3 ms·
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 fundame
by diyer22 1y ago
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.
- randomNumber7 1y agoWas there a specific reason for this choice?
- diyer22 1y ago1x1 convolution is the most lightweight operator for transforming features into outputs. 3x3 convolution is the most common operator used to provide basic computational power.