源码分析——迁移学习Inception V3网络重训练实现图片分类

1. 前言

近些年来，随着以卷积神经网络（CNN）为代表的深度学习在图像识别领域的突破，越来越多的图像识别算法不断涌现。在去年，我们初步成功尝试了图像识别在测试领域的应用：将网站样式错乱问题、无线领域机型适配问题转换为“特定场景下的正常图片和异常图片的二分类问题”，并借助Goolge开源的Inception V3网络进行迁移学习，重训练出对应场景下的图片分类模型，问题图片的准确率达到95%以上。

过去一年，我们在图片智能识别做的主要工作包括：

• 模型的落地和参数调优
• 模型的服务化
• 模型服务的优化（包括数据库连接池的引入、gunicorn容器的引入、docker化等）

本篇文章主要是对模型重训练的源码进行学习和分析，加深对模型训练过程的理解，以便后续在对模型训练过程进行调整时有的放矢。

这边对迁移学习做个简单解释：图像识别往往包含数以百万计的参数，从头训练需要大量打好标签的图片，还需要大量的计算力（往往数百小时的GPU时间）。对此，迁移学习是一个捷径，它可以在已经训练好的相似工作模型基础上，继续训练新的模型。

2. retrain.py源码分析

目前我们使用的图像智能服务，对于迁移学习的代码，是参考的开源代码 github: tensorflow/hub/image_retraining/retrain.py。

下面是对源码的学习和解读：

2.1 执行主入口main：

if __name__ == '__main__':
  parser = argparse.ArgumentParser()
  parser.add_argument(
      '--image_dir',
      type=str,
      default='',
      help='Path to folders of labeled images.'
  )
  parser.add_argument(
      '--output_graph',
      type=str,
      default='/tmp/output_graph.pb',
      help='Where to save the trained graph.'
  )
  ......省略......
  parser.add_argument(
      '--logging_verbosity',
      type=str,
      default='INFO',
      choices=['DEBUG', 'INFO', 'WARN', 'ERROR', 'FATAL'],
      help='How much logging output should be produced.')
  FLAGS, unparsed = parser.parse_known_args()
  tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)

可以看到，程序main入口主要是对输入参数的声明和解析，实际执行时传入的参数会存入到FLAGS变量中，然后执行 tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)开始正式训练。

2.2 main(_)方法

def main(_):
  # Needed to make sure the logging output is visible.
  # See https://github.com/tensorflow/tensorflow/issues/3047

  ## 设置log级别
  logging_verbosity = logging_level_verbosity(FLAGS.logging_verbosity)
  tf.logging.set_verbosity(logging_verbosity)

  ## 判断image_dir参数是否传入，该参数表示用于训练的图片集路径
  if not FLAGS.image_dir:
    tf.logging.error('Must set flag --image_dir.')
    return -1

  # Prepare necessary directories that can be used during training
  ## 重建summaries_dir，并确保intermediate_output_graphs_dir存在
  prepare_file_system()

  # Look at the folder structure, and create lists of all the images.
  ## 根据输入的图片集路径、测试图片占比、验证图片占比来划分输入的图集，将图集划分为训练集、测试集、验证集
  image_lists = create_image_lists(FLAGS.image_dir, FLAGS.testing_percentage,
                                   FLAGS.validation_percentage)

  ## 根据image_dir下的子目录个数，判断要分类的数量。每个子目录为一个类别，每个类别会各自分为训练集、测试集、验证集。如果类别数为0或1，则返回错误，因为分类问题至少要有2个类。
  class_count = len(image_lists.keys())
  if class_count == 0:
    tf.logging.error('No valid folders of images found at ' + FLAGS.image_dir)
    return -1
  if class_count == 1:
    tf.logging.error('Only one valid folder of images found at ' +
                     FLAGS.image_dir +
                     ' - multiple classes are needed for classification.')
    return -1

  # See if the command-line flags mean we're applying any distortions.
  ## 根据传入的参数判断是否要对图片进行一些调整
  do_distort_images = should_distort_images(
      FLAGS.flip_left_right, FLAGS.random_crop, FLAGS.random_scale,
      FLAGS.random_brightness)

