yolov3_coco.h5 + YOLOv3 model

Author: sonalrpatel

configs.py

@@ -34,7 +34,7 @@
 DIR_TEST = DIR_DATA + "test/"
 PATH_CLASSES = DIR_TRAIN + "_classes.txt"
 PATH_ANCHORS = "data/yolo_anchors.txt"
-PATH_WEIGHT = "data/yolov3_weights.h5"
+PATH_WEIGHT = "data/yolov3_coco.h5"
 PATH_DARKNET_WEIGHT = "data/yolov3.weights"
 
 # TRAIN options

convert2.py

@@ -33,7 +33,7 @@
 #   https://machinelearningmastery.com/how-to-perform-object-detection-with-yolov3-in-keras/
 #   https://github.com/experiencor/keras-yolo3/blob/master/yolo3_one_file_to_detect_them_all.py
 #   Input: yolov3 keras model, yolov3.weights
-#   Output: yolov3_weights.h5
+#   Output: yolov3_coco.h5
 # ===================================================================
 class WeightReader:
     def __init__(self, weight_file):
@@ -131,5 +131,5 @@ def _main(args):
 
 if __name__ == '__main__':
     # run following command (as per current folder structure) on terminal
-    # python convert2.py data/yolov3.weights data/yolov3_weights.h5
+    # python convert2.py data/yolov3.weights data/yolov3_coco.h5
     _main(parser.parse_args())

loss/loss_functional.py

@@ -127,9 +127,13 @@ def loop_body(idx, ignore_mask):
 
         # =========================================================
         #   calculate xy loss using binary_crossentropy
-        #   K.sigmoid calculation additionally added in this repository compared to original source
         # =========================================================
-        xy_loss = object_mask * box_loss_scale * K.binary_crossentropy(raw_true_xy, K.sigmoid(raw_pred[..., 0:2]),
+        #   Adding K.sigmoid calculation additionally (which is not in original repo) shifted the final prediction
+        #   TODO: need to investigate
+        # =========================================================
+        # xy_loss = object_mask * box_loss_scale * K.binary_crossentropy(raw_true_xy, K.sigmoid(raw_pred[..., 0:2]),
+        #                                                                from_logits=True)
+        xy_loss = object_mask * box_loss_scale * K.binary_crossentropy(raw_true_xy, raw_pred[..., 0:2],
                                                                        from_logits=True)
 
         # =========================================================

predict.py

@@ -216,7 +216,7 @@ def draw_boxes(image, boxes, labels, obj_thresh):
 
 def _main(args):
     image_path = args.image_path if args.image_path is not None else "data/apple.jpg"
-    weight_path = "data/yolov3_weights.h5"
+    weight_path = "data/yolov3.h5"
 
     # set some parameters
     net_h, net_w = 416, 416
@@ -238,9 +238,9 @@ def _main(args):
     # anchors = [[116, 90, 156, 198, 373, 326], [30, 61, 62, 45, 59, 119], [10, 13, 16, 30, 33, 23]]
 
     # make the yolov3 model to predict 80 classes on COCO
-    yolov3 = YOLOv3((None, None, 3), num_classes)
+    # yolov3 = YOLOv3((None, None, 3), num_classes)
     # yolov3 = make_yolov3_model((None, None, 3))
-    # yolov3 = yolo_body((None, None, 3), anchors_mask, num_classes)
+    yolov3 = yolo_body((None, None, 3), anchors_mask, num_classes)
 
     # run model summary
     # yolov3.summary()

predict2.py

@@ -56,7 +56,7 @@ def __init__(self, args):
         #   If the shape does not match, pay attention to the modification of the model_path
         #       and classes_path parameters during training
         # =====================================================================
-        self.weight_path = args.weight_path if args.weight_path is not None else 'data/yolov3_weights.h5'
+        self.weight_path = args.weight_path if args.weight_path is not None else 'data/yolov3_coco.h5'
         self.classes_path = args.classes_path if args.classes_path is not None else 'data/coco_classes.txt'
 
         # =====================================================================
@@ -448,6 +448,7 @@ def _main(args):
 if __name__ == '__main__':
     # run following command (as per current folder structure) on terminal
     # python predict2.py [-i] <image_path>
-    # python predict2.py -w data/yolov3_weights_license.h5 -c data/license_classes.txt -i data/florida_license.jpg
+    # python predict2.py -w data/yolov3_license.h5 -c data/license_classes.txt -i data/florida_license.jpg
+    # python predict2.py -w data/yolov3_fruits.h5 -c data/fruits/train/_classes.txt -i data/apple.jpg
     _main(parser.parse_args())