From ae1768e5831caa95214b93b08ee711aede36df07 Mon Sep 17 00:00:00 2001
From: AlexeyAB <alexeyab84@gmail.com>
Date: Mon, 05 Mar 2018 20:26:09 +0000
Subject: [PATCH] Removed random=1 from resnet152_yolo.cfg. Until resize_network() isn't supported for [shortcut] layer
---
src/detector.c | 206 ++++++++++++++++++++++++++++++++++++++++++++++++++-
1 files changed, 202 insertions(+), 4 deletions(-)
diff --git a/src/detector.c b/src/detector.c
index d7cc8a0..f8515d4 100644
--- a/src/detector.c
+++ b/src/detector.c
@@ -10,7 +10,9 @@
#ifdef OPENCV
#include "opencv2/highgui/highgui_c.h"
#include "opencv2/core/core_c.h"
+//#include "opencv2/core/core.hpp"
#include "opencv2/core/version.hpp"
+#include "opencv2/imgproc/imgproc_c.h"
#ifndef CV_VERSION_EPOCH
#include "opencv2/videoio/videoio_c.h"
@@ -23,11 +25,13 @@
#pragma comment(lib, "opencv_highgui" OPENCV_VERSION ".lib")
#endif
-#endif
+IplImage* draw_train_chart(float max_img_loss, int max_batches, int number_of_lines, int img_size);
+void draw_train_loss(IplImage* img, int img_size, float avg_loss, float max_img_loss, int current_batch, int max_batches);
+#endif // OPENCV
static int coco_ids[] = {1,2,3,4,5,6,7,8,9,10,11,13,14,15,16,17,18,19,20,21,22,23,24,25,27,28,31,32,33,34,35,36,37,38,39,40,41,42,43,44,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,67,70,72,73,74,75,76,77,78,79,80,81,82,84,85,86,87,88,89,90};
-void train_detector(char *datacfg, char *cfgfile, char *weightfile, int *gpus, int ngpus, int clear)
+void train_detector(char *datacfg, char *cfgfile, char *weightfile, int *gpus, int ngpus, int clear, int dont_show)
{
list *options = read_data_cfg(datacfg);
char *train_images = option_find_str(options, "train", "data/train.list");
@@ -92,6 +96,15 @@
args.saturation = net.saturation;
args.hue = net.hue;
+#ifdef OPENCV
+ IplImage* img = NULL;
+ float max_img_loss = 5;
+ int number_of_lines = 100;
+ int img_size = 1000;
+ if (!dont_show)
+ img = draw_train_chart(max_img_loss, net.max_batches, number_of_lines, img_size);
+#endif //OPENCV
+
pthread_t load_thread = load_data(args);
clock_t time;
int count = 0;
@@ -157,6 +170,12 @@
i = get_current_batch(net);
printf("%d: %f, %f avg, %f rate, %lf seconds, %d images\n", get_current_batch(net), loss, avg_loss, get_current_rate(net), sec(clock()-time), i*imgs);
+
+#ifdef OPENCV
+ if(!dont_show)
+ draw_train_loss(img, img_size, avg_loss, max_img_loss, i, net.max_batches);
+#endif // OPENCV
+
//if (i % 1000 == 0 || (i < 1000 && i % 100 == 0)) {
if (i % 100 == 0) {
#ifdef GPU
@@ -174,6 +193,9 @@
char buff[256];
sprintf(buff, "%s/%s_final.weights", backup_directory, base);
save_weights(net, buff);
+
+ //cvReleaseImage(&img);
+ //cvDestroyAllWindows();
}
@@ -772,7 +794,7 @@
}
}
}
- //printf("point = %d, cur_recall = %.4f, cur_precision = %.4f \n", point, cur_recall, cur_precision);
+ //printf("class_id = %d, point = %d, cur_recall = %.4f, cur_precision = %.4f \n", i, point, cur_recall, cur_precision);
