User defined layers in CNTK python

Question

I am trying to create a custom layer to use in CNTK using the python interface. I am following this guide, but keep having a TypeError exception thrown right in the __init__ of my class. Note that I just copy-pasted the example in the linked guide.

import cntk as C
import numpy as np

class MySigmoid(UserFunction):
    def __init__(self, arg, name='MySigmoid'):
        super(MySigmoid, self).__init__([arg], name=name)

    def forward(self, argument, device=None, outputs_to_retain=None):
        sigmoid_x = 1 / (1 + np.exp(-argument))
        return sigmoid_x, sigmoid_x

    def backward(self, state, root_gradients):
        sigmoid_x = state
        return root_gradients * sigmoid_x * (1 - sigmoid_x)

    def infer_outputs(self):
        return [output_variable(self.inputs[0].shape, self.inputs[0].dtype,
            self.inputs[0].dynamic_axes)]

    @staticmethod
    def deserialize(inputs, name, state):
        return MySigmoid(inputs[0], name)
model = C.layers.Sequential(C.layers.Dense(10), C.user_function(layers_extensions.MySigmoid(3)))

And this is the error I get:

  File "...\layers_extensions.py", line 30, in __init__
    super(MySigmoid, self).__init__([arg], name=name)
  File "c:\repos\cntk\bindings\python\cntk\ops\functions.py", line 1286, in __init__
    super(UserFunction, self).__init__(inputs, name)
  File "c:\repos\cntk\bindings\python\cntk\ops\functions.py", line 109, in __init__
    super(Function, self).__init__(*args, **kwargs)
  File "c:\repos\cntk\bindings\python\cntk\cntk_py.py", line 1698, in __init__
    this = _cntk_py.new_Function(_self, *args)
TypeError: cannot convert list element to CNTK::Variable

I tried to google this error but nothing comes up. Can you help me?

Answer 1

For some reason, CNTK is passing the argument parameter in the forward(...) method as a list, even if it is a single parameter. I ended up making it work by taking the first from the list. You will find the working example here.