fix rescale bug

README.md

@@ -76,11 +76,11 @@
 - Result images are saved in `save_path` setting in `configs/pretrain_coco_path.cfg`.
 
 ### Sample results
-<img src='docs/figs/pretrained/im0.png' height='350px'>
+<img src='docs/figs/pretrained/im0.png' height='300px'>
 <!--<img src='docs/figs/pretrained/im1.png' height='350px'>-->
-<img src='docs/figs/pretrained/im2.png' height='350px'>
-<img src='docs/figs/pretrained/im3.png' height='400px'>
-<img src='docs/figs/pretrained/im4.png' height='400px'>
+<img src='docs/figs/pretrained/im2.png' height='300px'>
+<img src='docs/figs/pretrained/im3.png' height='300px'>
+<img src='docs/figs/pretrained/im4.png' height='300px'>
 
 ## Train on [VOC2012](http://host.robots.ox.ac.uk/pascal/VOC/voc2012/) dataset (20 classes)
 ### Prepare dataset and pre-trained feature extractor
@@ -125,7 +125,7 @@
 - Validation image are all rescaled to 416 * 416 without augmentation for validation.
 - The learning rate schedule needs to be further tuned, but the current setting is: 0.1 (1-50 epochs), 0.01 (51-100 epochs) and 0.001 (101-150 epochs).
 - Tensorboard summary includes losses and sample predictions for both training set (every 100 steps) and validation set (every epoch) are saved in `save_path` in `configs/config_path.cfg`. Note that non-maximum suppression does not used in sample predictions and only top 20 predicted bounding boxes based on class score are shown. You can see how the model is doing during training: 
-  <img src='docs/figs/train/tensorboard.png' height='350px'>
+  <img src='docs/figs/train/tensorboard.png' height='400px'>
 
 ### Sample results
 - Prediction after 150 epochs. Performance evaluation will be added soon.

experiment/configs/config_path.cfg

@@ -1,6 +1,6 @@
 [path]
 coco_pretrained_npy = /home/qge2/workspace/data/pretrain/yolo/yolov3.npy
-save_path = /home/qge2/workspace/data/out/yolo/yolo_15/
+save_path = /home/qge2/workspace/data/out/yolo/yolo_16/
 test_image_path = ../data/yolo/
 test_image_name = .jpg
 yolo_feat_pretraind_npy = /home/qge2/workspace/data/pretrain/yolo/yolov3_feat.npy

experiment/darknet.py

@@ -38,14 +38,18 @@ def predict():
     label_dict = loader.load_imagenet1k_label_darknet()
     # 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)
+        im_name=FLAGS.im_name, 
+        n_channel=3,
+        batch_size=1,
+        rescale=FLAGS.rescale,
+        data_dir=FLAGS.data_dir, )
 
     # 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:

experiment/loader.py

@@ -75,11 +75,11 @@ def load_VOC(rescale_shape_list,
     #     im_dir = '/home/qge2/workspace/data/dataset/VOCdevkit/VOC2012/JPEGImages/'
     #     xml_dir = '/home/qge2/workspace/data/dataset/VOCdevkit/VOC2012/Annotations/'
     # elif platform.node() == 'aros04':
-    #     im_dir = 'E:/Dataset/VOCdevkit/VOC2007/JPEGImages/'
-    #     xml_dir = 'E:/Dataset/VOCdevkit/VOC2007/Annotations/'
+    #     im_dir = 'E:/Dataset/VOCdevkit/test/JPEGImages/'
+    #     xml_dir = 'E:/Dataset/VOCdevkit/test/Annotations/'
     # else:
-    #     im_dir = '/Users/gq/workspace/Dataset/VOCdevkit/VOC2007/JPEGImages/'
-    #     xml_dir = '/Users/gq/workspace/Dataset/VOCdevkit/VOC2007/Annotations/'
+    #     im_dir = '/Users/gq/workspace/Dataset/VOCdevkit/test/JPEGImages/'
+    #     xml_dir = '/Users/gq/workspace/Dataset/VOCdevkit/test/Annotations/'
 
     class_name_dict, class_id_dict = get_class_dict_from_xml(xml_dir)
 
@@ -94,6 +94,7 @@ def normalize_im(im, *args):
             im = im / 255.
         return np.clip(im, 0., 1.)
 
+    # Load entire dataset
     voc_data = VOC(
         class_dict=class_name_dict,
         image_dir=im_dir,
@@ -105,11 +106,13 @@ def normalize_im(im, *args):
         )
     voc_data.setup(epoch_val=0, batch_size=1)
 
+    # divide dataset into training and validation set
     train_data, valid_date = dfoper.divide_dataflow(voc_data, divide_list=[train_percentage, 1 - train_percentage], shuffle=True)
 
     if n_class is None:
         n_class = len(class_name_dict)
 
+    # create training set Generator
     train_data_generator = generator.Generator(
         dataflow=train_data, 
         n_channle=3,
@@ -124,6 +127,7 @@ def normalize_im(im, *args):
         max_num_bbox_per_im=max_num_bbox_per_im)
     train_data_generator.reset_im_scale(scale=416)
 
+    # create validation set Generator
     valid_data_generator = generator.Generator(
         dataflow=valid_date, 
         n_channle=3,

experiment/test.py

@@ -13,7 +13,6 @@
 sys.path.append('../')
 import loader
 import configs.parsecfg as parscfg
-import src.bbox.bboxgt as bboxgt
 
 import src.utils.image as image
 import src.dataflow.augmentation as augment
@@ -160,7 +159,6 @@ def test_input():
          num_parallel_preprocess=2,
          h_flip=True, crop=True, color=True, affine=True,
          max_num_bbox_per_im=45)
-
     print(valid_data_generator.batch_data)
 
     with tf.Session() as sess:
@@ -169,18 +167,19 @@ def test_input():
 
         for epoch in range(10):
             print('epoch: {}'.format(epoch))
-            train_data_generator.init_iterator(sess)
+            train_data_generator.init_iterator(sess, reset_scale=True)
             while True:
                 try:
+                    # train_data_generator.init_iterator(sess, reset_scale=True)
+                    # train_data_generator.reset_im_scale()
                     start_time = time.time()
                     t = sess.run(train_data_generator.batch_data)
                     print(time.time() - start_time)
                     print(t[0].shape)
                     print(t[-1].shape)
-                    viz.draw_bounding_box(t[0][0]*255, t[-1][0], label_list=None, box_type='xyxy')
+                    viz.draw_bounding_box(t[0][0]*255, t[2][0], label_list=None, box_type='xyxy')
                 except tf.errors.OutOfRangeError:
                     break
-            pre.set_output_scale(416)
 
 # def test_divide_df():
 #     import src.utils.viz as viz

experiment/yolov3.py

@@ -114,17 +114,22 @@ def train():
             #     category_index, 
             #     config.save_path, 
             #     run_type='epoch')
-
-            train_data_generator.init_iterator(sess)
+
+            if i > 0 and i % 10 == 0:
+                train_data_generator.init_iterator(sess, reset_scale=True)
+            else:
+                train_data_generator.init_iterator(sess, reset_scale=False)
+
             train_model.train_epoch(sess, lr, summary_writer=writer)
 
             valid_data_generator.init_iterator(sess)
             valid_model.valid_epoch(sess, summary_writer=writer)
 
             saver.save(sess, '{}/yolov3_epoch_{}'.format(config.save_path, i))
 
-            if i > 0 and i % 10 == 0:
-                train_data_generator.reset_im_scale()
+            # if i > 0 and i % 10 == 0:
+            #     train_data_generator.reset_im_scale()
+
     writer.close()
 
 def detect():