sizes = "None" failing for Keras DynamicGraphConvolutionKerasLayer implementations

Running Environment

• Ubuntu 19.10, TF 2.1, TF graphics built from repo ~14 days ago.

Issue

• Documentation for the Keras Layer implementations suggests that the layers can be run with sizes = None, (which is the default). Instead, if sizes = None, or the default is passed, a value error is eventually raised:
  ValueError: The channel dimension of the inputs should be defined. Found 'None'.
  This occurs in the _edge_convolution_template function of convolution/graph_convolution ln. 266

Related
This is similar in nature to the issue raised in issue 13, but differs in that:

• If input sizes are passed, conversion to dense tensor is consuming unfeasible amounts of memory,
• it appears that the layers would run with None input at the time of that issue, which is not the case here.

Minimal example reproducing the undesired behaviour

from tensorflow_graphics.nn.layer import graph_convolution as g_nn
import tensorflow as tf
from tensorflow import keras
def min_example():
    input = keras.Input(shape=(1000, 3))

    sparse = tf.sparse.reshape(tf.sparse.eye(1000), shape=[1, 1000, 1000])
    outputs = g_nn.DynamicGraphConvolutionKerasLayer(num_output_channels=64, reduction='weighted', name='EdgeConv0')(
        inputs=[input, sparse],
    )
    model = keras.Model(inputs=[input], outputs=[outputs])

    return model

if __name__ == "__main__":
    min_example()