  # Set up the pre-trained graph.
  ## 载入module，默认使用inception v3，可以用参数--tfhub_module调整为使用其他已训练的模型
  module_spec = hub.load_module_spec(FLAGS.tfhub_module)
  ## 创建模型图graph
  graph, bottleneck_tensor, resized_image_tensor, wants_quantization = (
      create_module_graph(module_spec))

  # Add the new layer that we'll be training.
  ## 调用add_final_retrain_ops方法获得训练步骤、交叉熵、瓶颈输入、真实的输入、最终的tensor
  with graph.as_default():
    (train_step, cross_entropy, bottleneck_input,
     ground_truth_input, final_tensor) = add_final_retrain_ops(
         class_count, FLAGS.final_tensor_name, bottleneck_tensor,
         wants_quantization, is_training=True)

  with tf.Session(graph=graph) as sess:
    # Initialize all weights: for the module to their pretrained values,
    # and for the newly added retraining layer to random initial values.
    ## 初始化变量
    init = tf.global_variables_initializer()
    sess.run(init)

    # Set up the image decoding sub-graph.
    ## 调用图片解码操作的函数获得输入的图片tensor和解码后的图片tensor
    jpeg_data_tensor, decoded_image_tensor = add_jpeg_decoding(module_spec)

    if do_distort_images:
      # We will be applying distortions, so set up the operations we'll need.
      (distorted_jpeg_data_tensor,
       distorted_image_tensor) = add_input_distortions(
           FLAGS.flip_left_right, FLAGS.random_crop, FLAGS.random_scale,
           FLAGS.random_brightness, module_spec)
    else:
      # We'll make sure we've calculated the 'bottleneck' image summaries and
      # cached them on disk.
      ## 创建各个image的bottlenecks，并缓存到磁盘disk
      cache_bottlenecks(sess, image_lists, FLAGS.image_dir,
                        FLAGS.bottleneck_dir, jpeg_data_tensor,
                        decoded_image_tensor, resized_image_tensor,
                        bottleneck_tensor, FLAGS.tfhub_module)

    # Create the operations we need to evaluate the accuracy of our new layer.
    ## 创建评估的operation
    evaluation_step, _ = add_evaluation_step(final_tensor, ground_truth_input)

    # Merge all the summaries and write them out to the summaries_dir
    ## 将summary merge并写到summaries_dir目录下
    merged = tf.summary.merge_all()
    train_writer = tf.summary.FileWriter(FLAGS.summaries_dir + '/train',
                                         sess.graph)

    validation_writer = tf.summary.FileWriter(
        FLAGS.summaries_dir + '/validation')

    # Create a train saver that is used to restore values into an eval graph
    # when exporting models.
    train_saver = tf.train.Saver()

    # Run the training for as many cycles as requested on the command line.
    ## 根据传入的迭代次数，开始训练
    for i in range(FLAGS.how_many_training_steps):
      # Get a batch of input bottleneck values, either calculated fresh every
      # time with distortions applied, or from the cache stored on disk.
      if do_distort_images:
        (train_bottlenecks,
         train_ground_truth) = get_random_distorted_bottlenecks(
             sess, image_lists, FLAGS.train_batch_size, 'training',
             FLAGS.image_dir, distorted_jpeg_data_tensor,
             distorted_image_tensor, resized_image_tensor, bottleneck_tensor)
      else:
        ## 获取用于training的图片bottlenecks值，默认train_batch_size=100，即每次迭代会批量取100张图片进行训练
        (train_bottlenecks,
         train_ground_truth, _) = get_random_cached_bottlenecks(
             sess, image_lists, FLAGS.train_batch_size, 'training',
             FLAGS.bottleneck_dir, FLAGS.image_dir, jpeg_data_tensor,
             decoded_image_tensor, resized_image_tensor, bottleneck_tensor,
             FLAGS.tfhub_module)
      # Feed the bottlenecks and ground truth into the graph, and run a training
      # step. Capture training summaries for TensorBoard with the `merged` op.
      ## 执行merge操作，并用feed_dict的内容填充placeholder
      train_summary, _ = sess.run(
          [merged, train_step],
          feed_dict={bottleneck_input: train_bottlenecks,
                     ground_truth_input: train_ground_truth})
      train_writer.add_summary(train_summary, i)

