Comment by diyer22
In DDN, 1×1 convolutions are used only in the output layers of the Discrete Distribution Layer (DDL). The NN blocks between DDLs, which supply the fundamental computational power and parameter count, adopt standard 3×3 convolutions.
Was there a specific reason for this choice?