||||
}
|
||||
float r1 = a.x + a.w/2;
|
||||
float r2 = b.x + b.w/2;
|
||||
if(r1 < r2){
|
||||
d.dx += 1;
|
||||
d.dw += .5;
|
||||
}
|
||||
if (l1 > r2) {
|
||||
d.dx = -1;
|
||||
d.dw = 0;
|
||||
}
|
||||
if (r1 < l2){
|
||||
d.dx = 1;
|
||||
d.dw = 0;
|
||||
}
|
||||
|
||||
d.dy = 0;
|
||||
d.dh = 0;
|
||||
float t1 = a.y - a.h/2;
|
||||
float t2 = b.y - b.h/2;
|
||||
if (t1 > t2){
|
||||
d.dy -= 1;
|
||||
d.dh += .5;
|
||||
}
|
||||
float b1 = a.y + a.h/2;
|
||||
float b2 = b.y + b.h/2;
|
||||
if(b1 < b2){
|
||||
d.dy += 1;
|
||||
d.dh += .5;
|
||||
}
|
||||
if (t1 > b2) {
|
||||
d.dy = -1;
|
||||
d.dh = 0;
|
||||
}
|
||||
if (b1 < t2){
|
||||
d.dy = 1;
|
||||
d.dh = 0;
|
||||
}
|
||||
return d;
|
||||
}
|
||||
|
||||
float overlap(float x1, float w1, float x2, float w2)
|
||||
{
|
||||
float l1 = x1 - w1/2;
|
||||
float l2 = x2 - w2/2;
|
||||
float left = l1 > l2 ? l1 : l2;
|
||||
float r1 = x1 + w1/2;
|
||||
float r2 = x2 + w2/2;
|
||||
float right = r1 < r2 ? r1 : r2;
|
||||
return right - left;
|
||||
}
|
||||
|
||||
float box_intersection(box a, box b)
|
||||
{
|
||||
float w = overlap(a.x, a.w, b.x, b.w);
|
||||
float h = overlap(a.y, a.h, b.y, b.h);
|
||||
if(w < 0 || h < 0) return 0;
|
||||
float area = w*h;
|
||||
return area;
|
||||
}
|
||||
|
||||
float box_union(box a, box b)
|
||||
{
|
||||
float i = box_intersection(a, b);
|
||||
float u = a.w*a.h + b.w*b.h - i;
|
||||
return u;
|
||||
}
|
||||
|
||||
float box_iou(box a, box b)
|
||||
{
|
||||
return box_intersection(a, b)/box_union(a, b);
|
||||
}
|
||||
|
||||
dbox dintersect(box a, box b)
|
||||
{
|
||||
float w = overlap(a.x, a.w, b.x, b.w);
|
||||
float h = overlap(a.y, a.h, b.y, b.h);
|
||||
dbox dover = derivative(a, b);
|
||||
dbox di;
|
||||
|
||||
di.dw = dover.dw*h;
|
||||
di.dx = dover.dx*h;
|
||||
di.dh = dover.dh*w;
|
||||
di.dy = dover.dy*w;
|
||||
if(h < 0 || w < 0){
|
||||
di.dx = dover.dx;
|
||||
di.dy = dover.dy;
|
||||
}
|
||||
return di;
|
||||
}
|
||||
|
||||
dbox dunion(box a, box b)
|
||||
{
|
||||
dbox du = {0,0,0,0};;
|
||||
float w = overlap(a.x, a.w, b.x, b.w);
|
||||
float h = overlap(a.y, a.h, b.y, b.h);
|
||||
if(w > 0 && h > 0){
|
||||
dbox di = dintersect(a, b);
|
||||
du.dw = h - di.dw;
|
||||
du.dh = w - di.dw;
|
||||
du.dx = -di.dx;
|
||||
du.dy = -di.dy;
|
||||
}
|
||||
return du;
|
||||
}
|
||||
|
||||
dbox diou(box a, box b)
|
||||
{
|
||||
float u = box_union(a,b);
|
||||
float i = box_intersection(a,b);
|
||||
dbox di = dintersect(a,b);
|
||||
dbox du = dunion(a,b);
|
||||
dbox dd = {0,0,0,0};
|
||||
if(i < 0) {
|
||||
dd.dx = b.x - a.x;
|
||||
dd.dy = b.y - a.y;
|
||||
dd.dw = b.w - a.w;
|
||||
dd.dh = b.h - a.h;
|
||||
return dd;
|
||||
}
|
||||
dd.dx = 2*pow((1-(i/u)),1)*(di.dx*u - du.dx*i)/(u*u);
