Tensorflow-@tf_export详解

@tf_export为函数取了个名字！

---

Tensorflow经常看到定义的函数前面加了“@tf_export”。例如，tensorflow/python/platform/app.py中有：

@tf_export('app.run')
def run(main=None, argv=None):
  """Runs the program with an optional 'main' function and 'argv' list."""

  # Define help flags.
  _define_help_flags()

  # Parse known flags.
  argv = flags.FLAGS(_sys.argv if argv is None else argv, known_only=True)

  main = main or _sys.modules['__main__'].main

  # Call the main function, passing through any arguments
  # to the final program.
  _sys.exit(main(argv))

 

首先，@tf_export是一个修饰符。修饰符的本质是一个函数，不懂可以撮戳这里。

tf_export的实现在tensorflow/python/util/tf_export.py中：

tf_export = functools.partial(api_export, api_name=TENSORFLOW_API_NAME)

  等号的右边的理解分两步：

  1.functools.partial

  2.api_export

   functools.partial是偏函数,它的本质简而言之是为函数固定某些参数。如：functools.partial(FuncA, p1)的作用是把函数FuncA的第一个参数固定为p1；又如functools.partial(FuncB, key1="Hello")的作用是把FuncB中的参数key1固定为“Hello"。  

    functools.partial(api_export, api_name=TENSORFLOW_API_NAME)的意思是把api_export的api_name这个参数固定为TENSORFLOW_API。其中TENSORFLOW_API_NAME = 'tensorflow'。

  api_export是实现了__call__()函数的类，不懂戳这里，简而言之是把类变得可以像函数一样调用。

  tf_export=unctools.partial(api_export, api_name=TENSORFLOW_API_NAME)的写法等效于：  

funcC = api_export(api_name=TENSORFLOW_API_NAME)
tf_export = funcC

对于funcC = api_export(api_name=TENSORFLOW_API_NAME)，会导致__init__(api_name=TENSORFLOW_API_NAME)被调用:

  def __init__(self, *args, **kwargs):
    self._names = args
    self._names_v1 = kwargs.get('v1', args)
    self._api_name = kwargs.get('api_name', TENSORFLOW_API_NAME)
    self._overrides = kwargs.get('overrides', [])
    self._allow_multiple_exports = kwargs.get('allow_multiple_exports', False)

  其中第4行self._api_name=kwargs.get('api_name', TENSORFLOW_API_NAME)的意思是获取api_name这个参数，如果未检测到该参数，则默认为TENSORFLOW_API_NAME。由此看，api_name这个参数传进来和默认的值都是TENSORFLOW_API_NAME,最终的结果是self._api_name=TENSORFLOW_API_NAME。

  然后调用像函数一样调用funcC()实际上就会调用__call__()：

  def __call__(self, func):
    api_names_attr = API_ATTRS[self._api_name].names       -----1
    api_names_attr_v1 = API_ATTRS_V1[self._api_name].names
    # Undecorate overridden names
    for f in self._overrides:
      _, undecorated_f = tf_decorator.unwrap(f)
      delattr(undecorated_f, api_names_attr)
      delattr(undecorated_f, api_names_attr_v1)

    _, undecorated_func = tf_decorator.unwrap(func)       -----2
    self.set_attr(undecorated_func, api_names_attr, self._names)  ----3
    self.set_attr(undecorated_func, api_names_attr_v1, self._names_v1)
    return func

因此@tf_export("app.run")最终的结果是用上面这个__call__()来作为修饰器。这是一个带参数的修饰器（真心有点复杂）！

 标注1：

  api_names_attr = API_ATTRS[self._api_name].names: 中的self._api_name即为__init__()中提到的TENSORFLOW_API_NAME。看看API_ATTRS中都有些什么：

_Attributes = collections.namedtuple(
    'ExportedApiAttributes', ['names', 'constants'])

# Attribute values must be unique to each API.
API_ATTRS = {
    TENSORFLOW_API_NAME: _Attributes(
        '_tf_api_names',
        '_tf_api_constants'),
    ESTIMATOR_API_NAME: _Attributes(
        '_estimator_api_names',
        '_estimator_api_constants')
}

collections.namedtuple()返回具有命名字段的元组的新子类。从“ExportedApiAttributes”可推测这是用来管理已输出的API的属性的。

  标注2：

    _, undecorated_func = tf_decorator.unwrap(func)

def unwrap(maybe_tf_decorator):
  """Unwraps an object into a list of TFDecorators and a final target.

  Args:
    maybe_tf_decorator: Any callable object.

  Returns:
    A tuple whose first element is an list of TFDecorator-derived objects that
    were applied to the final callable target, and whose second element is the
    final undecorated callable target. If the `maybe_tf_decorator` parameter is
    not decorated by any TFDecorators, the first tuple element will be an empty
    list. The `TFDecorator` list is ordered from outermost to innermost
    decorators.
  """
  decorators = []
  cur = maybe_tf_decorator
  while True:
    if isinstance(cur, TFDecorator):
      decorators.append(cur)
    elif hasattr(cur, '_tf_decorator'):
      decorators.append(getattr(cur, '_tf_decorator'))
    else:
      break
    cur = decorators[-1].decorated_target
  return decorators, cur

将对象展开到tfdecorator列表和最终目标列表中。undecorated_func获得的返回对象就是我们有@tf_export修饰的函数。

标注3：self.set_attr(undecorated_func, api_names_attr, self._names) 设置属性。

 

总结：@tf_export修饰器为所修饰的函数取了个名字！