tf.train.Example 记录

在阅读代码时候看到这个片段.

example = dataset_utils.image_to_tfexample(
    image_data, b'jpg', height, width, class_id)
tfrecord_writer.write(example.SerializeToString())

内部实现是这样的.

def image_to_tfexample(image_data, image_format, height, width, class_id):
  return tf.train.Example(features=tf.train.Features(feature={
      'image/encoded': bytes_feature(image_data),
      'image/format': bytes_feature(image_format),
      'image/class/label': int64_feature(class_id),
      'image/height': int64_feature(height),
      'image/width': int64_feature(width),
  }))

这里涉及的知识点有这些.

• tf.train.BytesList / tf.train.Int64List / tf.train.FloatList
• tf.train.Feature / tf.train.Features
• tf.train.Example

首先是第一类,他总共三个,分别处理不同数据.

import tensorflow as tf

a = 0.1
b = 3
c = "hello"

tf_a = tf.train.FloatList(value=[a])
tf_b = tf.train.Int64List(value=[b])
tf_c = tf.train.BytesList(value=[bytes(c, encoding='utf-8')])

print([tf_a, tf_b, tf_c])

输出内容.

[value: 0.10000000149011612
, value: 3
, value: "hello"
]

通过Feature做Features,代码如下.

feature_dict = {
    "a": tf.train.Feature(float_list=tf_a),
    "b": tf.train.Feature(int64_list=tf_b),
    "c": tf.train.Feature(bytes_list=tf_c)
}

features = tf.train.Features(feature=feature_dict)

print(features)

结果:

feature {
  key: "a"
  value {
    float_list {
      value: 0.10000000149011612
    }
  }
}
feature {
  key: "b"
  value {
    int64_list {
      value: 3
    }
  }
}
feature {
  key: "c"
  value {
    bytes_list {
      value: "hello"
    }
  }
}

最后引出主角Example方法,通常为了储存,都会序列化他.

example = tf.train.Example(features=features)
example_str = example.SerializeToString()

结果当然是一串byte.

还原方法就是用FromString,逐步还原,不过这个就是pb格式文件,如果按照特定格式做也就是TFRecord.