Tomasz Sobczyk

Home * People * Tomasz Sobczyk

Tomasz Sobczyk, (Sopel97, Sopel) a Polish computer scientist involved in recent Stockfish development and documentation concerning NNUE . He introduced new net architectures using the HalfKA, HalfKAv2 and HalfKAv2_hm  feature sets, optimized the NNUE inference code for various SIMD instruction sets  , and contributed to Gary Linscott's Pytorch NNUE project to train Stockfish, mostly working on optimizations which allowed nets to be trained within hours instead of days . He also heavily contributed to the original NNUE trainer and data generator by Yu Nasu, cleaning up the codebase, updating the data generator, optimizing the trainer, and adding other useful tools which are now available in the tools branch of the official-stockfish repository. He also introduced the Binpack storage format for training data, comprising of position evaluations from chess games, which utilizes efficient delta encoding, and reduces the sizes of the datasets by 10 to 20 times compared to the previous solutions. His contributions span more than 200 pull requests over multiple repositories. He's also an author of a chess engine Fat Titz, which is based on CFish and is a parody of Fat Fritz.

=Forum Posts=

2020 ...

 * Re: Removing Large Arrays by Sopel, CCC, March 12, 2020
 * Re: Compiler Optimization Question by Sopel, CCC, April 14, 2020
 * Re: An actual interesting computer chess read about FF2 by Sopel, CCC, February 28, 2021
 * Re: It walks like a clone, it quacks like a clone ... by Sopel, CCC, March 14, 2021
 * Re: larger nets for SF? by Sopel, CCC, April 16, 2021
 * Re: NNUE Question - King Placements by Sopel, CCC, July 01, 2021 » NNUE Question
 * Fat Titz 1.0 released by Sopel, CCC, Aug 26, 2021
 * Fat Titz 1.1 released! by Sopel, CCC, Aug 31, 2021

2022 ...

 * Fat Titz 2 released! by Tomasz Sobczyk, CCC, January 13, 2022

=External Links=
 * Sopel97 (Tomasz Sobczyk) · GitHub

Stockfish

 * GitHub - Sopel97/Stockfish: UCI chess engine
 * Update default net to nn-8a08400ed089.nnue by Sopel97 · Pull Request #3474 · official-stockfish/Stockfish · GitHub
 * Reduce the number of accumulator states from 3 to 2. Make the intent of the states clearer. by Sopel97 · Pull Request #3548 · official-stockfish/Stockfish · GitHub
 * New NNUE architecture and net by Sopel97 · Pull Request #3646 · official-stockfish/Stockfish · GitHub

Pytorch NNUE

 * GitHub - Sopel97/nnue-pytorch: NNUE (Chess evaluation) trainer in Pytorch
 * Defer data preparation to native code. Use sparse input tensors. by Sopel97 · Pull Request #1 · glinscott/nnue-pytorch · GitHub
 * A custom kernel for the feature transformer. by Sopel97 · Pull Request #96 · glinscott/nnue-pytorch · GitHub
 * Update and enhance information about additional feature factors. by Sopel97 · Pull Request #109 · glinscott/nnue-pytorch · GitHub
 * Update trainer to the new architecture. by Sopel97 · Pull Request #110 · glinscott/nnue-pytorch · GitHub

=References= Up one level