视频的 TFRecords

Question

我正在尝试从自定义视频数据集创建 TFRecords，但我无法完全理解如何设置它们。

为了准备存储数据，我编写了一个脚本，针对给定的视频源，输出形状为 [N_FRAMES, WIDTH, HEIGHT, CHANNEL] 的 3D 立方体。此后我创建一个 tfrecord 如下：

def _int64_feature(self, value):
  return tf.train.Feature(int64_list=tf.train.Int64List(value=[value]))

def _bytes_feature(self, value):
  return tf.train.Feature(bytes_list=tf.train.BytesList(value=[value]))

def createDataRecord(self, file_name, locations, categories):
    writer = tf.python_io.TFRecordWriter(file_name)

    feature = {}

    for loc, category in zip(locations, categories):
        data = self.3DVideo(loc) # the final array of shape [N_FRAMES, WIDTH, HEIGHT, CHANNEL]

        feature['height'] = self._int64_feature(self.height)
        feature['width'] = self._int64_feature(self.width)
        feature['depth'] = self._int64_feature(self.depth)
        feature['data'] = self._bytes_feature(data.tostring())
        feature['category'] = self._int64_feature(category)

        example = tf.train.Example(features=tf.train.Features(feature=feature))
        writer.write(example.SerializeToString())

    writer.close()

然后我当前的解析器函数看起来像这样

def readDataRecord(self, record):
  filename_queue = tf.train.string_input_producer([record], num_epochs=1)

  reader = tf.TFRecordReader()
  _, serialized_example = reader.read(filename_queue)

  feature =
  {'height': tf.FixedLenFeature([], tf.int64),
    'width': tf.FixedLenFeature([], tf.int64),
    'depth': tf.FixedLenFeature([], tf.int64),
    'data': tf.FixedLenFeature([], tf.string),
    'category': tf.FixedLenFeature([], tf.int64),
  }

  example = tf.parse_single_example(serialized_example, features=feature)

  video3D_buffer = tf.reshape(example['data'], shape=[])
  video3D = tf.decode_raw(video3D_buffer, tf.uint8)

  label = tf.cast(example['category'], tf.int32)

  return video3D, label

话虽如此，我的问题是：

我知道 readDataRecord() 是错误的，因为它适用于单个帧。我如何准确地将其添加到 return 形状 [N_FRAMES, WIDTH, HEIGHT, CHANNEL] 的单个 3D 立方体及其各自的类别？
简单地保存整个 3D 立方体是个好主意吗？

任何帮助或指导将不胜感激:)

PS: 我研究了其他方法，包括 video2tfrecord 但大多数方法似乎都为每个方法保存了单独的帧视频，我不想要那个。

Answer 1

所以这就是我最终所做的，无需对单个帧进行编码。

我最终将立方体展平，然后将其写出来，如下所示：

def _cube_feature(self, value):
    return tf.train.Feature(float_list=tf.train.FloatList(value=value))

def createDataRecord(self, name, locations, categories):

    writer = tf.python_io.TFRecordWriter(name)

    feature = {}

    for loc, category in zip(locations, categories):
        data = self.3DVideo(loc)
        .............
        feature['data'] = self._cube_feature(data.flatten())
        feature['category'] = self._int64_feature(category)

        example = tf.train.Example(features=tf.train.Features(feature=feature))
        writer.write(example.SerializeToString())

    writer.close()

生成的解析器是：

def readDataRecord(self, record):
    ..........
    feature = \
    {'height': tf.FixedLenFeature([], tf.int64),
     'width': tf.FixedLenFeature([], tf.int64),
     'depth': tf.FixedLenFeature([], tf.int64),
     'data': tf.FixedLenFeature((NUM_FRAMES, WIDTH, HEIGHT, CHANNEL), tf.float32),
     'category': tf.FixedLenFeature([], tf.int64),
    }

    example = tf.parse_single_example(serialized_example, features=feature)

    cube = tf.cast(example['data'], tf.uint8)
    label = tf.cast(example['category'], tf.int32)

    return cube, label

Answer 2

已接受答案的缺点是您必须在某处存储数组的维度（NUM_FRAMES、WIDTH、HEIGHT、CHANNEL）。解决方法是使用 tf.io.serialize_tensor(array.astype(...)) 序列化整个 3D 立方体，将其作为字节字符串特征保存到 TFRecord，然后（在加载 TFRecord 之后）使用 tf.io.parse_tensor(bytestring_array_feature, out_type=...) 恢复它。在这里看到一个很好的解释：（向下滚动到关于 _bytes_feature 的段落）

视频的 TFRecords

TFRecords for videos

python-2.7

deep-learning

tensorflow

tfrecord

tensorflow-datasets