From 564877ad6a3f53d3d866b0015237d07f4af2eaa2 Mon Sep 17 00:00:00 2001
From: vinjn <vinjn.z@gmail.com>
Date: Sat, 07 Jul 2018 04:30:45 +0000
Subject: [PATCH] cuda.h - converts tab to space
---
src/yolo_v2_class.cpp | 232 ++++++++++++++++++++++++++++++++++++++++++++++++---------
1 files changed, 194 insertions(+), 38 deletions(-)
diff --git a/src/yolo_v2_class.cpp b/src/yolo_v2_class.cpp
index ea13ea3..4df9be5 100644
--- a/src/yolo_v2_class.cpp
+++ b/src/yolo_v2_class.cpp
@@ -22,38 +22,92 @@
#define FRAMES 3
-struct detector_gpu_t{
- float **probs;
- box *boxes;
+//static Detector* detector = NULL;
+static std::unique_ptr<Detector> detector;
+
+int init(const char *configurationFilename, const char *weightsFilename, int gpu)
+{
+ detector.reset(new Detector(configurationFilename, weightsFilename, gpu));
+ return 1;
+}
+
+int detect_image(const char *filename, bbox_t_container &container)
+{
+ std::vector<bbox_t> detection = detector->detect(filename);
+ for (size_t i = 0; i < detection.size() && i < C_SHARP_MAX_OBJECTS; ++i)
+ container.candidates[i] = detection[i];
+ return detection.size();
+}
+
+int detect_mat(const uint8_t* data, const size_t data_length, bbox_t_container &container) {
+#ifdef OPENCV
+ std::vector<char> vdata(data, data + data_length);
+ cv::Mat image = imdecode(cv::Mat(vdata), 1);
+
+ std::vector<bbox_t> detection = detector->detect(image);
+ for (size_t i = 0; i < detection.size() && i < C_SHARP_MAX_OBJECTS; ++i)
+ container.candidates[i] = detection[i];
+ return detection.size();
+#else
+ return -1;
+#endif // OPENCV
+}
+
+int dispose() {
+ //if (detector != NULL) delete detector;
+ //detector = NULL;
+ detector.reset();
+ return 1;
+}
+
+#ifdef GPU
+void check_cuda(cudaError_t status) {
+ if (status != cudaSuccess) {
+ const char *s = cudaGetErrorString(status);
+ printf("CUDA Error Prev: %s\n", s);
+ }
+}
+#endif
+
+struct detector_gpu_t {
network net;
image images[FRAMES];
float *avg;
float *predictions[FRAMES];
+ int demo_index;
+ unsigned int *track_id;
};
-
-YOLODLL_API Detector::Detector(std::string cfg_filename, std::string weight_filename, int gpu_id)
+YOLODLL_API Detector::Detector(std::string cfg_filename, std::string weight_filename, int gpu_id) : cur_gpu_id(gpu_id)
{
+ wait_stream = 0;
int old_gpu_index;
- cudaGetDevice(&old_gpu_index);
+#ifdef GPU
+ check_cuda( cudaGetDevice(&old_gpu_index) );
+#endif
detector_gpu_ptr = std::make_shared<detector_gpu_t>();
- detector_gpu_t &detector_gpu = *reinterpret_cast<detector_gpu_t *>(detector_gpu_ptr.get());
+ detector_gpu_t &detector_gpu = *static_cast<detector_gpu_t *>(detector_gpu_ptr.get());
- cudaSetDevice(gpu_id);
+#ifdef GPU
+ //check_cuda( cudaSetDevice(cur_gpu_id) );
+ cuda_set_device(cur_gpu_id);
+ printf(" Used GPU %d \n", cur_gpu_id);
+#endif
network &net = detector_gpu.net;
- net.gpu_index = gpu_id;
+ net.gpu_index = cur_gpu_id;
//gpu_index = i;
char *cfgfile = const_cast<char *>(cfg_filename.data());
char *weightfile = const_cast<char *>(weight_filename.data());
- net = parse_network_cfg(cfgfile);
+ net = parse_network_cfg_custom(cfgfile, 1);
if (weightfile) {
load_weights(&net, weightfile);
}
set_batch_network(&net, 1);
- net.gpu_index = gpu_id;
+ net.gpu_index = cur_gpu_id;
+ fuse_conv_batchnorm(net);
layer l = net.layers[net.n - 1];
int j;
@@ -62,42 +116,58 @@
for (j = 0; j < FRAMES; ++j) detector_gpu.predictions[j] = (float *)calloc(l.outputs, sizeof(float));
for (j = 0; j < FRAMES; ++j) detector_gpu.images[j] = make_image(1, 1, 3);
- detector_gpu.boxes = (box *)calloc(l.w*l.h*l.n, sizeof(box));
- detector_gpu.probs = (float **)calloc(l.w*l.h*l.n, sizeof(float *));
