使用WeightNormalization附加层时出现TypeError问题求助

问题复现 #

以下代码在Colab环境中执行时触发TypeError，单独使用Dense层无异常，问题出在WeightNormalization包装器：

import numpy as np
import tensorflow as tf
import tensorflow_addons as addon_layers

x = np.random.rand(1, 10, 10, 1)
dense = addon_layers.WeightNormalization(tf.keras.layers.Dense(10), data_init=False)
y = dense(x)
y.shape

报错信息 #

完整报错堆栈：

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-47-a2d44c85c018> in <module>
      2 
      3 x = np.random.rand(1, 10, 10, 1)
----> 4 dense = addon_layers.WeightNormalization(tf.keras.layers.Dense(10))
      5 y = dense(x)
      6 y.shape

2 frames
/usr/local/lib/python3.9/dist-packages/tensorflow_addons/layers/wrappers.py in __init__(self, layer, data_init, **kwargs)
     57 
     58     @typechecked
---&gt; 59     def __init__(self, layer: tf.keras.layers, data_init: bool = True, **kwargs):
     60         super().__init__(layer, **kwargs)
     61         self.data_init = data_init

/usr/local/lib/python3.9/dist-packages/typeguard/_functions.py in check_argument_types(memo)
    111             value = memo.arguments[argname]
    112             try:
--&gt; 113                 check_type_internal(value, expected_type, memo=memo)
    114             except TypeCheckError as exc:
    115                 qualname = qualified_name(value, add_class_prefix=True)

/usr/local/lib/python3.9/dist-packages/typeguard/_checkers.py in check_type_internal(value, annotation, memo)
    668             return
    669 
--&gt; 670     if not isinstance(value, origin_type):
    671         raise TypeCheckError(f"is not an instance of {qualified_name(origin_type)}")
    672 

TypeError: isinstance() arg 2 must be a type or tuple of types

问题原因 #

错误根源是TensorFlow Addons的WeightNormalization类在__init__方法的类型注解中，错误地将layer参数的类型指定为tf.keras.layers（模块），而非正确的tf.keras.layers.Layer（层类）。这导致typeguard库在执行类型检查时，无法用isinstance验证实例是否属于一个模块，从而抛出TypeError。

解决方法 #

方法一：降级TensorFlow Addons版本 #

该bug在0.20.0及以上版本出现，降级到0.19.0版本即可修复：

pip install tensorflow-addons==0.19.0

方法二：临时禁用类型检查 #

如果需要保留当前版本，可在代码开头添加以下语句关闭typeguard的类型检查：

from typeguard import typechecked
typechecked.off()

内容的提问来源于stack exchange，提问作者Bhavesh