ipynb version for google colab & fixed more errors

@@ -24,10 +24,10 @@ anchor_path = data_path + 'yolo_anchors.txt'  # The path of the anchor txt file.
 class_name_path = data_path + 'classes.txt'  # The path of the class names.
 class_name_path = data_path + 'classes.txt'  # The path of the class names.
 
 
 ### Training releated numbers
 ### Training releated numbers
-batch_size = 6
+batch_size = 10
 img_size = [416, 416]  # Images will be resized to `img_size` and fed to the network, size format: [width, height]
 img_size = [416, 416]  # Images will be resized to `img_size` and fed to the network, size format: [width, height]
 letterbox_resize = True  # Whether to use the letterbox resize, i.e., keep the original aspect ratio in the resized image.
 letterbox_resize = True  # Whether to use the letterbox resize, i.e., keep the original aspect ratio in the resized image.
-total_epoches = 50
+total_epoches = 20
 train_evaluation_step = 10  # Evaluate on the training batch after some steps.
 train_evaluation_step = 10  # Evaluate on the training batch after some steps.
 val_evaluation_epoch = 2  # Evaluate on the whole validation dataset after some epochs. Set to None to evaluate every epoch.
 val_evaluation_epoch = 2  # Evaluate on the whole validation dataset after some epochs. Set to None to evaluate every epoch.
 save_epoch = 5  # Save the model after some epochs.
 save_epoch = 5  # Save the model after some epochs.
@@ -73,7 +73,7 @@ use_label_smooth = True # Whether to use class label smoothing strategy.
 use_focal_loss = True  # Whether to apply focal loss on the conf loss.
 use_focal_loss = True  # Whether to apply focal loss on the conf loss.
 use_mix_up = True  # Whether to use mix up data augmentation strategy. 
 use_mix_up = True  # Whether to use mix up data augmentation strategy. 
 use_warm_up = True  # whether to use warm up strategy to prevent from gradient exploding.
 use_warm_up = True  # whether to use warm up strategy to prevent from gradient exploding.
-warm_up_epoch = 3  # Warm up training epoches. Set to a larger value if gradient explodes.
+warm_up_epoch = 2  # Warm up training epoches. Set to a larger value if gradient explodes.
 
 
 ### some constants in validation
 ### some constants in validation
 # nms
 # nms

@@ -2,8 +2,13 @@ changes from https://github.com/wizyoung/YOLOv3_TensorFlow
 
 
 by Seongju Kim, kareus1@khu.ac.kr
 by Seongju Kim, kareus1@khu.ac.kr
 
 
+I only tested in colab environment yet (2020.05.16),
+so let me know if there are some errors/problems in python code version
+(##last changed: 2020.05.16)
+
 1] changed TextLineDataset to TFRecordDataset. (also changed data parsing in data utils and eval utils)
 1] changed TextLineDataset to TFRecordDataset. (also changed data parsing in data utils and eval utils)
 2] fixed restore-does-not-exist problem in train/eval mode
 2] fixed restore-does-not-exist problem in train/eval mode
 3] fixed saver to save the parameter only when save-optimizer option is true
 3] fixed saver to save the parameter only when save-optimizer option is true
 4] changed parameter 'mode' to bool value 'is_training' in data util functions (string value 'mode' is passed as byte string, so functions do not evaluate if-clauses as expected. ex) 'train' != b'train')
 4] changed parameter 'mode' to bool value 'is_training' in data util functions (string value 'mode' is passed as byte string, so functions do not evaluate if-clauses as expected. ex) 'train' != b'train')
-5] wrote TFRecord binary iterator, which runs without tf session (references: https://github.com/pgmmpk/tfrecord )
\ No newline at end of file
+5] wrote TFRecord binary iterator, which runs without tf session (references: https://github.com/pgmmpk/tfrecord )
+6] removed logging/tenorboard summary code. (I will add it later if necessary)
\ No newline at end of file

@@ -9,22 +9,20 @@ import random
 PY_VERSION = sys.version_info[0]
 PY_VERSION = sys.version_info[0]
 iter_cnt = 0
 iter_cnt = 0
 
