1.3 label_map.pbtxt檔案

xx_label_map.pbtxt檔案中的內容如下：

1. item {
2. id: 1
3. name: 'Abyssinian'
4. }
5.  
6. item {
7. id: 2
8. name: 'american_bulldog'
9. }
10.  
11. item {
12. id: 3
13. name: 'american_pit_bull_terrier'
14. }
15.  

1.4 建立tf_record檔案

先建立一個create_xx_tf_record.py

修改起始引數配置：

• data_dir: 資料目錄，包含了圖片和標註的目錄
• output_dir:輸出目錄，圖片轉換為tf_record之後儲存的位置
• label_map_path:上面提到的xx_label_map.pbtxt

修改dict_to_tf_example

參考你的標準xml檔案，有些地方需要修改。

修改main

確保你的標註檔案，圖片目錄對應的目錄。標註檔案目錄下是否存在 trainval.txt檔案是否存在，這個需要自己生成。我生成的列表（注意：沒有帶字尾）為：

執行完之後會在對應目錄下生成 tf_record檔案。

我的檔案：

# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================

r"""Convert raw PASCAL dataset to TFRecord for object_detection.

Example usage:
    python object_detection/dataset_tools/create_pascal_tf_record.py \
        --data_dir=/home/user/VOCdevkit \
        --year=VOC2012 \
        --output_path=/home/user/pascal.record
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import hashlib
import io
import logging
import os

from lxml import etree
import PIL.Image
import tensorflow as tf

from object_detection.utils import dataset_util
from object_detection.utils import label_map_util

flags = tf.app.flags
flags.DEFINE_string('data_dir', '/media/chenyu/6046436546433AD4', 'Root directory to raw PASCAL VOC dataset.')
flags.DEFINE_string('set', 'train', 'Convert training set, validation set or '
                    'merged set.')
flags.DEFINE_string('annotations_dir', 'Annotations',
                    '(Relative) path to annotations directory.')
flags.DEFINE_string('year', 'VOC2007', 'Desired challenge year.')
flags.DEFINE_string('output_path', '/home/chenyu/Documents/models/research/object_detection/date/train/pascal_train.record', 'Path to output TFRecord')
flags.DEFINE_string('label_map_path', '/home/chenyu/Documents/models/research/object_detection/date/pascal_label_map.pbtxt',
                    'Path to label map proto')
flags.DEFINE_boolean('ignore_difficult_instances', False, 'Whether to ignore '
                     'difficult instances')
FLAGS = flags.FLAGS

SETS = ['train', 'val', 'trainval', 'test']
YEARS = ['VOC2007', 'VOC2012', 'merged']

def dict_to_tf_example(data,
                       dataset_directory,
                       label_map_dict,
                       ignore_difficult_instances=False,
                       image_subdirectory='JPEGImages'):
  """Convert XML derived dict to tf.Example proto.

  Notice that this function normalizes the bounding box coordinates provided
  by the raw data.

  Args:
    data: dict holding PASCAL XML fields for a single image (obtained by
      running dataset_util.recursive_parse_xml_to_dict)
    dataset_directory: Path to root directory holding PASCAL dataset
    label_map_dict: A map from string label names to integers ids.
    ignore_difficult_instances: Whether to skip difficult instances in the
      dataset  (default: False).
    image_subdirectory: String specifying subdirectory within the
      PASCAL dataset directory holding the actual image data.

  Returns:
    example: The converted tf.Example.

  Raises:
    ValueError: if the image pointed to by data['filename'] is not a valid JPEG
  """#data['folder'], 
  img_path = os.path.join('VOC2007',image_subdirectory, data['filename'][:-4]+'.jpg')  #此處修改
  full_path = os.path.join(dataset_directory, img_path)
  with tf.gfile.GFile(full_path, 'rb') as fid:
    encoded_jpg = fid.read()
  encoded_jpg_io = io.BytesIO(encoded_jpg)
  image = PIL.Image.open(encoded_jpg_io)
  if image.format != 'JPEG':
    raise ValueError('Image format not JPEG')
  key = hashlib.sha256(encoded_jpg).hexdigest()

