It doesnt seem like the decoder prenet matches the architecture described in the paper. Can anyone shed light on why this is?