 
-FEATURE_DESCRIPTION = {
-    'index': tf.FixedLenFeature([], tf.int64),
-    'image': tf.FixedLenFeature([], tf.string),
-    'width': tf.FixedLenFeature([], tf.int64),
-    'height': tf.FixedLenFeature([], tf.int64),
-    'boxes': tf.VarLenFeature(tf.int64)
-}
+def _parse_tfrecord(data):
+    example = tf.train.Example()
+    example.ParseFromString(data)
+    features = example.features.feature
+    return features
 
 
 def parse_tfrecord(data):
 def parse_tfrecord(data):
     # tfrecord parser for TFRecordDataset (raw data)
     # tfrecord parser for TFRecordDataset (raw data)
-    features = tf.parse_single_example(data, FEATURE_DESCRIPTION)
-    index = int(features['index'])
-    encoded_image = np.frombuffer(features['image'], dtype = np.uint8)
-    width = int(features['width'])
-    height = int(features['height'])
-    boxes = features['boxes'].eval()
+    features = _parse_tfrecord(data)
+    index = features['index'].int64_list.value[0]
+    encoded_image = np.frombuffer(features['image'].bytes_list.value[0], dtype = np.uint8)
+    width = features['width'].int64_list.value[0]
+    height = features['height'].int64_list.value[0]
+    boxes = features['boxes'].int64_list.value
 
 
     assert len(boxes) % 5 == 0, 'Annotation error occured in box array.'
     assert len(boxes) % 5 == 0, 'Annotation error occured in box array.'
     box_cnt = len(boxes) // 5
     box_cnt = len(boxes) // 5
@@ -33,7 +31,7 @@ def parse_tfrecord(data):
     labels = []
     labels = []
 
 
     for i in range(box_cnt):
     for i in range(box_cnt):
-        label, x_min, y_min, x_max, y_max = int(boxes[i * 5]), float(boxes[i * 5 + 1]), float(boxes[i * 5 + 2]), float(boxes[i * 5 + 3]) ## do we need to change int to float? is there float rectangle sample?
+        label, x_min, y_min, x_max, y_max = int(boxes[i * 5]), float(boxes[i * 5 + 1]), float(boxes[i * 5 + 2]), float(boxes[i * 5 + 3]), float(boxes[i * 5 + 4]) ## do we need to change int to float? is there float rectangle sample?
         aligned_boxes.append([x_min, y_min, x_max, y_max])
         aligned_boxes.append([x_min, y_min, x_max, y_max])
         labels.append(label)
         labels.append(label)
 
 

@@ -99,6 +99,8 @@ with tf.Session() as sess:
     sess.run([tf.global_variables_initializer()])
     sess.run([tf.global_variables_initializer()])
     if os.path.exists(args.restore_path):
     if os.path.exists(args.restore_path):
         saver_to_restore.restore(sess, args.restore_path)
         saver_to_restore.restore(sess, args.restore_path)
+    else:
+        raise ValueError('there is no model to evaluate. You should move/create the checkpoint file to restore path')
 
 
     print('\nStart evaluation...\n')
     print('\nStart evaluation...\n')
 
 

@@ -22,18 +22,18 @@ pred_scores_flag = tf.placeholder(tf.float32, [1, None, None])
 gpu_nms_op = gpu_nms(pred_boxes_flag, pred_scores_flag, args.class_num, args.nms_topk, args.score_threshold, args.nms_threshold)
 gpu_nms_op = gpu_nms(pred_boxes_flag, pred_scores_flag, args.class_num, args.nms_topk, args.score_threshold, args.nms_threshold)
 
 
 ### tf.data pipeline
 ### tf.data pipeline
-train_dataset = tf.data.TFRecordDataset(filenames=train_file, compression_type='GZIP')
-train_dataset = train_dataset.shuffle(train_img_cnt)
-train_dataset = train_dataset.batch(batch_size)
+train_dataset = tf.data.TFRecordDataset(filenames=args.train_file, compression_type='GZIP')
+train_dataset = train_dataset.shuffle(args.train_img_cnt)
+train_dataset = train_dataset.batch(args.batch_size)
 train_dataset = train_dataset.map(
 train_dataset = train_dataset.map(
     lambda x: tf.py_func(get_batch_data,
     lambda x: tf.py_func(get_batch_data,
                          inp=[x, args.class_num, args.img_size, args.anchors, True, args.multi_scale_train, args.use_mix_up, args.letterbox_resize],
                          inp=[x, args.class_num, args.img_size, args.anchors, True, args.multi_scale_train, args.use_mix_up, args.letterbox_resize],
                          Tout=[tf.int64, tf.float32, tf.float32, tf.float32, tf.float32]),
                          Tout=[tf.int64, tf.float32, tf.float32, tf.float32, tf.float32]),
     num_parallel_calls=args.num_threads
     num_parallel_calls=args.num_threads
 )
 )
-train_dataset = train_dataset.prefetch(prefetech_buffer)
+train_dataset = train_dataset.prefetch(args.prefetech_buffer)
 