  width = int(data['size']['width'])
  height = int(data['size']['height'])

  xmin = []
  ymin = []
  xmax = []
  ymax = []
  classes = []
  classes_text = []
 # truncated = []
  #poses = []
  difficult_obj = []
  if 'object' in data:
    for obj in data['object']:
      difficult = bool(int(obj['difficult']))
      if ignore_difficult_instances and difficult:
        continue

      difficult_obj.append(int(difficult))

      xmin.append(float(obj['bndbox']['xmin']) / width)
      ymin.append(float(obj['bndbox']['ymin']) / height)
      xmax.append(float(obj['bndbox']['xmax']) / width)
      ymax.append(float(obj['bndbox']['ymax']) / height)
      classes_text.append(obj['name'].encode('utf8'))
      classes.append(label_map_dict[obj['name']])
     # truncated.append(int(obj['truncated']))    #此處修改
      #poses.append(obj['pose'].encode('utf8'))   #此處修改

  example = tf.train.Example(features=tf.train.Features(feature={
      'image/height': dataset_util.int64_feature(height),
      'image/width': dataset_util.int64_feature(width),
      'image/filename': dataset_util.bytes_feature(
          data['filename'].encode('utf8')),
      'image/source_id': dataset_util.bytes_feature(
          data['filename'].encode('utf8')),
      'image/key/sha256': dataset_util.bytes_feature(key.encode('utf8')),
      'image/encoded': dataset_util.bytes_feature(encoded_jpg),
      'image/format': dataset_util.bytes_feature('jpeg'.encode('utf8')),
      'image/object/bbox/xmin': dataset_util.float_list_feature(xmin),
      'image/object/bbox/xmax': dataset_util.float_list_feature(xmax),
      'image/object/bbox/ymin': dataset_util.float_list_feature(ymin),
      'image/object/bbox/ymax': dataset_util.float_list_feature(ymax),
      'image/object/class/text': dataset_util.bytes_list_feature(classes_text),
      'image/object/class/label': dataset_util.int64_list_feature(classes),
      'image/object/difficult': dataset_util.int64_list_feature(difficult_obj),
     # 'image/object/truncated': dataset_util.int64_list_feature(truncated),  #此處修改
     # 'image/object/view': dataset_util.bytes_list_feature(poses),         #此處修改
  }))
  return example

def main(_):
  if FLAGS.set not in SETS:
    raise ValueError('set must be in : {}'.format(SETS))
  if FLAGS.year not in YEARS:
    raise ValueError('year must be in : {}'.format(YEARS))

  data_dir = FLAGS.data_dir
  years = ['VOC2007', 'VOC2012']
  if FLAGS.year != 'merged':
    years = [FLAGS.year]

  writer = tf.python_io.TFRecordWriter(FLAGS.output_path)

  label_map_dict = label_map_util.get_label_map_dict(FLAGS.label_map_path)

  for year in years:
    logging.info('Reading from PASCAL %s dataset.', year)
    examples_path = os.path.join(data_dir,year, 'ImageSets', 'Main',
                                 FLAGS.set + '.txt')       #此處修改
    annotations_dir = os.path.join(data_dir, year, FLAGS.annotations_dir)  
    examples_list = dataset_util.read_examples_list(examples_path)
    for idx, example in enumerate(examples_list):
      if idx % 100 == 0:
        logging.info('On image %d of %d', idx, len(examples_list))
      path = os.path.join(annotations_dir, example + '.xml')
      print(path)
      with tf.gfile.GFile(path, 'r') as fid:
        xml_str = fid.read()
      xml = etree.fromstring(xml_str.encode('utf-8'))   #此處修改
     # xml = etree.fromstring(xml_str)
      data = dataset_util.recursive_parse_xml_to_dict(xml)['annotation']

      tf_example = dict_to_tf_example(data, FLAGS.data_dir, label_map_dict,
                                      FLAGS.ignore_difficult_instances)
      writer.write(tf_example.SerializeToString())

  writer.close()

if __name__ == '__main__':
  tf.app.run()