tensorflow_probability子类化JointDistributionNamed __init__行为

Question

我正在尝试在tensorflow_probability库(tensorflow v2.0.0，tensorflow_probability v0.8.0)中从JointDistributionNamed创建一个派生类。然而，super().__init__函数以一种我不理解的奇怪方式运行。也许我只是错误地使用了super()，但对于其他类，它似乎可以像我期望的那样工作。不管怎样，这里有一个例子：

from tensorflow_probability import distributions as tfd

models = {'normal': tfd.Normal(loc=0, scale=1)}
joint = tfd.JointDistributionNamed(models) # Works perfectly fine
print("joint:",joint) 

class Test(tfd.JointDistributionNamed):
    def __init__(self,name,models):
        self.myname = name
        self.models = models
        super().__init__(models) #(1) Works
        #super().__init__(self.models) #(2) Doesn't work

t = Test('hello',models)
print("t:", t)

将models赋给super().__init__时的行为是不同的，无论我是直接传入models还是先将其赋给self.models。为什么？在后一种情况下，我得到以下错误：

Traceback (most recent call last):
  File "test_jointdistnamed.py", line 18, in <module>
    t = Test('hello',models)
  File "</home/farmer/anaconda3/envs/tensorflow/lib/python3.7/site-packages/decorator.py:decorator-gen-244>", line 2, in __init__
  File "/home/farmer/anaconda3/envs/tensorflow/lib/python3.7/site-packages/tensorflow_probability/python/distributions/distribution.py", line 276, in wrapped_init
    default_init(self_, *args, **kwargs)
  File "test_jointdistnamed.py", line 16, in __init__
    super().__init__(self.models) #doesn't work
  File "</home/farmer/anaconda3/envs/tensorflow/lib/python3.7/site-packages/decorator.py:decorator-gen-138>", line 2, in __init__
  File "/home/farmer/anaconda3/envs/tensorflow/lib/python3.7/site-packages/tensorflow_probability/python/distributions/distribution.py", line 276, in wrapped_init
    default_init(self_, *args, **kwargs)
  File "/home/farmer/anaconda3/envs/tensorflow/lib/python3.7/site-packages/tensorflow_probability/python/distributions/joint_distribution_named.py", line 170, in __init__
    model, validate_args, name or 'JointDistributionNamed')
  File "</home/farmer/anaconda3/envs/tensorflow/lib/python3.7/site-packages/decorator.py:decorator-gen-70>", line 2, in __init__
  File "/home/farmer/anaconda3/envs/tensorflow/lib/python3.7/site-packages/tensorflow_probability/python/distributions/distribution.py", line 276, in wrapped_init
    default_init(self_, *args, **kwargs)
  File "/home/farmer/anaconda3/envs/tensorflow/lib/python3.7/site-packages/tensorflow_probability/python/distributions/joint_distribution_sequential.py", line 211, in __init__
    self._model_unflatten(self._model_flatten(model))
  File "/home/farmer/anaconda3/envs/tensorflow/lib/python3.7/site-packages/tensorflow_probability/python/distributions/joint_distribution_named.py", line 186, in _model_unflatten
    return type(self.model)(**kwargs)
TypeError: __init__() got an unexpected keyword argument 'normal'

由于某种原因，它就像是试图在类层次结构中的某个地方解压models字典。但是，为什么这会因为我是否首先赋值给self而不同呢？无论哪种方式，我传递的不是对完全相同的字典的引用吗？有什么不同？这是一个奇怪的bug，还是我做错了什么？如果我编写自己的简单定制类而不是JointDistributionNamed，同样的事情似乎也可以很好地工作。

Answer 1

我的直觉是与字段的tf.Module依赖跟踪包装器对象有关。什么是type(self.models)？如果您更改为self._models = self._no_dependency(models)，它可以工作吗？