使用自定義層保存 Keras 模型

Question

我正在嘗試將 Keras model 保存在 H5 文件中。 Keras model 有一個自定義層。 當我嘗試恢復 model時，出現以下錯誤：

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-5-0fbff9b56a9d> in <module>()
      1 model.save('model.h5')
      2 del model
----> 3 model = tf.keras.models.load_model('model.h5')

8 frames
/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/utils/generic_utils.py in class_and_config_for_serialized_keras_object(config, module_objects, custom_objects, printable_module_name)
    319   cls = get_registered_object(class_name, custom_objects, module_objects)
    320   if cls is None:
--> 321     raise ValueError('Unknown ' + printable_module_name + ': ' + class_name)
    322 
    323   cls_config = config['config']

ValueError: Unknown layer: CustomLayer

您能否告訴我應該如何保存和加載所有自定義 Keras 層的權重？ （另外，保存時沒有警告，是否可以從我已經保存但現在無法加載的H5文件中加載模型？）

這是此錯誤的最小工作代碼示例 (MCVE)，以及完整的擴展消息： Google Colab Notebook

為了完整起見，這是我用來制作自定義層的代碼。 get_config和from_config都工作正常。

class CustomLayer(tf.keras.layers.Layer):
    def __init__(self, k, name=None):
        super(CustomLayer, self).__init__(name=name)
        self.k = k

    def get_config(self):
        return {'k': self.k}

    def call(self, input):
        return tf.multiply(input, 2)

model = tf.keras.models.Sequential([
    tf.keras.Input(name='input_layer', shape=(10,)),
    CustomLayer(10, name='custom_layer'),
    tf.keras.layers.Dense(1, activation='sigmoid', name='output_layer')
])
model.save('model.h5')
model = tf.keras.models.load_model('model.h5')

Answer 1

更正編號1是在loading Saved Model時使用Custom_Objects ，即替換代碼，

new_model = tf.keras.models.load_model('model.h5') 

和

new_model = tf.keras.models.load_model('model.h5', custom_objects={'CustomLayer': CustomLayer})

由於我們使用Custom Layers來build Model並且在Saving之前，我們應該在Loading時使用Custom Objects 。

更正2是在Custom Layer的__init__ function中添加**kwargs like

def __init__(self, k, name=None, **kwargs):
        super(CustomLayer, self).__init__(name=name)
        self.k = k
        super(CustomLayer, self).__init__(**kwargs)

完整的工作代碼如下所示：

import tensorflow as tf

class CustomLayer(tf.keras.layers.Layer):
    def __init__(self, k, name=None, **kwargs):
        super(CustomLayer, self).__init__(name=name)
        self.k = k
        super(CustomLayer, self).__init__(**kwargs)


    def get_config(self):
        config = super(CustomLayer, self).get_config()
        config.update({"k": self.k})
        return config

    def call(self, input):
        return tf.multiply(input, 2)

model = tf.keras.models.Sequential([
    tf.keras.Input(name='input_layer', shape=(10,)),
    CustomLayer(10, name='custom_layer'),
    tf.keras.layers.Dense(1, activation='sigmoid', name='output_layer')
])
tf.keras.models.save_model(model, 'model.h5')
new_model = tf.keras.models.load_model('model.h5', custom_objects={'CustomLayer': CustomLayer})

print(new_model.summary())

上述代碼的Output如下所示：

WARNING:tensorflow:No training configuration found in the save file, so the model was *not* compiled. Compile it manually.
Model: "sequential_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
custom_layer_1 (CustomLayer) (None, 10)                0         
_________________________________________________________________
output_layer (Dense)         (None, 1)                 11        
=================================================================
Total params: 11
Trainable params: 11
Non-trainable params: 0

希望這可以幫助。 快樂學習！

Answer 2

您可以在答案https://stackoverflow.com/a/62326857/8056572中提到的load_model方法中手動提供映射custom_objects但是當您有很多自定義層（或定義的任何自定義可調用對象）時，它可能會很乏味。例如指標，損失，優化器，...）。

Tensorflow 提供了一個工具 function 來自動完成： tf.keras.utils.register_keras_serializable

您必須按如下方式更新您的CustomLayer ：

import tensorflow as tf

@tf.keras.utils.register_keras_serializable()
class CustomLayer(tf.keras.layers.Layer):
    def __init__(self, k, **kwargs):
        self.k = k
        super(CustomLayer, self).__init__(**kwargs)

    def get_config(self):
        config = super().get_config()
        config["k"] = self.k
        return config

    def call(self, input):
        return tf.multiply(input, 2)

---

這是完整的工作代碼：

import tensorflow as tf


@tf.keras.utils.register_keras_serializable()
class CustomLayer(tf.keras.layers.Layer):
    def __init__(self, k, **kwargs):
        self.k = k
        super(CustomLayer, self).__init__(**kwargs)

    def get_config(self):
        config = super().get_config()
        config["k"] = self.k
        return config

    def call(self, input):
        return tf.multiply(input, 2)


def main():
    model = tf.keras.models.Sequential(
        [
            tf.keras.Input(name='input_layer', shape=(10,)),
            CustomLayer(10, name='custom_layer'),
            tf.keras.layers.Dense(1, activation='sigmoid', name='output_layer')
        ]
    )
    print("SUMMARY OF THE MODEL CREATED")
    print("-" * 60)
    print(model.summary())
    model.save('model.h5')

    del model

    print()
    print()

    model = tf.keras.models.load_model('model.h5')
    print("SUMMARY OF THE MODEL LOADED")
    print("-" * 60)
    print(model.summary())

if __name__ == "__main__":
    main()

以及對應的output：

SUMMARY OF THE MODEL CREATED
------------------------------------------------------------
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
custom_layer (CustomLayer)   (None, 10)                0         
_________________________________________________________________
output_layer (Dense)         (None, 1)                 11        
=================================================================
Total params: 11
Trainable params: 11
Non-trainable params: 0
_________________________________________________________________
None


WARNING:tensorflow:No training configuration found in the save file, so the model was *not* compiled. Compile it manually.
SUMMARY OF THE MODEL LOADED
------------------------------------------------------------
Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
custom_layer (CustomLayer)   (None, 10)                0         
_________________________________________________________________
output_layer (Dense)         (None, 1)                 11        
=================================================================
Total params: 11
Trainable params: 11
Non-trainable params: 0
_________________________________________________________________
None

Answer 3

此處提供的另一個答案在實際示例中不起作用。 代碼示例在同一進程中運行，因此加載CustomLayer時，CustomLayer 在內存中可用。 如果您嘗試在新進程中加載 model，它將失敗並要求您使用custom_objects參數。 我仍在尋找一種從源代碼中導出自定義對象的干凈方法。