add np_mAP

experiment/cluster_bbox.py

@@ -34,6 +34,10 @@ def kmeans_clustering(xml_dir, k):
     avg_iou = kmeans_bbox.mean_iou()
     print(centroid)
     print(avg_iou)
+    def bbox_area(bbox): 
+        return bbox[2] * bbox[3]
+    sorted_centroid = sorted(centroid, key=bbox_area)
+    print(list(np.reshape([[b[2], b[3]] for b in sorted_centroid], -1)))
 
 def get_args():
     parser = argparse.ArgumentParser()

experiment/configs/parsecfg.py

@@ -8,6 +8,12 @@
 import numpy as np
 
 
+def Config(object):
+    def __init__(self):
+        pass
+
+
+
 def parse_cfg(file_name):
     # section_counters = defaultdict(int)
     cfg_stream = open(file_name)
@@ -28,6 +34,7 @@ def parse_cfg(file_name):
             cfg_dict['obj_score_thr'] = float(cfg_parser[section]['obj_score_thresh'])
             cfg_dict['nms_iou_thr'] = float(cfg_parser[section]['nms_iou_thresh'])
             cfg_dict['n_class'] = int(cfg_parser[section]['classes'])
+            cfg_dict['ignore_thr'] = float(cfg_parser[section]['ignore_thr'])
 
             anchors = list(map(float, (cfg_parser[section]['anchors']).split(',')))
             anchor_mask = list(map(int, (cfg_parser[section]['anchor_mask']).split(',')))

experiment/configs/voc.cfg

@@ -8,7 +8,8 @@ multiscale = 320, 352, 384, 416, 448, 480, 512, 544, 576, 608
 
 [yolo]
 anchor_mask = 2, 2, 2, 1, 1, 1, 0, 0, 0 
-anchors = 10,13,  16,30,  33,23,  30,61,  62,45,  59,119,  116,90,  156,198,  373,326
+anchors = 20.0, 26.0, 31.0, 67.0, 59.0, 42.0, 57.0, 115.0, 112.0, 91.0, 107.0, 191.0, 244.0, 142.0, 192.0, 280.0, 373.0, 298.0
 classes=20
 obj_score_thresh=0.8
-nms_iou_thresh=0.45
+nms_iou_thresh=0.45
+ignore_thr=0.5

experiment/convert/convert_model.py

@@ -2,7 +2,7 @@
 # -*- coding: utf-8 -*-
 # File: convert_model.py
 # Author: Qian Ge <[email protected]>
-# Modified from 
+# modified from:
 # https://github.com/qqwweee/keras-yolo3/blob/master/convert.py
 # reference:
 # https://github.com/pjreddie/darknet/blob/b13f67bfdd87434e141af532cdb5dc1b8369aa3b/src/parser.c#L958
@@ -17,20 +17,18 @@
 
 
 def unique_config_sections(config_path):
-    """Convert all config sections to have unique names.
-    Adds unique suffixes to config sections for compability with configparser.
+    """ Convert all config sections to have unique names.
+        Adds unique suffixes to config sections for compability with configparser.
     """
     section_counters = defaultdict(int)
     output_stream = io.StringIO()
     yolo_id = 0
-    # prev_dim = 3
-    # prev_dim_dict = {}
+
     with open(config_path) as fin:
         for line in fin:
             if line.startswith('['):
                 section = line.strip().strip('[]')
                 n_section = section
-                # out_dim = prev_dim
                 if section == 'yolo':
                     n_section = section
                     yolo_id += 1
@@ -50,15 +48,25 @@ def unique_config_sections(config_path):
                     _section = n_section + '_' + str(section_counters[n_section]-1)
                 else:
                     _section = n_section + '_' + str(section_counters[n_section])
-                # prev_dim_dict[_section] = prev_dim
-                # prev_dim = out_dim
+
                 print(_section)
                 line = line.replace(section, _section)
             output_stream.write(line)
     output_stream.seek(0)
     return output_stream
 
 def parse_conv(weights_file, cfg_parser, section, layer_dict):
+    """ parse conv layer
+
+        Args:
+            weights_file (file object): file object of .weights file
+            cfg_parser (ConfigParser object): ConfigParser object of .cfg file for net
+            section (str): name of conv layer
+            layer_dict (dictionary): dict storing layer info
+
+        Returns:
+            dict storing layer info and weights values
+    """
     prev_layer_channel = layer_dict['prev_layer_channel']
     count = layer_dict['count']
 