      # Every so often, print out how well the graph is training.
      ## 判断是否最后一步训练
      is_last_step = (i + 1 == FLAGS.how_many_training_steps)

      ## 默认eval_step_interval=10，即每训练10次或训练全部完成，打印一下当前的训练结果
      if (i % FLAGS.eval_step_interval) == 0 or is_last_step:
      ## 打印训练精确度和交叉熵
        train_accuracy, cross_entropy_value = sess.run(
            [evaluation_step, cross_entropy],
            feed_dict={bottleneck_input: train_bottlenecks,
                       ground_truth_input: train_ground_truth})
        tf.logging.info('%s: Step %d: Train accuracy = %.1f%%' %
                        (datetime.now(), i, train_accuracy * 100))
        tf.logging.info('%s: Step %d: Cross entropy = %f' %
                        (datetime.now(), i, cross_entropy_value))
        # TODO: Make this use an eval graph, to avoid quantization
        # moving averages being updated by the validation set, though in
        # practice this makes a negligable difference.
        ## 获取验证集的图片的bottleneck值，也是每批次取100
        validation_bottlenecks, validation_ground_truth, _ = (
            get_random_cached_bottlenecks(
                sess, image_lists, FLAGS.validation_batch_size, 'validation',
                FLAGS.bottleneck_dir, FLAGS.image_dir, jpeg_data_tensor,
                decoded_image_tensor, resized_image_tensor, bottleneck_tensor,
                FLAGS.tfhub_module))
        # Run a validation step and capture training summaries for TensorBoard
        # with the `merged` op.
        validation_summary, validation_accuracy = sess.run(
            [merged, evaluation_step],
            feed_dict={bottleneck_input: validation_bottlenecks,
                       ground_truth_input: validation_ground_truth})
        validation_writer.add_summary(validation_summary, i)

        ## 打印验证集的测试精确度和测试的图片数
        tf.logging.info('%s: Step %d: Validation accuracy = %.1f%% (N=%d)' %
                        (datetime.now(), i, validation_accuracy * 100,
                         len(validation_bottlenecks)))

      # Store intermediate results
      ## 存储瞬时结果
      intermediate_frequency = FLAGS.intermediate_store_frequency

      if (intermediate_frequency > 0 and (i % intermediate_frequency == 0)
          and i > 0):
        # If we want to do an intermediate save, save a checkpoint of the train
        # graph, to restore into the eval graph.
        train_saver.save(sess, CHECKPOINT_NAME)
        intermediate_file_name = (FLAGS.intermediate_output_graphs_dir +
                                  'intermediate_' + str(i) + '.pb')
        tf.logging.info('Save intermediate result to : ' +
                        intermediate_file_name)
        save_graph_to_file(intermediate_file_name, module_spec,
                           class_count)

    # After training is complete, force one last save of the train checkpoint.
    train_saver.save(sess, CHECKPOINT_NAME)

    # We've completed all our training, so run a final test evaluation on
    # some new images we haven't used before.
    ## 执行最终的评估
    run_final_eval(sess, module_spec, class_count, image_lists,
                   jpeg_data_tensor, decoded_image_tensor, resized_image_tensor,
                   bottleneck_tensor)

    # Write out the trained graph and labels with the weights stored as
    # constants.
    tf.logging.info('Save final result to : ' + FLAGS.output_graph)
    if wants_quantization:
      tf.logging.info('The model is instrumented for quantization with TF-Lite')
    save_graph_to_file(FLAGS.output_graph, module_spec, class_count)
    with tf.gfile.GFile(FLAGS.output_labels, 'w') as f:
      f.write('\n'.join(image_lists.keys()) + '\n')

