Rather than adding 2 extra 1x1 conv's to reduce channels for the 3x3 conv's, try just reducing the output channels of the first conv in the residual block.
This should have better latency and is easier to reuse existing lc0 optimizations, but won't reduce the weight count as much...
Rather than adding 2 extra 1x1 conv's to reduce channels for the 3x3 conv's, try just reducing the output channels of the first conv in the residual block. This should have better latency and is easier to reuse existing lc0 optimizations, but won't reduce the weight count as much...