- for (j = 0; j < l.w*l.h*l.n; ++j) detector_gpu.probs[j] = (float *)calloc(l.classes, sizeof(float));
+ detector_gpu.track_id = (unsigned int *)calloc(l.classes, sizeof(unsigned int));
+ for (j = 0; j < l.classes; ++j) detector_gpu.track_id[j] = 1;
- cudaSetDevice(old_gpu_index);
+#ifdef GPU
+ check_cuda( cudaSetDevice(old_gpu_index) );
+#endif
}
YOLODLL_API Detector::~Detector()
{
- detector_gpu_t &detector_gpu = *reinterpret_cast<detector_gpu_t *>(detector_gpu_ptr.get());
+ detector_gpu_t &detector_gpu = *static_cast<detector_gpu_t *>(detector_gpu_ptr.get());
layer l = detector_gpu.net.layers[detector_gpu.net.n - 1];
+ free(detector_gpu.track_id);
+
free(detector_gpu.avg);
for (int j = 0; j < FRAMES; ++j) free(detector_gpu.predictions[j]);
for (int j = 0; j < FRAMES; ++j) if(detector_gpu.images[j].data) free(detector_gpu.images[j].data);
- for (int j = 0; j < l.w*l.h*l.n; ++j) free(detector_gpu.probs[j]);
- free(detector_gpu.boxes);
- free(detector_gpu.probs);
-
int old_gpu_index;
+#ifdef GPU
cudaGetDevice(&old_gpu_index);
- cudaSetDevice(detector_gpu.net.gpu_index);
+ cuda_set_device(detector_gpu.net.gpu_index);
+#endif
free_network(detector_gpu.net);
+#ifdef GPU
cudaSetDevice(old_gpu_index);
+#endif
+}
+
+YOLODLL_API int Detector::get_net_width() const {
+ detector_gpu_t &detector_gpu = *static_cast<detector_gpu_t *>(detector_gpu_ptr.get());
+ return detector_gpu.net.w;
+}
+YOLODLL_API int Detector::get_net_height() const {
+ detector_gpu_t &detector_gpu = *static_cast<detector_gpu_t *>(detector_gpu_ptr.get());
+ return detector_gpu.net.h;
+}
+YOLODLL_API int Detector::get_net_color_depth() const {
+ detector_gpu_t &detector_gpu = *static_cast<detector_gpu_t *>(detector_gpu_ptr.get());
+ return detector_gpu.net.c;
}
-YOLODLL_API std::vector<bbox_t> Detector::detect(std::string image_filename, float thresh)
+YOLODLL_API std::vector<bbox_t> Detector::detect(std::string image_filename, float thresh, bool use_mean)
{
std::shared_ptr<image_t> image_ptr(new image_t, [](image_t *img) { if (img->data) free(img->data); delete img; });
*image_ptr = load_image(image_filename);
- return detect(*image_ptr, thresh);
+ return detect(*image_ptr, thresh, use_mean);
}
static image load_image_stb(char *filename, int channels)
@@ -144,17 +214,21 @@
}
}
-YOLODLL_API std::vector<bbox_t> Detector::detect(image_t img, float thresh)
+YOLODLL_API std::vector<bbox_t> Detector::detect(image_t img, float thresh, bool use_mean)
{
-
- detector_gpu_t &detector_gpu = *reinterpret_cast<detector_gpu_t *>(detector_gpu_ptr.get());
+ detector_gpu_t &detector_gpu = *static_cast<detector_gpu_t *>(detector_gpu_ptr.get());
network &net = detector_gpu.net;
int old_gpu_index;
+#ifdef GPU
cudaGetDevice(&old_gpu_index);
- cudaSetDevice(net.gpu_index);
+ if(cur_gpu_id != old_gpu_index)
+ cudaSetDevice(net.gpu_index);
+
+ net.wait_stream = wait_stream; // 1 - wait CUDA-stream, 0 - not to wait
+#endif
//std::cout << "net.gpu_index = " << net.gpu_index << std::endl;
- float nms = .4;
+ //float nms = .4;
image im;
im.c = img.c;
@@ -162,23 +236,42 @@
im.h = img.h;
im.w = img.w;
- image sized = resize_image(im, net.w, net.h);
+ image sized;
+
+ if (net.w == im.w && net.h == im.h) {
+ sized = make_image(im.w, im.h, im.c);
+ memcpy(sized.data, im.data, im.w*im.h*im.c * sizeof(float));
+ }
+ else
+ sized = resize_image(im, net.w, net.h);
+
layer l = net.layers[net.n - 1];
float *X = sized.data;
- network_predict(net, X);
+ float *prediction = network_predict(net, X);
- get_region_boxes(l, 1, 1, thresh, detector_gpu.probs, detector_gpu.boxes, 0, 0);
- if (nms) do_nms_sort(detector_gpu.boxes, detector_gpu.probs, l.w*l.h*l.n, l.classes, nms);
- //draw_detections(im, l.w*l.h*l.n, thresh, boxes, probs, names, alphabet, l.classes);
