This implements propagation of sparse input layer invented by AndrovT for Stockfish.
It also switches to new network which is functional identical but with input weights sorted by activation probability.
Speed varies from "almost identical" to ~8% faster depending on the CPU.
This implements propagation of sparse input layer invented by AndrovT for Stockfish. It also switches to new network which is functional identical but with input weights sorted by activation probability. Speed varies from "almost identical" to ~8% faster depending on the CPU.