关于tensorflow,load_model函数加载带有'leaky_relu'激活函数bug处理
今天在使用
tensorflow2.5(其他版本这个问题有没有不知道)的load_model函数加载之前保存的模型时发生报错，报错详情如下：
问题展示
Traceback (most recent call last):
File "C:\Users\Administrator\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\keras\saving\save.py", line 202, in load_model
compile)
File "C:\Users\Administrator\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\keras\saving\hdf5_format.py", line 181, in load_model_from_hdf5
custom_objects=custom_objects)
File "C:\Users\Administrator\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\keras\saving\model_config.py", line 59, in model_from_config
return deserialize(config, custom_objects=custom_objects)
File "C:\Users\Administrator\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\keras\layers\serialization.py", line 163, in deserialize
printable_module_name='layer')
File "C:\Users\Administrator\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\keras\utils\generic_utils.py", line 672, in deserialize_keras_object
list(custom_objects.items())))
File "C:\Users\Administrator\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\keras\engine\sequential.py", line 498, in from_config
custom_objects=custom_objects)
File "C:\Users\Administrator\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\keras\layers\serialization.py", line 163, in deserialize
printable_module_name='layer')
File "C:\Users\Administrator\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\keras\utils\generic_utils.py", line 675, in deserialize_keras_object
deserialized_obj = cls.from_config(cls_config)
File "C:\Users\Administrator\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\keras\engine\base_layer.py", line 740, in from_config
return cls(**config)
File "C:\Users\Administrator\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\keras\layers\convolutional.py", line 669, in __init__
activation=activations.get(activation),
File "C:\Users\Administrator\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\util\dispatch.py", line 206, in wrapper
return target(*args, **kwargs)
File "C:\Users\Administrator\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\keras\activations.py", line 587, in get
return deserialize(identifier)
File "C:\Users\Administrator\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\util\dispatch.py", line 206, in wrapper
return target(*args, **kwargs)
File "C:\Users\Administrator\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\keras\activations.py", line 550, in deserialize
printable_module_name='activation function')
File "C:\Users\Administrator\AppData\Roaming\Python\Python36\site-packages\tensorflow\python\keras\utils\generic_utils.py", line 703, in deserialize_keras_object
.format(printable_module_name, object_name))
ValueError: Unknown activation function: leaky_relu. Please ensure this object is passed to the `custom_objects` argument. See https://www.tensorflow.org/guide/keras/save_and_serialize#registering_the_custom_object for details.
其中最关键的信息就是ValueError: Unknown activation function: leaky_relu. Please ensure this object is passed to the 'custom_objects' argument，意思是: ValueError: Unknown activation function: leaky_relu。请确保将此对象传递给'custom objects'参数
这里开启调试模式重新运行看看
在
tensorflow.python.keras.utils.generic_utils.deserialize_keras_object这个函数中将字符串leaky_relu转为对应的tensorflow.python.ops.nn_ops.leaky_relu，首先它判断是否是自定义对象，然后判断是否是全局自定义对象，最后才在module_objects字典对象中找，我们看看这个dict有什么。
在
module_objects字典对象中没有leaky_relukey，所以找不到obj就会赋值为None，后面就报错了。
解决方法
从源代码中可以发现解决方法有多种：
- 直接使用
module_object字典有的做激活函数，如：relu。这样就是要重新训练模型。 - 将
leaky_relu当作自定义的函数，然后通过load_model的custom_objects参数传入。 - 使用
LeakyRelu层，不使用层内激活函数，直接使用层。
对于上面第二种解决方法，可以直接看源码：
image.png
对于上面最后一种解决方法，可以直接看源码：
以上就是最后一种解决方法的由来。
如果模型训练费时很久，不想重新训练，就可以使用第二种解决方法。
