Keras: Custom layer without inputs

Question

I want to implement a Keras custom layer without any input, just trainable weights.

Here is the code so far:

class Simple(Layer):

    def __init__(self, output_dim, **kwargs):
       self.output_dim = output_dim
       super(Simple, self).__init__(**kwargs)

    def build(self):
       self.kernel = self.add_weight(name='kernel', shape=self.output_dim, initializer='uniform', trainable=True)
       super(Simple, self).build()  

    def call(self):
       return self.kernel

    def compute_output_shape(self):
       return self.output_dim

X = Simple((1, 784))()

I am getting an error message:

__call__() missing 1 required positional argument: 'inputs'

Is there a workaround for building a custom layer without inputs in Keras?

Answer 1

You can do the following,

from tensorflow.keras.layers import Layer

class Simple(Layer):

    def __init__(self, output_dim, **kwargs):
       self.output_dim = output_dim
       super(Simple, self).__init__(**kwargs)

    def build(self, input_shapes):
       self.kernel = self.add_weight(name='kernel', shape=self.output_dim, initializer='uniform', trainable=True)
       super(Simple, self).build(input_shapes)  

    def call(self, inputs):
       return self.kernel

    def compute_output_shape(self):
       return self.output_dim

X = Simple((1, 784))([])
print(X.shape)

Which produces

>>> (1, 784)