@@ -69,11 +77,6 @@ def parse_conv(weights_file, cfg_parser, section, layer_dict):
     activation = cfg_parser[section]['activation']
     batch_normalize = 'batch_normalize' in cfg_parser[section]
 
-    # Setting weights.
-    # Darknet serializes convolutional weights as:
-    # [bias/beta, [gamma, mean, variance], conv_weights]
-    # prev_layer_shape = K.int_shape(prev_layer)
-
     weights_shape = (size, size, prev_layer_channel, filters)
     darknet_w_shape = (filters, weights_shape[2], size, size)
     weights_size = np.product(weights_shape)
@@ -125,6 +128,16 @@ def parse_conv(weights_file, cfg_parser, section, layer_dict):
     return layer_dict
 
 def convert(weights_path, config_path, save_path):
+    """ convert .weight file to .npy file
+        The converted .npy file will be saved in save_path
+
+        Args:
+            weights_path (str): path of .weight file
+            config_path (str): path of configuration .cfg file
+            save_path (str): path for saving .npy file
+
+    """
+    # load weights file
     weights_file = open(weights_path, 'rb')
     major, minor, revision = np.ndarray(
         shape=(3, ), dtype='int32', buffer=weights_file.read(12))
@@ -134,7 +147,7 @@ def convert(weights_path, config_path, save_path):
     else:
         seen = np.ndarray(shape=(1,), dtype='int32', buffer=weights_file.read(4))
     print('Weights Header: ', major, minor, revision, seen)
-
+    # parse net configuration
     net_config = unique_config_sections(config_path)
     cfg_parser = configparser.ConfigParser()
     cfg_parser.read_file(net_config)
@@ -144,29 +157,23 @@ def convert(weights_path, config_path, save_path):
     dim_list = []
     layer_dict['prev_layer_channel'] = 3
     layer_dict['count'] = 0
-    # layer_id = 0
     for section in cfg_parser.sections():
         print('Parsing section {}'.format(section))
         if section.startswith('conv'):
             save_weight_dict[section] = {}
-
             layer_dict = parse_conv(weights_file, cfg_parser, section, layer_dict)
             save_weight_dict[section]['weights'] = layer_dict['conv_weights']
-
             if len(layer_dict['bn_weight_list']) > 0:
                save_weight_dict[section]['bn'] = layer_dict['bn_weight_list']
             if len(layer_dict['conv_bias']) > 0:
                 save_weight_dict[section]['biases'] = layer_dict['conv_bias']
         elif section.startswith('route'):
             route_layers = list(map(int, (cfg_parser[section]['layers']).split(',')))
             layer_dict['prev_layer_channel'] = sum([dim_list[layer_] for layer_ in route_layers])
-            # print(route_layers)
         dim_list.append(layer_dict['prev_layer_channel'])
-        # layer_id += 1
     remaining_weights = len(weights_file.read()) / 4
     print('Load {} of {} from weights.'.format(layer_dict['count'], remaining_weights + layer_dict['count']))
     weights_file.close()
-    # print(dim_list)
     np.save(save_path, save_weight_dict)
 
 def get_args():
@@ -192,20 +199,22 @@ def get_args():
         weights_dir = FLAGS.weights_dir
         save_dir = FLAGS.save_dir
 
