adversarial attack - minor improvement

src/detector.c

@@ -1709,6 +1709,8 @@ void draw_object(char *datacfg, char *cfgfile, char *weightfile, char *filename,
         if (letter_box) sized = letterbox_image(im, net.w, net.h);
         else sized = resize_image(im, net.w, net.h);
 
+        image src_sized = copy_image(sized);
+
         layer l = net.layers[net.n - 1];
         net.num_boxes = l.max_boxes;
         int num_truth = l.truths;
@@ -1716,8 +1718,9 @@ void draw_object(char *datacfg, char *cfgfile, char *weightfile, char *filename,
 
         int *it_num_set = (int *)xcalloc(1, sizeof(int));
         float *lr_set = (float *)xcalloc(1, sizeof(float));
+        int *boxonly = (int *)xcalloc(1, sizeof(int));
 
-        cv_draw_object(sized, truth_cpu, net.num_boxes, num_truth, it_num_set, lr_set, l.classes, names);
+        cv_draw_object(sized, truth_cpu, net.num_boxes, num_truth, it_num_set, lr_set, boxonly, l.classes, names);
 
         net.learning_rate = *lr_set;
         it_num = *it_num_set;
@@ -1736,17 +1739,6 @@ void draw_object(char *datacfg, char *cfgfile, char *weightfile, char *filename,
         int iteration;
         for (iteration = 0; iteration < it_num; ++iteration)
         {
-            /*
-            free_network(net);
-            net = parse_network_cfg(cfgfile);// parse_network_cfg_custom(cfgfile, 1, 1); // set batch=1
-            net.learning_rate = *lr_set;
-            net.adversarial = 1;
-            set_batch_network(&net, 1);
-            if (weightfile) {
-                load_weights(&net, weightfile);
-            }
-            */
-
             if (iteration == it_num - 1) {
                 net.train = 0;
                 quantize_image(sized);
@@ -1758,55 +1750,38 @@ void draw_object(char *datacfg, char *cfgfile, char *weightfile, char *filename,
             draw_train_loss(windows_name, img, img_size, avg_loss, max_img_loss, iteration, it_num, 0, 0, "mAP%", dont_show, 0, 0);
             //quantize_image(sized);
 
-            show_image_cv(sized, "image_optimization");
+            if (*boxonly) {
-            wait_key_cv(20);
+                int dw = truth_cpu[2] * sized.w, dh = truth_cpu[3] * sized.h;
+                int dx = truth_cpu[0] * sized.w - dw / 2, dy = truth_cpu[1] * sized.h - dh / 2;
+                image crop = crop_image(sized, dx, dy, dw, dh);
+                copy_image_inplace(src_sized, sized);
+                embed_image(crop, sized, dx, dy);
             }
 
-        {
+            show_image_cv(sized, "image_optimization");
-        int nboxes = 0;
+            wait_key_cv(20);
-        detection *dets = get_network_boxes(&net, im.w, im.h, thresh, 0, 0, 1, &nboxes, letter_box);
-        if (nms) {
-            if (l.nms_kind == DEFAULT_NMS) do_nms_sort(dets, nboxes, l.classes, nms);
-            else diounms_sort(dets, nboxes, l.classes, nms, l.nms_kind, l.beta_nms);
-        }
-        draw_detections_v3(im, dets, nboxes, thresh, names, alphabet, l.classes, 1);
-        save_image(im, "pre_predictions");
-        if (!dont_show) {
-            show_image(im, "pre_predictions");
-        }
         }
 
         quantize_image(sized);
         save_image_png(sized, "drawn");
         //sized = load_image("drawn.png", 0, 0, net.c);
 
-        free_network(net);
-        net = parse_network_cfg_custom(cfgfile, 1, 1); // set batch=1
-        if (weightfile) {
-            load_weights(&net, weightfile);
-        }
-        net.benchmark_layers = benchmark_layers;
-
-
-        double time = get_time_point();
-        network_predict(net, X);
-        printf("%s: Predicted in %lf milli-seconds.\n", input, ((double)get_time_point() - time) / 1000);
-
         int nboxes = 0;
-        detection *dets = get_network_boxes(&net, im.w, im.h, thresh, 0, 0, 1, &nboxes, letter_box);
+        detection *dets = get_network_boxes(&net, sized.w, sized.h, thresh, 0, 0, 1, &nboxes, letter_box);
         if (nms) {
             if (l.nms_kind == DEFAULT_NMS) do_nms_sort(dets, nboxes, l.classes, nms);
             else diounms_sort(dets, nboxes, l.classes, nms, l.nms_kind, l.beta_nms);
         }
-        draw_detections_v3(im, dets, nboxes, thresh, names, alphabet, l.classes, 1);
+        draw_detections_v3(sized, dets, nboxes, thresh, names, alphabet, l.classes, 1);
-        save_image(im, "predictions");
+        save_image(sized, "pre_predictions");
         if (!dont_show) {
-            show_image(im, "predictions");
+            show_image(sized, "pre_predictions");
         }
 
         free_detections(dets, nboxes);
         free_image(im);
         free_image(sized);
+        free_image(src_sized);
 
         if (!dont_show) {
             wait_until_press_key_cv();

src/image.c

@@ -654,6 +654,11 @@ void normalize_image2(image p)
     free(max);
 }
 
+void copy_image_inplace(image src, image dst)
+{
+    memcpy(dst.data, src.data, src.h*src.w*src.c * sizeof(float));
+}
+
 image copy_image(image p)
 {
     image copy = p;