    ## 保存训练的graph
    if FLAGS.saved_model_dir:
      export_model(module_spec, class_count, FLAGS.saved_model_dir)

main方法中的一些细节解释已经用中文备注在上述代码（使用“##”开头）中，它的主要步骤是：

• 设置log级别
• 准备workspace
• 从image_dir载入输入图片集，并创建image_lists，该image_lists是一个字段，key为各个类别，value为对应类别的图集（包含训练集、测试集、验证集，划分比例默认为0.8、0.1、0.1）
• 载入在ImageNet上已经训练好的Inception V3网络的特征张量
• 针对每个图片，调用图片解码操作获得图片的原始张量和解码后张量
• 针对每个图片的jpeg_data_tensor和decoded_image_tensor，创建其对应的bottlenects（实际上是1*2048维的张量），并缓存到磁盘
• 获取训练步骤、交叉熵
• 开始迭代训练
• 每迭代10次，打印训练的精度和交叉熵，打印验证集的测试结果。默认情况下训练集和测试集都是取100张图
• 训练完成后，使用测试集进行最后的评估
• 结果的打印和保存

2.3 其它方法

分析完代码的主要执行路径，下面解读下其它方法。因为总的代码非常的长，篇幅有限，下面按照顺序简单介绍下其它方法的内容。

2.3.1 create_image_lists

def create_image_lists(image_dir, testing_percentage, validation_percentage):
    ...... 省略......
	result[label_name] = {
        'dir': dir_name,
        'training': training_images,
        'testing': testing_images,
        'validation': validation_images,
    }
  return result

根据image_dir的地址，testing_percentage和testing_percentage的比例划分图集，返回的格式类似如下：

{
	'correct': {
		'dir': correct_image_dir,
		'training': correct_training_images,
		'testing': correct_testing_images,
		'validation': correct_validation_images
	},
	'error': {
		'dir': error_image_dir,
		'training': error_training_images,
		'testing': error_testing_images,
		'validation': error_validation_images
	}
}

每个training/testing/validation对应的value为image的file_name list。

2.3.2 get_image_path

获取图片的全路径

2.3.3 get_bottleneck_path

获得不同类别（training、testing、validation）的bottleneck路径

2.3.4 create_module_graph

根据给定的已训练好的模型Hub Module，创建模型的图

2.3.5 run_bottleneck_on_image

def run_bottleneck_on_image(sess, image_data, image_data_tensor,
                            decoded_image_tensor, resized_input_tensor,
                            bottleneck_tensor):
  """Runs inference on an image to extract the 'bottleneck' summary layer.
  Args:
    sess: Current active TensorFlow Session.
    image_data: String of raw JPEG data.
    image_data_tensor: Input data layer in the graph.
    decoded_image_tensor: Output of initial image resizing and preprocessing.
    resized_input_tensor: The input node of the recognition graph.
    bottleneck_tensor: Layer before the final softmax.
  Returns:
    Numpy array of bottleneck values.
  """
  # First decode the JPEG image, resize it, and rescale the pixel values.
  resized_input_values = sess.run(decoded_image_tensor,
                                  {image_data_tensor: image_data})
  # Then run it through the recognition network.
  bottleneck_values = sess.run(bottleneck_tensor,
                               {resized_input_tensor: resized_input_values})
  bottleneck_values = np.squeeze(bottleneck_values)
  return bottleneck_values

根据给定的输入图片解码后的tensor，计算bottleneck_values，并执行squeeze操作（删除单维度条目，把shape中为1的维度去掉）

2.3.6 ensure_dir_exists

确保目录存在：如果目录不存在，则创建目录

2.3.7 create_bottleneck_file

调run_bottleneck_on_image方法计算bottleneck值，并缓存到磁盘文件

2.3.8 get_or_create_bottleneck

批量获取一组图片的bottleneck值

2.3.9 cache_bottlenecks

批量缓存bottleneck

2.3.10 get_random_cached_bottlenecks

随机获取一批缓存的bottlenecks，以及其对应的真实标ground_truths和文件名filenames

2.3.11 add_final_retrain_ops

def add_final_retrain_ops(class_count, final_tensor_name, bottleneck_tensor,
                          quantize_layer, is_training):