-
     FLAGS = get_args()
     if FLAGS.model == 'darknet':
+        # convert Darknet53 for classification
         config_path = 'darknet53.cfg'
         weights_path = os.path.join(weights_dir, 'darknet53_448.weights')
         save_path = os.path.join(save_dir, 'darknet53_448.npy')
     elif FLAGS.model == 'yolov3_feat':
+        # convert Darknet53 first 52 conv layers for feature extration in yolov3
         config_path = 'yolov3_feat.cfg'
         weights_path = os.path.join(weights_dir, 'yolov3.weights')
         save_path = os.path.join(save_dir, 'yolov3_feat.npy')
     elif FLAGS.model == 'yolo':
+        # convert yolov3 trained on COCO dataset
         config_path = 'yolov3.cfg'
         weights_path = os.path.join(weights_dir, 'yolov3.weights')
         save_path = os.path.join(save_dir, 'yolov3.npy')
+
     convert(weights_path, config_path, save_path)
-    # unique_config_sections(config_path)
 

experiment/darknet.py

@@ -34,14 +34,18 @@ def get_args():
 
 def predict():
     FLAGS = get_args()
+    # load class id and the corresponding class name
     label_dict = loader.load_imagenet1k_label_darknet()
-    # Create a Dataflow object for test images
+    # create a Dataflow object for test images
     image_data = loader.read_image(
         im_name=FLAGS.im_name, n_channel=3,
-        data_dir=FLAGS.data_dir, batch_size=1, rescale=FLAGS.rescale)
+        data_dir=FLAGS.data_dir, batch_size=1,
+        rescale=FLAGS.rescale)
 
+    # create test model
     test_model = DarkNet53(
-        n_channel=3, n_class=1000, pre_trained_path=FLAGS.pretrained_path, trainable=False)
+        n_channel=3,n_class=1000,
+        pre_trained_path=FLAGS.pretrained_path, trainable=False)
     test_model.create_valid_model()
 
     with tf.Session() as sess:
@@ -63,4 +67,4 @@ def predict():
 
 
 if __name__ == '__main__':
-    predict()
+    predict()

experiment/test.py

@@ -15,8 +15,10 @@
 import src.bbox.bboxgt as bboxgt
 import src.utils.viz as viz
 import src.utils.image as image
+import src.evaluate.np_eval as np_eval
 
-if __name__ == "__main__":
+
+def test_target_anchor():
     pathconfig = parscfg.parse_cfg('configs/{}_path.cfg'.format(platform.node()))
     pretrained_path = pathconfig['coco_pretrained_npy_path']
     data_dir = pathconfig['test_image_path']
@@ -86,4 +88,25 @@
     viz.draw_bounding_box(o_im, gt_bbox_para, label_list=None, box_type='xyxy')
     viz.draw_bounding_box(rescale_im, target_anchor_batch[0], label_list=None, box_type='xyxy')
 
+def test_mAP():
+    pred_bboxes = [[25,35,45,55], [35,45,55,65],[250,350,450,550],[250,350,450,550],[250,350,450,550],
+                    [45,65,55,75],[15,25,35,45],[250,350,450,550],[250,350,450,550],[35,25,55,45],[15,25,35,45],
+                    ]
+    gt_bboxes = [[25,35,45,55], [35,45,55,65],[45,65,55,75],[15,25,35,45],[35,25,55,45],[45,65,55,75],[15,25,35,45],[35,25,55,45]]
+    pred_classes = [1,1,1,1,1,1,1,1,1,1, 2]
+    gt_classes = [1,1,1,1,1,2,2,2]
+    pred_conf = [0.9, 0.9, 0.8,0.7,0.6,0.5,0.4,0.3,0.2,0.1, 0.9]
+    IoU_thr = 0.5
+    pred_im_size = 1
+    gt_im_size = 1
+
+    re = np_eval.mAP(
+        pred_bboxes, pred_classes, pred_conf, gt_bboxes,
+        gt_classes, IoU_thr, pred_im_size, gt_im_size)
+
+    print(re)
+
+
+if __name__ == "__main__":
+    test_mAP()