keras - 自定义 Keras 层失败

Question

我想要自定义 Keras 层，实现两个模型的输出分配不同的权重，权重可以训练如下

prediction1=model1.output
prediction2=model2.output
class WeightedSum(Layer):
    def __init__(self,**kwargs):
        super(WeightedSum, self).__init__(**kwargs)
    def build(self, input_shape):
        self.weights =K.variable(np.random.random(1))
        self.trainable_weights=[self.weights]
    def call(self, two_model_outputs):
        return self.weights * two_model_outputs[0] + (1 - self.weights) * two_model_outputs[1]
    def compute_output_shape(self, input_shape):
        return input_shape[0]
final_pred=WeightedSum()([prediction1,prediction2])

但我写错了，不知道该怎么做。

    Traceback (most recent call last):
      File "test-paper3.py", line 182, in <module>
        final_pred=WeightedSum()([prediction1,prediction2])
      File "/root/anaconda3/lib/python3.7/site-packages/keras/engine/base_layer.py", line 431, in __call__
        self.build(unpack_singleton(input_shapes))
      File "test-paper3.py", line 162, in build
        self.weights =K.variable(np.random.random(1))
    AttributeError: can't set attribute

标签： kerasneural-networkkeras-layer

Answer 1

也许 Keras 是在保护自己不让你使用它认为以某种方式保留的词？

尝试以标准方式添加权重并使用另一个变量名称：

def build(self, input_shape):
    self.kernel = self.add_weight(name='kernel', 
                                  shape=(1,),
                                  initializer='uniform',
                                  #I suggest a constraint here, see below
                                  trainable=True)

    #this works as an initializer for the weights
    K.set_value(self.kernel, np.array([0.5])) 
        #you can use np.random here, but it seems safer to go with 0.5

    #this tells keras that the layer is build in fact
    super(WeightedSum, self).build(shapes)

当然，您需要在方法中weights替换为。kernelcall

---

不相关：

我建议您还使用约束将内核保持在 0 和 1 之间。

from keras.constraints import MinMaxNorm


........
    self.kernel = self.add_weight(name='kernel', 
                                  shape=(1,),
                                  initializer='uniform',
                                  constraint = MinMaxNorm(0,1)
                                  trainable=True)
........