avg_precision += cur_precision;
}
@@ -804,6 +826,177 @@
fprintf(stderr, "Total Detection Time: %f Seconds\n", (double)(time(0) - start));
}
+#ifdef OPENCV
+void calc_anchors(char *datacfg, int num_of_clusters, int final_width, int final_height, int show)
+{
+ printf("\n num_of_clusters = %d, final_width = %d, final_height = %d \n", num_of_clusters, final_width, final_height);
+
+ //float pointsdata[] = { 1,1, 2,2, 6,6, 5,5, 10,10 };
+ float *rel_width_height_array = calloc(1000, sizeof(float));
+
+ list *options = read_data_cfg(datacfg);
+ char *train_images = option_find_str(options, "train", "data/train.list");
+ list *plist = get_paths(train_images);
+ int number_of_images = plist->size;
+ char **paths = (char **)list_to_array(plist);
+
+ int number_of_boxes = 0;
+ printf(" read labels from %d images \n", number_of_images);
+
+ int i, j;
+ for (i = 0; i < number_of_images; ++i) {
+ char *path = paths[i];
+ char labelpath[4096];
+ find_replace(path, "images", "labels", labelpath);
+ find_replace(labelpath, "JPEGImages", "labels", labelpath);
+ find_replace(labelpath, ".jpg", ".txt", labelpath);
+ find_replace(labelpath, ".JPEG", ".txt", labelpath);
+ find_replace(labelpath, ".png", ".txt", labelpath);
+ int num_labels = 0;
+ box_label *truth = read_boxes(labelpath, &num_labels);
+ //printf(" new path: %s \n", labelpath);
+ for (j = 0; j < num_labels; ++j)
+ {
+ number_of_boxes++;
+ rel_width_height_array = realloc(rel_width_height_array, 2 * number_of_boxes * sizeof(float));
+ rel_width_height_array[number_of_boxes * 2 - 2] = truth[j].w * final_width;
+ rel_width_height_array[number_of_boxes * 2 - 1] = truth[j].h * final_height;
+ printf("\r loaded \t image: %d \t box: %d", i+1, number_of_boxes);
+ }
+ }
+ printf("\n all loaded. \n");
+
+ CvMat* points = cvCreateMat(number_of_boxes, 2, CV_32FC1);
+ CvMat* centers = cvCreateMat(num_of_clusters, 2, CV_32FC1);
+ CvMat* labels = cvCreateMat(number_of_boxes, 1, CV_32SC1);
+
+ for (i = 0; i < number_of_boxes; ++i) {
+ points->data.fl[i * 2] = rel_width_height_array[i * 2];
+ points->data.fl[i * 2 + 1] = rel_width_height_array[i * 2 + 1];
+ //cvSet1D(points, i * 2, cvScalar(rel_width_height_array[i * 2], 0, 0, 0));
+ //cvSet1D(points, i * 2 + 1, cvScalar(rel_width_height_array[i * 2 + 1], 0, 0, 0));
+ }
+
+
+ const int attemps = 10;
+ double compactness;
+
+ enum {
+ KMEANS_RANDOM_CENTERS = 0,
+ KMEANS_USE_INITIAL_LABELS = 1,
+ KMEANS_PP_CENTERS = 2
+ };
+
+ printf("\n calculating k-means++ ...");
+ // Should be used: distance(box, centroid) = 1 - IoU(box, centroid)
+ cvKMeans2(points, num_of_clusters, labels,
+ cvTermCriteria(CV_TERMCRIT_EPS+CV_TERMCRIT_ITER, 10000, 0), attemps,
+ 0, KMEANS_PP_CENTERS,
+ centers, &compactness);
+
+ //orig 2.0 anchors = 1.08,1.19, 3.42,4.41, 6.63,11.38, 9.42,5.11, 16.62,10.52
+ //float orig_anch[] = { 1.08,1.19, 3.42,4.41, 6.63,11.38, 9.42,5.11, 16.62,10.52 };
+ // worse than ours (even for 19x19 final size - for input size 608x608)