+ if (use_mean) {
+ memcpy(detector_gpu.predictions[detector_gpu.demo_index], prediction, l.outputs * sizeof(float));
+ mean_arrays(detector_gpu.predictions, FRAMES, l.outputs, detector_gpu.avg);
+ l.output = detector_gpu.avg;
+ detector_gpu.demo_index = (detector_gpu.demo_index + 1) % FRAMES;
+ }
+ //get_region_boxes(l, 1, 1, thresh, detector_gpu.probs, detector_gpu.boxes, 0, 0);
+ //if (nms) do_nms_sort(detector_gpu.boxes, detector_gpu.probs, l.w*l.h*l.n, l.classes, nms);
+
+ int nboxes = 0;
+ int letterbox = 0;
+ float hier_thresh = 0.5;
+ detection *dets = get_network_boxes(&net, im.w, im.h, thresh, hier_thresh, 0, 1, &nboxes, letterbox);
+ if (nms) do_nms_sort(dets, nboxes, l.classes, nms);
std::vector<bbox_t> bbox_vec;
- for (size_t i = 0; i < (l.w*l.h*l.n); ++i) {
- box b = detector_gpu.boxes[i];
- int const obj_id = max_index(detector_gpu.probs[i], l.classes);
- float const prob = detector_gpu.probs[i][obj_id];
+ for (size_t i = 0; i < nboxes; ++i) {
+ box b = dets[i].bbox;
+ int const obj_id = max_index(dets[i].prob, l.classes);
+ float const prob = dets[i].prob[obj_id];
if (prob > thresh)
{
@@ -189,15 +282,78 @@
bbox.h = b.h*im.h;
bbox.obj_id = obj_id;
bbox.prob = prob;
+ bbox.track_id = 0;
bbox_vec.push_back(bbox);
}
}
+ free_detections(dets, nboxes);
if(sized.data)
free(sized.data);
- cudaSetDevice(old_gpu_index);
+#ifdef GPU
+ if (cur_gpu_id != old_gpu_index)
+ cudaSetDevice(old_gpu_index);
+#endif
return bbox_vec;
+}
+
+YOLODLL_API std::vector<bbox_t> Detector::tracking_id(std::vector<bbox_t> cur_bbox_vec, bool const change_history,
+ int const frames_story, int const max_dist)
+{
+ detector_gpu_t &det_gpu = *static_cast<detector_gpu_t *>(detector_gpu_ptr.get());
+
+ bool prev_track_id_present = false;
+ for (auto &i : prev_bbox_vec_deque)
+ if (i.size() > 0) prev_track_id_present = true;
+
+ if (!prev_track_id_present) {
+ for (size_t i = 0; i < cur_bbox_vec.size(); ++i)
+ cur_bbox_vec[i].track_id = det_gpu.track_id[cur_bbox_vec[i].obj_id]++;
+ prev_bbox_vec_deque.push_front(cur_bbox_vec);
+ if (prev_bbox_vec_deque.size() > frames_story) prev_bbox_vec_deque.pop_back();
+ return cur_bbox_vec;
+ }
+
+ std::vector<unsigned int> dist_vec(cur_bbox_vec.size(), std::numeric_limits<unsigned int>::max());
+
+ for (auto &prev_bbox_vec : prev_bbox_vec_deque) {
+ for (auto &i : prev_bbox_vec) {
+ int cur_index = -1;
+ for (size_t m = 0; m < cur_bbox_vec.size(); ++m) {
+ bbox_t const& k = cur_bbox_vec[m];
+ if (i.obj_id == k.obj_id) {
+ float center_x_diff = (float)(i.x + i.w/2) - (float)(k.x + k.w/2);
+ float center_y_diff = (float)(i.y + i.h/2) - (float)(k.y + k.h/2);
+ unsigned int cur_dist = sqrt(center_x_diff*center_x_diff + center_y_diff*center_y_diff);
+ if (cur_dist < max_dist && (k.track_id == 0 || dist_vec[m] > cur_dist)) {
+ dist_vec[m] = cur_dist;
+ cur_index = m;
+ }
+ }
+ }
+
+ bool track_id_absent = !std::any_of(cur_bbox_vec.begin(), cur_bbox_vec.end(),
+ [&i](bbox_t const& b) { return b.track_id == i.track_id && b.obj_id == i.obj_id; });
+
+ if (cur_index >= 0 && track_id_absent){
+ cur_bbox_vec[cur_index].track_id = i.track_id;
+ cur_bbox_vec[cur_index].w = (cur_bbox_vec[cur_index].w + i.w) / 2;
+ cur_bbox_vec[cur_index].h = (cur_bbox_vec[cur_index].h + i.h) / 2;
+ }
+ }
+ }
+
+ for (size_t i = 0; i < cur_bbox_vec.size(); ++i)
+ if (cur_bbox_vec[i].track_id == 0)
+ cur_bbox_vec[i].track_id = det_gpu.track_id[cur_bbox_vec[i].obj_id]++;
+
+ if (change_history) {
+ prev_bbox_vec_deque.push_front(cur_bbox_vec);
+ if (prev_bbox_vec_deque.size() > frames_story) prev_bbox_vec_deque.pop_back();
+ }
+
+ return cur_bbox_vec;
}
\ No newline at end of file
--
Gitblit v1.10.0