This is an arch change for Berserk which increases the network size by ~33%. The new layout is (2x1024)->16->32->1. In order to use such a large FT, Berserk now relies on sparse matrix multiplication after the FT. Further optimizations will be made to this network as it remains unsorted and sparsity untouched outside the defaults.
This is the first network in Berserk trained by Grapheus.
Bench: 4886939
This is an arch change for Berserk which increases the network size by ~33%. The new layout is (2x1024)->16->32->1. In order to use such a large FT, Berserk now relies on sparse matrix multiplication after the FT. Further optimizations will be made to this network as it remains unsorted and sparsity untouched outside the defaults.
This is the first network in Berserk trained by Grapheus.
ELO | -0.21 +- 5.17 (95%) CONF | 8.0+0.08s Threads=1 Hash=8MB GAMES | N: 8384 W: 2028 L: 2033 D: 4323 http://chess.grantnet.us/test/33528/
ELO | 6.39 +- 3.82 (95%) SPRT | 60.0+0.60s Threads=1 Hash=64MB LLR | 2.97 (-2.94, 2.94) [0.00, 3.00] GAMES | N: 14512 W: 3447 L: 3180 D: 7885 http://chess.grantnet.us/test/33529/