|
||||
dd.dy = 2*pow((1-(i/u)),1)*(di.dy*u - du.dy*i)/(u*u);
|
||||
dd.dw = 2*pow((1-(i/u)),1)*(di.dw*u - du.dw*i)/(u*u);
|
||||
dd.dh = 2*pow((1-(i/u)),1)*(di.dh*u - du.dh*i)/(u*u);
|
||||
return dd;
|
||||
}
|
||||
|
||||
void test_box()
|
||||
{
|
||||
box a = {1, 1, 1, 1};
|
||||
box b = {0, 0, .5, .2};
|
||||
int count = 0;
|
||||
while(count++ < 300){
|
||||
dbox d = diou(a, b);
|
||||
printf("%f %f %f %f\n", a.x, a.y, a.w, a.h);
|
||||
a.x += .1*d.dx;
|
||||
a.w += .1*d.dw;
|
||||
a.y += .1*d.dy;
|
||||
a.h += .1*d.dh;
|
||||
printf("inter: %f\n", box_intersection(a, b));
|
||||
printf("union: %f\n", box_union(a, b));
|
||||
printf("IOU: %f\n", box_iou(a, b));
|
||||
if(d.dx==0 && d.dw==0 && d.dy==0 && d.dh==0) {
|
||||
printf("break!!!\n");
|
||||
break;
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
void forward_detection_layer(const detection_layer layer, network_state state)
|
||||
{
|
||||
int in_i = 0;
|
||||
@ -92,31 +259,63 @@ void forward_detection_layer(const detection_layer layer, network_state state)
|
||||
layer.output[out_i++] = mask*state.input[in_i++];
|
||||
}
|
||||
}
|
||||
/*
|
||||
int count = 0;
|
||||
for(i = 0; i < layer.batch*locations; ++i){
|
||||
for(j = 0; j < layer.classes+layer.background; ++j){
|
||||
printf("%f, ", layer.output[count++]);
|
||||
}
|
||||
printf("\n");
|
||||
for(j = 0; j < layer.coords; ++j){
|
||||
printf("%f, ", layer.output[count++]);
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
*/
|
||||
/*
|
||||
if(layer.background || 1){
|
||||
if(layer.does_cost){
|
||||
*(layer.cost) = 0;
|
||||
int size = get_detection_layer_output_size(layer) * layer.batch;
|
||||
memset(layer.delta, 0, size * sizeof(float));
|
||||
for(i = 0; i < layer.batch*locations; ++i){
|
||||
int index = i*(layer.classes+layer.coords+layer.background);
|
||||
for(j= 0; j < layer.classes; ++j){
|
||||
if(state.truth[index+j+layer.background]){
|
||||
//dark_zone(layer, j, index, state);
|
||||
}
|
||||
int classes = layer.nuisance+layer.classes;
|
||||
int offset = i*(classes+layer.coords);
|
||||
for(j = offset; j < offset+classes; ++j){
|
||||
*(layer.cost) += pow(state.truth[j] - layer.output[j], 2);
|
||||
layer.delta[j] = state.truth[j] - layer.output[j];
|
||||
}
|
||||
box truth;
|
||||
truth.x = state.truth[j+0];
|
||||
truth.y = state.truth[j+1];
|
||||
truth.w = state.truth[j+2];
|
||||
truth.h = state.truth[j+3];
|
||||
box out;
|
||||
out.x = layer.output[j+0];
|
||||
out.y = layer.output[j+1];
|
||||
out.w = layer.output[j+2];
|
||||
out.h = layer.output[j+3];
|
||||
if(!(truth.w*truth.h)) continue;
|
||||
float iou = box_iou(truth, out);
|
||||
//printf("iou: %f\n", iou);
|
||||
*(layer.cost) += pow((1-iou), 2);
|
||||
dbox d = diou(out, truth);
|
||||
