tensorflow - 爱德华的源代码中 Normal 类的位置在哪里？

Question

在使用 edward 的时候，我们总是使用from edward.models import Normal，但是我没有Normal在github中找到声明

谁能告诉我它在哪里

Answer 1

They are defined in edward/models/random_variables.py.

You import the Normal class like this:

from edward.models import Normal

This suggests looking in edward/models/__init__.py, which has this line:

from edward.models.random_variables import *

Looking in edward/models/random_variables.py we find this code:

from edward.models.random_variable import RandomVariable as _RandomVariable
from tensorflow.contrib import distributions as _distributions

# Automatically generate random variable classes from classes in
# tf.contrib.distributions.
_globals = globals()
for _name in sorted(dir(_distributions)):
  _candidate = getattr(_distributions, _name)
  if (_inspect.isclass(_candidate) and
          _candidate != _distributions.Distribution and
          issubclass(_candidate, _distributions.Distribution)):

    # to use _candidate's docstring, must write a new __init__ method
    def __init__(self, *args, **kwargs):
      _RandomVariable.__init__(self, *args, **kwargs)
    __init__.__doc__ = _candidate.__init__.__doc__
    _params = {'__doc__': _candidate.__doc__,
               '__init__': __init__}
    _globals[_name] = type(_name, (_RandomVariable, _candidate), _params)

del _candidate

This goes through the tensorflow.contrib.distributions module looking for classes derived from tensorflow.contrib.distributions.Distribution (ignoring other attributes like e.g. the __file__ member of the module, or the base Distribution class itself). For each one, it does a bit of hacking (which only affects the generated documentation) then executes this key line:

_globals[_name] = type(_name, (_RandomVariable, _candidate), _params)

The type() built-in function creates a new type i.e. declares a new class. The second parameter is the list of base classes, which here is edward's RandomVariable class and the TensorFlow random variable class. Earlier it defined _globals to be globals(), which is a built-in function returning the dictionary of the module's variables. Therefore, in the case you're interested in, the line above is equivalent to the following:

from edward.models.random_variable import RandomVariable as EdRandVar
from tensorflow.contrib.distributions import Normal as TfNormal
Normal = type("Normal", (EdRandVar, TfNormal), {...})

Which in turn is equivalent to this (if you ignore the docstring stuff):

from edward.models.random_variable import RandomVariable as EdRandVar
from tensorflow.contrib.distributions import Normal as TfNormal
class Normal(EdRandVar, TfNormal):
    pass