找不到匹配功能来调用已加载模型

Question

我使用自定义对象训练了一个深神网络，并能够成功训练它，但是当我保存并尝试重新加载模型以进行再培训时，我将帖子的主题视为错误。 我已经尝试使用TensorFlow创建一个get_config来

class SkipCon(keras.layers.Layer):
  def __init__(self, size, reduce = True, deep = 3, skip_when=0, activation="relu", **kwargs):
   
    super().__init__(**kwargs)
    self.activation = keras.activations.get(activation) # used to combine
    # skip connections and cascaded dense layers
    self.main_layers =[]
    self.skip_when = skip_when #to be used in call as a control
    if reduce:
      for _ in range(deep):
        self.main_layers.extend([
          keras.layers.Dense(size, activation=activation, 
                              use_bias=True),
          keras.layers.BatchNormalization()])

        # Reduce the input size by two each time, if the
        # network is to be designed deeper and narrow
        size = size/2
    else:
      for _ in range(deep):
        self.main_layers.extend([
        keras.layers.Dense(size, activation=activation, 
                            use_bias=True),
        keras.layers.BatchNormalization()])
        
    self.skip_layers = []
    if skip_when > 0:
      if reduce:
        size = size*2 # since the size of skipped connection  
                      # should match with cascaded dense
      self.skip_layers = [
          keras.layers.Dense(size, activation=activation, 
                          use_bias=True),
          keras.layers.BatchNormalization()]

  def call(self, inputs):
    Z = inputs
    for layer in self.main_layers:
      Z = layer(Z)
    if not self.skip_when:
      return self.activation(Z)
    skip_Z = inputs
    for layer in self.skip_layers:
      skip_Z = layer(skip_Z)
    return self.activation(Z + skip_Z)
  
  def get_config(self):
        return {"Z": self.activation}
  @classmethod  
  def from_config(cls, config):
        return cls(**config)

跳过连接创建的

user_layer = keras.layers.Dense(256, activation='relu', use_bias = True)(feature_layer_users)
# Add a Skip Connection
user_layer = SkipCon(size = 256, deep = 2, reduce = False, skip_when=1, activation="relu")(user_layer)
user_layer = keras.layers.Dropout(0.5)(user_layer) # Way to handle overfitting
user_layer = SkipCon(size = 256, deep = 2, reduce = True, skip_when=1, activation="relu")(user_layer)
user_layer = keras.layers.Dropout(0.5)(user_layer)
user_layer = SkipCon(size = 64, deep = 2, reduce = True, skip_when=0, activation="relu")(user_layer)


item_layer = keras.layers.Dense(256, activation='relu', use_bias = True)(feature_layer_items)
# Add a Skip Connection
item_layer = SkipCon(size = 256, deep = 2, reduce = False, skip_when=1, activation="relu")(item_layer)
item_layer = keras.layers.Dropout(0.5)(item_layer) # Way to handle overfitting
item_layer = SkipCon(size = 256, deep = 2, reduce = True, skip_when=1, activation="relu")(item_layer)
item_layer = keras.layers.Dropout(0.5)(item_layer)
item_layer = SkipCon(size = 64, deep = 2, reduce = True, skip_when=0, activation="relu")(item_layer)


combined = tf.concat([user_layer, item_layer], axis =-1)

both_layer = SkipCon(size = 128, deep = 2, reduce = False, skip_when=1, activation="relu")(combined)
both_layer = keras.layers.Dropout(0.5)(both_layer)
both_layer = SkipCon(size = 64, deep = 2, reduce = False, skip_when=1, activation="relu")(combined)
both_layer = keras.layers.Dropout(0.5)(both_layer)
both_layer = SkipCon(size = 16, deep = 2, reduce = False, skip_when=0, activation="relu")(both_layer)

z = keras.layers.Dense(3, activation="softmax")(both_layer)

model = keras.Model(inputs=[input_user, input_items], outputs=z)

保存模型

model.save("my_model")

重新加载模型

model = keras.models.load_model('my_model', custom_objects={"SkipCon":SkipCon})

错误：

ValueError: Could not find matching function to call loaded from the SavedModel. Got:
  Positional arguments (2 total):

Answer 1

我相信这是因为您的get_config（self）函数应与自定义层的参数匹配。我无法完全复制您的示例，因为我错过了您的输入层，但是调整GET_CONFIG功能使我可以在保存模型后正确加载模型。

class SkipCon(keras.layers.Layer):
  def __init__(self, size, reduce = True, deep = 3, skip_when=0, activation="relu", **kwargs):
   
    super().__init__(**kwargs)
    self.size = size

... 

  def get_config(self):
        return {"size": self.size}

https://github.com/tensorflow/tensorflow/tensorflow/tensorflow/37973