layer.delta[j+0] = d.dx;
|
||||
layer.delta[j+1] = d.dy;
|
||||
layer.delta[j+2] = d.dw;
|
||||
layer.delta[j+3] = d.dh;
|
||||
}
|
||||
}
|
||||
*/
|
||||
/*
|
||||
int count = 0;
|
||||
for(i = 0; i < layer.batch*locations; ++i){
|
||||
for(j = 0; j < layer.classes+layer.background; ++j){
|
||||
printf("%f, ", layer.output[count++]);
|
||||
}
|
||||
printf("\n");
|
||||
for(j = 0; j < layer.coords; ++j){
|
||||
printf("%f, ", layer.output[count++]);
|
||||
}
|
||||
printf("\n");
|
||||
}
|
||||
*/
|
||||
/*
|
||||
if(layer.background || 1){
|
||||
for(i = 0; i < layer.batch*locations; ++i){
|
||||
int index = i*(layer.classes+layer.coords+layer.background);
|
||||
for(j= 0; j < layer.classes; ++j){
|
||||
if(state.truth[index+j+layer.background]){
|
||||
//dark_zone(layer, j, index, state);
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
*/
|
||||
}
|
||||
|
||||
void backward_detection_layer(const detection_layer layer, network_state state)
|
||||
@ -164,6 +363,7 @@ void forward_detection_layer_gpu(const detection_layer layer, network_state stat
|
||||
cpu_state.input = in_cpu;
|
||||
forward_detection_layer(layer, cpu_state);
|
||||
cuda_push_array(layer.output_gpu, layer.output, layer.batch*outputs);
|
||||
cuda_push_array(layer.delta_gpu, layer.delta, layer.batch*outputs);
|
||||
free(cpu_state.input);
|
||||
if(cpu_state.truth) free(cpu_state.truth);
|
||||
}
|
||||
|
||||
@ -11,6 +11,8 @@ typedef struct {
|
||||
int background;
|
||||
int rescore;
|
||||
int nuisance;
|
||||
int does_cost;
|
||||
float *cost;
|
||||
float *output;
|
||||
float *delta;
|
||||
#ifdef GPU
|
||||
|
||||
@ -47,7 +47,7 @@ void train_imagenet(char *cfgfile, char *weightfile)
|
||||
printf("%d: %f, %f avg, %lf seconds, %d images\n", i, loss, avg_loss, sec(clock()-time), net.seen);
|
||||
free_data(train);
|
||||
//if(i%100 == 0 && net.learning_rate > .00001) net.learning_rate *= .97;
|
||||
if(i%100==0){
|
||||
if(i%1000==0){
|
||||
char buff[256];
|
||||
sprintf(buff, "/home/pjreddie/imagenet_backup/%s_%d.weights",base, i);
|
||||
save_weights(net, buff);
|
||||
|
||||
@ -186,6 +186,9 @@ float get_network_cost(network net)
|
||||
if(net.types[net.n-1] == COST){
|
||||
return ((cost_layer *)net.layers[net.n-1])->output[0];
|
||||
}
|
||||
if(net.types[net.n-1] == DETECTION){
|
||||
return ((detection_layer *)net.layers[net.n-1])->cost[0];
|
||||
}
|
||||
return 0;
|
||||
}
|
||||
|
||||
|
||||
@ -36,5 +36,9 @@ float variance_array(float *a, int n);
|
||||
float mag_array(float *a, int n);
|
||||
float **one_hot_encode(float *a, int n, int k);
|
||||
float sec(clock_t clocks);
|
||||
|
||||
typedef struct{
|
||||
float x, y, w, h;
|
||||
} box;
|
||||
#endif
|
||||
|
||||
|
||||
Loading…
Reference in New Issue
Block a user