Wrap general python function in tensorflow

My goal is to create a custom layer in keras with a weight matrix that is larger than the input (output size is the same), while additional parameters are used to handle the function (weight %%%). The function is NOT written in tensorflow and it would be quite tedious to do so, so I tried using tensorflow py_func. However, NN aborts with "ValueError: None values ​​not supported".

The master code for my own layer is Dense with a few tweaks. W - matrix of weights with dimension [input_dim + n, output_dim]. Suppose there is a generic function "func" with n parameters, and it should be evaluated on (W% *% Input).

def func(self,output):  
      do something
      return something

      

        
        
        
      

    

We now wrap this function into a call:

def call(self, x, mask=None):
      output = K.dot(x, self.W[self.n:,:])
      if self.use_bias:
            output = K.bias_add(output, self.bias)
      if self.activation is not None:
            output = tf.reshape(tf.py_func(self.func,[output],tf.float32),(-1,self.units))
            output = self.activation(output)
      return output

      

        
        
        
      

    

The "No" error occurs in self.activation. py_func works flawlessly when using its own session, but it looks like this nesting in the Keras layer won't do the expected result. It also fails if func(x)=x*x

.