  batch_size, bottleneck_tensor_size = bottleneck_tensor.get_shape().as_list()
  assert batch_size is None, 'We want to work with arbitrary batch size.'
  with tf.name_scope('input'):
    bottleneck_input = tf.placeholder_with_default(
        bottleneck_tensor,
        shape=[batch_size, bottleneck_tensor_size],
        name='BottleneckInputPlaceholder')

    ground_truth_input = tf.placeholder(
        tf.int64, [batch_size], name='GroundTruthInput')

  # Organizing the following ops so they are easier to see in TensorBoard.
  layer_name = 'final_retrain_ops'
  with tf.name_scope(layer_name):
    with tf.name_scope('weights'):
      initial_value = tf.truncated_normal(
          [bottleneck_tensor_size, class_count], stddev=0.001)
      layer_weights = tf.Variable(initial_value, name='final_weights')
      variable_summaries(layer_weights)

    with tf.name_scope('biases'):
      layer_biases = tf.Variable(tf.zeros([class_count]), name='final_biases')
      variable_summaries(layer_biases)

    with tf.name_scope('Wx_plus_b'):
      logits = tf.matmul(bottleneck_input, layer_weights) + layer_biases
      tf.summary.histogram('pre_activations', logits)

  final_tensor = tf.nn.softmax(logits, name=final_tensor_name)

  # The tf.contrib.quantize functions rewrite the graph in place for
  # quantization. The imported model graph has already been rewritten, so upon
  # calling these rewrites, only the newly added final layer will be
  # transformed.
  if quantize_layer:
    if is_training:
      tf.contrib.quantize.create_training_graph()
    else:
      tf.contrib.quantize.create_eval_graph()

  tf.summary.histogram('activations', final_tensor)

  # If this is an eval graph, we don't need to add loss ops or an optimizer.
  if not is_training:
    return None, None, bottleneck_input, ground_truth_input, final_tensor

  with tf.name_scope('cross_entropy'):
    cross_entropy_mean = tf.losses.sparse_softmax_cross_entropy(
        labels=ground_truth_input, logits=logits)

  tf.summary.scalar('cross_entropy', cross_entropy_mean)

  with tf.name_scope('train'):
    optimizer = tf.train.GradientDescentOptimizer(FLAGS.learning_rate)
    train_step = optimizer.minimize(cross_entropy_mean)

  return (train_step, cross_entropy_mean, bottleneck_input, ground_truth_input,
          final_tensor)

在结尾处添加一个新的softmax层和全连接层（y=WX+b），用于训练和评估。此处与logistic模型是一样的，采用梯度下降的方式来最小化交叉熵进行迭代训练。

2.3.12 add_evaluation_step

def add_evaluation_step(result_tensor, ground_truth_tensor):
  with tf.name_scope('accuracy'):
    with tf.name_scope('correct_prediction'):
	  ## 对每组向量按列找到最大值的index
      prediction = tf.argmax(result_tensor, 1)
	  ## 将每组张量比较预测的结果和实际的结果的一致性，一致则为True，否则为False
      correct_prediction = tf.equal(prediction, ground_truth_tensor)
    with tf.name_scope('accuracy'):
	  ## 将True或False转为float格式，并计算平均值
      evaluation_step = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
  tf.summary.scalar('accuracy', evaluation_step)
  return evaluation_step, prediction

注解见上述代码，返回最终的accuracy和预测的值list。

2.3.13 run_final_eval

执行最终的评估，使用测试集进行结果评估。如果传入参数print_misclassified_test_images，则会打印评估出错的图片的名字和识别结果。

2.3.14 save_graph_to_file

将graph保存到文件

2.3.15 prepare_file_system

准备workspace

2.3.16 add_jpeg_decoding

将输入图片解析为张量，并进行解码

2.3.17 export_model

输出模型