Converter synapse not applied to neurons-to-TensorNode connections

[xchoo]

If you convert a Keras model where there is a connection from a neurons layer (e.g., dense, conv2D, etc.) to a layer that cannot be natively converted to a Nengo object (i.e., has to be replaced by a TensorNode), the connection to the TensorNode does not have the converter's synapse value applied to it.
Note: The NengoDL converter docs state:

Synaptic filter to be applied on the output of all neurons.

Minimal code to reproduce issue:

inp = tf.keras.Input(shape=(28, 28, 1))
conv0 = tf.keras.layers.Conv2D(
    filters=32,
    kernel_size=3,
    activation=tf.nn.relu,
)(inp)
max_pool = tf.keras.layers.MaxPool2D()(conv0)
# max_pool = tf.keras.layers.AvgPool2D()(conv0)  # Uncomment to see expected output
model = tf.keras.Model(inputs=inp, outputs=max_pool)

converter = nengo_dl.Converter(model, synapse=0.1)

# Find connection wehere pre is conv0
for conn in converter.net.all_connections:
    if conn.pre_obj == converter.layers[conv0]:
        break
print(conn, conn.synapse)

Expected output: <Connection from <Neurons of <Ensemble 'conv2d.0'>> to <TensorNode 'max_pooling2d'>> Lowpass(tau=0.1)
Observed output: <Connection from <Neurons of <Ensemble 'conv2d.0'>> to <TensorNode 'max_pooling2d'>> None

To see the difference between a TensorNode and native Nengo object, change MaxPool2D layer to AvgPool2D.

Probable fix:
The ConvertFallback function needs to be made aware of a non-default synapse value if the input_obj object is an Ensemble:

    def convert(self, node_id):
        input_obj = self.get_input_obj(node_id)

        # Identify the correct synapse value to use
        synapse = (
            self.converter.synapse
            if isinstance(input_obj, nengo.ensemble.Neurons)
            else None
        )

        output = self.tensor_layer(
            input_obj,
            shape_in=self.input_shape(node_id),
            label=self.layer.name,
            synapse=synapse,  # Pass synapse value to TensorNode
        )

[R-Gaurav]

Just wondering if I should keep an eye for a future release of NengoDL where this would be fixed? For now I am explicitly synapsing the connections :) .