 
-val_dataset = tf.data.TFRecordDataset(filenames=val_file, compression_type='GZIP')
+val_dataset = tf.data.TFRecordDataset(filenames=args.val_file, compression_type='GZIP')
 val_dataset = val_dataset.batch(1)
 val_dataset = val_dataset.batch(1)
 val_dataset = val_dataset.map(
 val_dataset = val_dataset.map(
     lambda x: tf.py_func(get_batch_data,
     lambda x: tf.py_func(get_batch_data,
@@ -41,7 +41,7 @@ val_dataset = val_dataset.map(
                          Tout=[tf.int64, tf.float32, tf.float32, tf.float32, tf.float32]),
                          Tout=[tf.int64, tf.float32, tf.float32, tf.float32, tf.float32]),
     num_parallel_calls=args.num_threads
     num_parallel_calls=args.num_threads
 )
 )
-val_dataset.prefetch(prefetech_buffer)
+val_dataset.prefetch(args.prefetech_buffer)
 
 
 iterator = tf.data.Iterator.from_structure(train_dataset.output_types, train_dataset.output_shapes)
 iterator = tf.data.Iterator.from_structure(train_dataset.output_types, train_dataset.output_shapes)
 train_init_op = iterator.make_initializer(train_dataset)
 train_init_op = iterator.make_initializer(train_dataset)
@@ -71,13 +71,13 @@ saver_to_restore = tf.train.Saver(var_list=tf.contrib.framework.get_variables_to
 update_vars = tf.contrib.framework.get_variables_to_restore(include=update_part)
 update_vars = tf.contrib.framework.get_variables_to_restore(include=update_part)
 
 
 
 
-global_step = tf.Variable(float(global_step), trainable=False, collections=[tf.GraphKeys.LOCAL_VARIABLES])
+global_step = tf.Variable(float(args.global_step), trainable=False, collections=[tf.GraphKeys.LOCAL_VARIABLES])
 if use_warm_up:
 if use_warm_up:
     learning_rate = tf.cond(tf.less(global_step, train_batch_num * warm_up_epoch), 
     learning_rate = tf.cond(tf.less(global_step, train_batch_num * warm_up_epoch), 
                             lambda: learning_rate_init * global_step / (train_batch_num * warm_up_epoch),
                             lambda: learning_rate_init * global_step / (train_batch_num * warm_up_epoch),
-                            lambda: config_learning_rate(global_step - args.train_batch_num * args.warm_up_epoch))
+                            lambda: config_learning_rate(args, global_step - args.train_batch_num * args.warm_up_epoch))
 else:
 else:
-    learning_rate = config_learning_rate(global_step)
+    learning_rate = config_learning_rate(args, global_step)
 
 
 optimizer = config_optimizer(args.optimizer_name, learning_rate)
 optimizer = config_optimizer(args.optimizer_name, learning_rate)
 
 
@@ -105,7 +105,7 @@ with tf.Session() as sess:
     if os.path.exists(args.restore_path):
     if os.path.exists(args.restore_path):
         saver_to_restore.restore(sess, args.restore_path)
         saver_to_restore.restore(sess, args.restore_path)
 
 
-    print('\nStart training...\n')
+    print('\nStart training...: Total epoches =', args.total_epoches, '\n')
 
 
     best_mAP = -np.Inf
     best_mAP = -np.Inf
 
 
@@ -163,7 +163,7 @@ with tf.Session() as sess:
 
 
             # calc mAP
             # calc mAP
             rec_total, prec_total, ap_total = AverageMeter(), AverageMeter(), AverageMeter()
             rec_total, prec_total, ap_total = AverageMeter(), AverageMeter(), AverageMeter()
-            gt_dict = parse_gt_rec(args.val_file, args.img_size, args.letterbox_resize)
+            gt_dict = parse_gt_rec(args.val_file, 'GZIP', args.img_size, args.letterbox_resize)
 
 
             info = '======> Epoch: {}, global_step: {}, lr: {:.6g} <======\n'.format(epoch, __global_step, __lr)
             info = '======> Epoch: {}, global_step: {}, lr: {:.6g} <======\n'.format(epoch, __global_step, __lr)