+
+ //orig anchors = 1.3221,1.73145, 3.19275,4.00944, 5.05587,8.09892, 9.47112,4.84053, 11.2364,10.0071
+ //float orig_anch[] = { 1.3221,1.73145, 3.19275,4.00944, 5.05587,8.09892, 9.47112,4.84053, 11.2364,10.0071 };
+ // orig (IoU=59.90%) better than ours (59.75%)
+
+ //gen_anchors.py = 1.19, 1.99, 2.79, 4.60, 4.53, 8.92, 8.06, 5.29, 10.32, 10.66
+ //float orig_anch[] = { 1.19, 1.99, 2.79, 4.60, 4.53, 8.92, 8.06, 5.29, 10.32, 10.66 };
+
+ // ours: anchors = 9.3813,6.0095, 3.3999,5.3505, 10.9476,11.1992, 5.0161,9.8314, 1.5003,2.1595
+ //float orig_anch[] = { 9.3813,6.0095, 3.3999,5.3505, 10.9476,11.1992, 5.0161,9.8314, 1.5003,2.1595 };
+ //for (i = 0; i < num_of_clusters * 2; ++i) centers->data.fl[i] = orig_anch[i];
+
+ //for (i = 0; i < number_of_boxes; ++i)
+ // printf("%2.2f,%2.2f, ", points->data.fl[i * 2], points->data.fl[i * 2 + 1]);
+
+ float avg_iou = 0;
+ for (i = 0; i < number_of_boxes; ++i) {
+ float box_w = points->data.fl[i * 2];
+ float box_h = points->data.fl[i * 2 + 1];
+ //int cluster_idx = labels->data.i[i];
+ int cluster_idx = 0;
+ float min_dist = 1000000;
+ for (j = 0; j < num_of_clusters; ++j) {
+ float anchor_w = centers->data.fl[j * 2];
+ float anchor_h = centers->data.fl[j * 2 + 1];
+ float w_diff = anchor_w - box_w;
+ float h_diff = anchor_h - box_h;
+ float distance = sqrt(w_diff*w_diff + h_diff*h_diff);
+ if (distance < min_dist) min_dist = distance, cluster_idx = j;
+ }
+
+ float anchor_w = centers->data.fl[cluster_idx * 2];
+ float anchor_h = centers->data.fl[cluster_idx * 2 + 1];
+ float min_w = (box_w < anchor_w) ? box_w : anchor_w;
+ float min_h = (box_h < anchor_h) ? box_h : anchor_h;
+ float box_intersect = min_w*min_h;
+ float box_union = box_w*box_h + anchor_w*anchor_h - box_intersect;
+ float iou = box_intersect / box_union;
+ if (iou > 1 || iou < 0) {
+ printf(" i = %d, box_w = %d, box_h = %d, anchor_w = %d, anchor_h = %d, iou = %f \n",
+ i, box_w, box_h, anchor_w, anchor_h, iou);
+ }
+ else avg_iou += iou;
+ }
+ avg_iou = 100 * avg_iou / number_of_boxes;
+ printf("\n avg IoU = %2.2f %% \n", avg_iou);
+
+ char buff[1024];
+ FILE* fw = fopen("anchors.txt", "wb");
+ printf("\nSaving anchors to the file: anchors.txt \n");
+ printf("anchors = ");
+ for (i = 0; i < num_of_clusters; ++i) {
+ sprintf(buff, "%2.4f,%2.4f", centers->data.fl[i * 2], centers->data.fl[i * 2 + 1]);
+ printf("%s, ", buff);
+ fwrite(buff, sizeof(char), strlen(buff), fw);
+ if (i + 1 < num_of_clusters) fwrite(", ", sizeof(char), 2, fw);;
+ }
+ printf("\n");
+ fclose(fw);
+
+ if (show) {
+ size_t img_size = 700;
+ IplImage* img = cvCreateImage(cvSize(img_size, img_size), 8, 3);
+ cvZero(img);
+ for (j = 0; j < num_of_clusters; ++j) {
+ CvPoint pt1, pt2;
+ pt1.x = pt1.y = 0;
+ pt2.x = centers->data.fl[j * 2] * img_size / final_width;
+ pt2.y = centers->data.fl[j * 2 + 1] * img_size / final_height;
+ cvRectangle(img, pt1, pt2, CV_RGB(255, 255, 255), 1, 8, 0);
+ }
+
+ for (i = 0; i < number_of_boxes; ++i) {
