Custom activation with parameter

Question

I'm trying to create an activation function in Keras that can take in a parameter beta like so:

from keras import backend as K
from keras.utils.generic_utils import get_custom_objects
from keras.layers import Activation

class Swish(Activation):

    def __init__(self, activation, beta, **kwargs):
        super(Swish, self).__init__(activation, **kwargs)
        self.__name__ = 'swish'
        self.beta = beta


def swish(x):
    return (K.sigmoid(beta*x) * x)

get_custom_objects().update({'swish': Swish(swish, beta=1.)})

It runs fine without the beta parameter, but how can I include the parameter in the activation definition? I also want this value to be saved when I do model.to_json() like for ELU activation.

---

Update: I wrote the following code based on @today's answer:

from keras.layers import Layer
from keras import backend as K

class Swish(Layer):
    def __init__(self, beta, **kwargs):
        super(Swish, self).__init__(**kwargs)
        self.beta = K.cast_to_floatx(beta)
        self.__name__ = 'swish'

    def call(self, inputs):
        return K.sigmoid(self.beta * inputs) * inputs

    def get_config(self):
        config = {'beta': float(self.beta)}
        base_config = super(Swish, self).get_config()
        return dict(list(base_config.items()) + list(config.items()))

    def compute_output_shape(self, input_shape):
        return input_shape

from keras.utils.generic_utils import get_custom_objects
get_custom_objects().update({'swish': Swish(beta=1.)})
gnn = keras.models.load_model("Model.h5")
arch = gnn.to_json()
with open(directory + 'architecture.json', 'w') as arch_file:
    arch_file.write(arch)

However, it does not currently save the beta value in the .json file. How can I make it save the value?

Answer 1

Since you want to save the parameters of activation function when serializing the model, I think it is better to define the activation function as a layer like the advanced activations which have been defined in Keras. You can do it like this:

from keras.layers import Layer
from keras import backend as K

class Swish(Layer):
    def __init__(self, beta, **kwargs):
        super(Swish, self).__init__(**kwargs)
        self.beta = K.cast_to_floatx(beta)

    def call(self, inputs):
        return K.sigmoid(self.beta * inputs) * inputs

    def get_config(self):
        config = {'beta': float(self.beta)}
        base_config = super(Swish, self).get_config()
        return dict(list(base_config.items()) + list(config.items()))

    def compute_output_shape(self, input_shape):
        return input_shape

Then you can use it the same way you use a Keras layer:

# ...
model.add(Swish(beta=0.3))

Since get_config() method has been implemented in its definition, the parameter beta would be saved when using methods like to_json() or save().