Added blur=1 for Classifier training if OPENCV=1 (removes textures like in stylized-ImageNet)

build/darknet/x64/cfg/efficientnet_b0.cfg

@@ -12,6 +12,7 @@ momentum=0.9
 decay=0.0005
 max_crop=256
 #mixup=4
+blur=1
 cutmix=1
 mosaic=1
 

cfg/efficientnet_b0.cfg

@@ -12,6 +12,7 @@ momentum=0.9
 decay=0.0005
 max_crop=256
 #mixup=4
+blur=1
 cutmix=1
 mosaic=1
 

src/classifier.c

@@ -84,6 +84,7 @@ void train_classifier(char *datacfg, char *cfgfile, char *weightfile, int *gpus,
     args.min = net.min_crop;
     args.max = net.max_crop;
     args.flip = net.flip;
+    args.blur = net.blur;
     args.angle = net.angle;
     args.aspect = net.aspect;
     args.exposure = net.exposure;

src/data.c

@@ -1289,7 +1289,7 @@ void *load_thread(void *ptr)
     if (a.type == OLD_CLASSIFICATION_DATA){
         *a.d = load_data_old(a.paths, a.n, a.m, a.labels, a.classes, a.w, a.h);
     } else if (a.type == CLASSIFICATION_DATA){
-        *a.d = load_data_augment(a.paths, a.n, a.m, a.labels, a.classes, a.hierarchy, a.flip, a.min, a.max, a.w, a.h, a.angle, a.aspect, a.hue, a.saturation, a.exposure, a.mixup, a.show_imgs);
+        *a.d = load_data_augment(a.paths, a.n, a.m, a.labels, a.classes, a.hierarchy, a.flip, a.min, a.max, a.w, a.h, a.angle, a.aspect, a.hue, a.saturation, a.exposure, a.mixup, a.blur, a.show_imgs);
     } else if (a.type == SUPER_DATA){
         *a.d = load_data_super(a.paths, a.n, a.m, a.w, a.h, a.scale);
     } else if (a.type == WRITING_DATA){
@@ -1434,7 +1434,7 @@ data load_data_super(char **paths, int n, int m, int w, int h, int scale)
     return d;
 }
 
-data load_data_augment(char **paths, int n, int m, char **labels, int k, tree *hierarchy, int use_flip, int min, int max, int w, int h, float angle, float aspect, float hue, float saturation, float exposure, int mixup, int show_imgs)
+data load_data_augment(char **paths, int n, int m, char **labels, int k, tree *hierarchy, int use_flip, int min, int max, int w, int h, float angle, float aspect, float hue, float saturation, float exposure, int mixup, int use_blur, int show_imgs)
 {
     char **paths_stored = paths;
     if(m) paths = get_random_paths(paths, n, m);
@@ -1557,6 +1557,28 @@ data load_data_augment(char **paths, int n, int m, char **labels, int k, tree *h
         }
     }
 
+#ifdef OPENCV
+    if (use_blur) {
+        int i;
+        for (i = 0; i < d.X.rows; ++i) {
+            if (random_gen() % 2) {
+                image im = make_empty_image(w, h, 3);
+                im.data = d.X.vals[i];
+                int ksize = use_blur;
+                if (use_blur == 1) ksize = 17;
+                image blurred = blur_image(im, ksize);
+                free_image(im);
+                d.X.vals[i] = blurred.data;
+                //if (i == 0) {
+                //    show_image(im, "Not blurred");
+                //    show_image(blurred, "blurred");
+                //    wait_until_press_key_cv();
+                //}
+            }
+        }
+    }
+#endif  // OPENCV
+
     if (show_imgs) {
         int i, j;
         for (i = 0; i < d.X.rows; ++i) {

src/data.h

@@ -91,7 +91,7 @@ data load_data_detection(int n, char **paths, int m, int w, int h, int c, int bo
 data load_data_tag(char **paths, int n, int m, int k, int use_flip, int min, int max, int w, int h, float angle, float aspect, float hue, float saturation, float exposure);
 matrix load_image_augment_paths(char **paths, int n, int use_flip, int min, int max, int w, int h, float angle, float aspect, float hue, float saturation, float exposure);
 data load_data_super(char **paths, int n, int m, int w, int h, int scale);
-data load_data_augment(char **paths, int n, int m, char **labels, int k, tree *hierarchy, int use_flip, int min, int max, int w, int h, float angle, float aspect, float hue, float saturation, float exposure, int mixup, int show_imgs);
+data load_data_augment(char **paths, int n, int m, char **labels, int k, tree *hierarchy, int use_flip, int min, int max, int w, int h, float angle, float aspect, float hue, float saturation, float exposure, int mixup, int use_blur, int show_imgs);
 data load_go(char *filename);
 
 box_label *read_boxes(char *filename, int *n);

src/image_opencv.cpp

@@ -1258,6 +1258,16 @@ void blend_images_cv(image new_img, float alpha, image old_img, float beta)
     cv::addWeighted(new_mat, alpha, old_mat, beta, 0.0, new_mat);
 }
 
+// bilateralFilter bluring
+image blur_image(image src_img, int ksize)
+{
+    cv::Mat src = image_to_mat(src_img);
+    cv::Mat dst;
+    cv::bilateralFilter(src, dst, ksize, 75, 75);
+    image dst_img = mat_to_image(dst);
+    return dst_img;
+}
+
 // ====================================================================
 // Show Anchors
 // ====================================================================

src/image_opencv.h

@@ -101,6 +101,9 @@ image image_data_augmentation(mat_cv* mat, int w, int h,
 // blend two images with (alpha and beta)
 void blend_images_cv(image new_img, float alpha, image old_img, float beta);
 
+// bilateralFilter bluring
+image blur_image(image src_img, int ksize);
+
 // 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

@@ -114,8 +114,8 @@ void backward_network_gpu(network net, network_state state)
 
         if (i != 0) {
             layer prev = net.layers[i - 1];
-            if (net.optimized_memory && state.delta && !prev.keep_delta_gpu && prev.delta_gpu != l.delta_gpu) {
-                simple_copy_ongpu(prev.outputs*prev.batch, state.delta, prev.delta_gpu);
+            if (net.optimized_memory && state.delta && !prev.keep_delta_gpu) {
+                if (prev.delta_gpu != state.delta) simple_copy_ongpu(prev.outputs*prev.batch, state.delta, prev.delta_gpu);
                 fill_ongpu(prev.outputs*prev.batch, 0, net.state_delta_gpu, 1);
             }
         }