+ CvPoint pt;
+ pt.x = points->data.fl[i * 2] * img_size / final_width;
+ pt.y = points->data.fl[i * 2 + 1] * img_size / final_height;
+ int cluster_idx = labels->data.i[i];
+ int red_id = (cluster_idx * (uint64_t)123 + 55) % 255;
+ int green_id = (cluster_idx * (uint64_t)321 + 33) % 255;
+ int blue_id = (cluster_idx * (uint64_t)11 + 99) % 255;
+ cvCircle(img, pt, 1, CV_RGB(red_id, green_id, blue_id), CV_FILLED, 8, 0);
+ //if(pt.x > img_size || pt.y > img_size) printf("\n pt.x = %d, pt.y = %d \n", pt.x, pt.y);
+ }
+ cvShowImage("clusters", img);
+ cvWaitKey(0);
+ cvReleaseImage(&img);
+ cvDestroyAllWindows();
+ }
+
+ free(rel_width_height_array);
+ cvReleaseMat(&points);
+ cvReleaseMat(¢ers);
+ cvReleaseMat(&labels);
+}
+#else
+void calc_anchors(char *datacfg, int num_of_clusters, int final_width, int final_height, int show) {
+ printf(" k-means++ can't be used without OpenCV, because there is used cvKMeans2 implementation \n");
+}
+#endif // OPENCV
+
void test_detector(char *datacfg, char *cfgfile, char *weightfile, char *filename, float thresh, int dont_show)
{
list *options = read_data_cfg(datacfg);
@@ -870,12 +1063,16 @@
void run_detector(int argc, char **argv)
{
int dont_show = find_arg(argc, argv, "-dont_show");
+ int show = find_arg(argc, argv, "-show");
int http_stream_port = find_int_arg(argc, argv, "-http_port", -1);
char *out_filename = find_char_arg(argc, argv, "-out_filename", 0);
char *prefix = find_char_arg(argc, argv, "-prefix", 0);
float thresh = find_float_arg(argc, argv, "-thresh", .24);
int cam_index = find_int_arg(argc, argv, "-c", 0);
int frame_skip = find_int_arg(argc, argv, "-s", 0);
+ int num_of_clusters = find_int_arg(argc, argv, "-num_of_clusters", 5);
+ int final_width = find_int_arg(argc, argv, "-final_width", 13);
+ int final_heigh = find_int_arg(argc, argv, "-final_heigh", 13);
if(argc < 4){
fprintf(stderr, "usage: %s %s [train/test/valid] [cfg] [weights (optional)]\n", argv[0], argv[1]);
return;
@@ -912,10 +1109,11 @@
if (weights[strlen(weights) - 1] == 0x0d) weights[strlen(weights) - 1] = 0;
char *filename = (argc > 6) ? argv[6]: 0;
if(0==strcmp(argv[2], "test")) test_detector(datacfg, cfg, weights, filename, thresh, dont_show);
- else if(0==strcmp(argv[2], "train")) train_detector(datacfg, cfg, weights, gpus, ngpus, clear);
+ else if(0==strcmp(argv[2], "train")) train_detector(datacfg, cfg, weights, gpus, ngpus, clear, dont_show);
else if(0==strcmp(argv[2], "valid")) validate_detector(datacfg, cfg, weights);
else if(0==strcmp(argv[2], "recall")) validate_detector_recall(datacfg, cfg, weights);
else if(0==strcmp(argv[2], "map")) validate_detector_map(datacfg, cfg, weights, thresh);
+ else if(0==strcmp(argv[2], "calc_anchors")) calc_anchors(datacfg, num_of_clusters, final_width, final_heigh, show);
else if(0==strcmp(argv[2], "demo")) {
list *options = read_data_cfg(datacfg);
int classes = option_find_int(options, "classes", 20);
--
Gitblit v1.10.0