src/image.h

@@ -82,6 +82,7 @@ image make_empty_image(int w, int h, int c);
 image float_to_image_scaled(int w, int h, int c, float *data);
 image float_to_image(int w, int h, int c, float *data);
 image copy_image(image p);
+void copy_image_inplace(image src, image dst);
 image load_image(char *filename, int w, int h, int c);
 image load_image_stb_resize(char *filename, int w, int h, int c);
 //LIB_API image load_image_color(char *filename, int w, int h);

src/image_opencv.cpp

@@ -1347,7 +1347,7 @@ void callback_mouse_click(int event, int x, int y, int flags, void* user_data)
     }
 }
 
-extern "C" void cv_draw_object(image sized, float *truth_cpu, int max_boxes, int num_truth, int *it_num_set, float *lr_set, int classes, char **names)
+extern "C" void cv_draw_object(image sized, float *truth_cpu, int max_boxes, int num_truth, int *it_num_set, float *lr_set, int *boxonly, int classes, char **names)
 {
     cv::Mat frame = image_to_mat(sized);
     if(frame.channels() == 3) cv::cvtColor(frame, frame, cv::COLOR_RGB2BGR);
@@ -1374,6 +1374,9 @@ extern "C" void cv_draw_object(image sized, float *truth_cpu, int max_boxes, int
     std::string const cl_trackbar_name = "class_id";
     int cl_tb_res = cv::createTrackbar(cl_trackbar_name, window_name, &cl_trackbar_value, classes-1);
 
+    std::string const bo_trackbar_name = "box-only";
+    int bo_tb_res = cv::createTrackbar(bo_trackbar_name, window_name, boxonly, 1);
+
     int i = 0;
 
     while (!selected) {

src/image_opencv.h

@@ -109,7 +109,7 @@ void blend_images_cv(image new_img, float alpha, image old_img, float beta);
 image blur_image(image src_img, int ksize);
 
 // draw objects for Adversarial attacks
-void cv_draw_object(image sized, float *truth_cpu, int max_boxes, int num_truth, int *it_num_set, float *lr_set, int classes, char **names);
+void cv_draw_object(image sized, float *truth_cpu, int max_boxes, int num_truth, int *it_num_set, float *lr_set, int *boxonly, int classes, char **names);
 
 // Show Anchors
 void show_acnhors(int number_of_boxes, int num_of_clusters, float *rel_width_height_array, model anchors_data, int width, int height);

src/network_kernels.cu

@@ -232,7 +232,6 @@ void backward_network_gpu(network net, network_state state)
         printf(" x_size = %d, original_delta = %p, original_input = %p, net.learning_rate = %d \n",
             x_size, original_delta, original_input, x_size, net.learning_rate);
         axpy_ongpu(x_size, net.learning_rate, original_delta, 1, original_input, 1);
-        //axpy_ongpu(x_size, 0.1, original_delta, 1, original_input, 1);
         constrain_min_max_ongpu(x_size, 0, 1, original_input, 1);
     }
 

src/yolo_layer.c

@@ -171,8 +171,8 @@ ious delta_yolo_box(box truth, float *x, float *biases, int n, int index, int i,
         float tw = log(truth.w*w / biases[2 * n]);
         float th = log(truth.h*h / biases[2 * n + 1]);
 
-        printf(" tx = %f, ty = %f, tw = %f, th = %f \n", tx, ty, tw, th);
+        //printf(" tx = %f, ty = %f, tw = %f, th = %f \n", tx, ty, tw, th);
-        printf(" x = %f, y = %f, w = %f, h = %f \n", x[index + 0 * stride], x[index + 1 * stride], x[index + 2 * stride], x[index + 3 * stride]);
+        //printf(" x = %f, y = %f, w = %f, h = %f \n", x[index + 0 * stride], x[index + 1 * stride], x[index + 2 * stride], x[index + 3 * stride]);
 
         // accumulate delta
         delta[index + 0 * stride] += scale * (tx - x[index + 0 * stride]) * iou_normalizer;
@@ -477,9 +477,8 @@ void forward_yolo_layer(const layer l, network_state state)
                 int class_index = entry_index(l, b, mask_n*l.w*l.h + j*l.w + i, 4 + 1);
                 delta_yolo_class(l.output, l.delta, class_index, class_id, l.classes, l.w*l.h, &avg_cat, l.focal_loss, l.label_smooth_eps, l.classes_multipliers);
 
-                printf(" label: class_id = %d, truth.x = %f, truth.y = %f, truth.w = %f, truth.h = %f \n", class_id, truth.x, truth.y, truth.w, truth.h);
+                //printf(" label: class_id = %d, truth.x = %f, truth.y = %f, truth.w = %f, truth.h = %f \n", class_id, truth.x, truth.y, truth.w, truth.h);
-                printf(" mask_n = %d, l.output[obj_index] = %f, l.output[class_index + class_id] = %f \n\n", mask_n, l.output[obj_index], l.output[class_index + class_id]);
+                //printf(" mask_n = %d, l.output[obj_index] = %f, l.output[class_index + class_id] = %f \n\n", mask_n, l.output[obj_index], l.output[class_index + class_id]);
-
 
                 ++count;
                 ++class_count;