Update main network to nn-71d6d32cb962.nnue
Using a net trained with Quantization Aware Training (QAT)
See:
- https://github.com/official-stockfish/nnue-pytorch/pull/477
- https://github.com/vondele/nettest/pull/346
CHANGES: To allow QAT some small changes had to be made.
- Quant scheme now only uses perfect power of two scales for anything but the very last output.
- Final output conversion happens in int64 to completely eliminate the chance of overflows.
REFACTORING: Some small refactoring was done to improve maintainability.
- Different weight scales are supported for different layers 64 or 128 instead of only 64).
- Changed order of weight segments in ft weights to simplify serialization.
Passed STC
LLR: 2.93 (-2.94,2.94) <0.00,2.00>
Total: 23904 W: 6271 L: 5970 D: 11663
Ptnml(0-2): 87, 2742, 5995, 3039, 89
https://tests.stockfishchess.org/tests/view/6a0f6911818cacc1db0abdea
Passed LTC
LLR: 3.01 (-2.94,2.94) <0.50,2.50>
Total: 82950 W: 21276 L: 20857 D: 40817
Ptnml(0-2): 50, 8976, 23015, 9373, 61
https://tests.stockfishchess.org/tests/view/6a0fe40d818cacc1db0abe88
closes https://github.com/official-stockfish/Stockfish/pull/6856
Bench: 3003571
Co-authored-by: Joost VandeVondele <Joost.VandeVondele@gmail.com>