stockfish-nnue-2020-08-30

Changes:

• Merged the official Stockfish master branch. (242a7d9)
• gensfen Don't allow LMP on PvNodes. Thank you, jhellis3! (#80)
• convert_bin Added check_invalid_fen option. It checks if fen is valid or not. Thank you, tttak! (#84)
• convert_bin Changed to compare "Piece placement field" between input_fen and pos.fen(). Thank you, tttak! (#84)
• learn Added winning_probability_coefficient option. It specifies the coefficient to calculate the winning probability from a value. (#71)
• convert_bin learn Added src_score_min_value, src_score_max_value, dest_score_min_value and dest_score_max_value options. They specify the scaling factor of scores. (#71)
• convert_bin Supported new score formats. Thank you, tttak! (#88)
  • { [%eval 0.25] [%clk 0:10:00] }
  • { +0.71/22 1.2s }
  • { 0.60 }
  • { book }
  • no eval
  • others
• convert_bin Added convert_no_eval_fens_as_score_zero option. If specifiyes the behavior to convert a fen whose score is {book], no eval or others. Thank you, tttak! (#88)
  • convert_no_eval_fens_as_score_zero 0 : Skip the conversion of the fen.
  • convert_no_eval_fens_as_score_zero 1 : Convert the fen as score 0.
• learn Added convert_teacher_signal_to_winning_probability option to convert the teacher signals to winning probabilities. (#71)
  • For example, the following command converts pgn files containing winning probabilities to a bin file, and train a net from the winning probabilities directly.
    • convert_bin src_score_min_value 0.0 src_score_max_value 1.0 dest_score_min_value -30000 dest_score_max_value 30000 ...
    • learn src_score_min_value -30000 src_score_max_value 30000 dest_score_min_value 0.0 dest_score_max_value 1.0 convert_teacher_signal_to_winning_probability 0 ...
• gensfen learn Added use_raw_nnue_eval option. It specifies if the training data generator or the trainer uses the raw NNUE eval values. (#89)
  • When initial training data are generated from the classic eval for supervised learning, set use_raw_nnue_eval to 0 and Use NNUE to 0.
  • When an initial net is trained for supervised learning, set use_raw_nnue_eval to 1 and Use NNUE to 1.
  • When training data are generated from NNUE for reinforcement learning, set use_raw_nnue_eval to 1 and Use NNUE to 1.
  • When a net is trained for reinforcement learning, set use_raw_nnue_eval to 1 and Use NNUE to 1.
• learn Added use_wdl option. It uses WDL to convert a score to a winning probability instead of sigmoid function. Thank you, tttak! (#90)
• Fixed a build error in Ubuntu 18.04. Thank you, hero2017! (#92)

The network parameter file is not included in this release. Please download the top most net file on fishtest (https://tests.stockfishchess.org/nns). If there are questions, ask in the Stockfish unofficial Discord server (https://discord.gg/ZzJwPv3) at first.

filename	CPU instructions	Supported CPUs
x86-64	Intel Streaming SIMD Extensions 2	>= AMD Athlon 64
x86-64-ssse3	Intel Supplemental Streaming SIMD Extensions 3	>= Intel Pentium 4 (Prescott model F)
x86-64-sse3-popcnt	Intel Streaming SIMD Extensions 3 + POPCNT	>= AMD Phenom
x86-64-modern	Intel Streaming SIMD Extensions 4.1	>= Intel Core 2 Duo (Penryn)
x86-64-sse41-popcnt	Intel Streaming SIMD Extensions 4.1 + POPCNT	>= AMD Bobcat
x86-64-avx2	Intel Advanced Vector Extensions 2	>= AMD APU (Carrizo)
x86-64-bmi2	Intel Advanced Vector Extensions 2 + BMI2	>= Intel Core i7 (Haswell)
x86-64-avx512	Intel Advanced Vector Extensions 512	>= Intel Core i9/i7 (Skylake-X)
x86-64-vnni256	Intel Advanced Vector Extensions 2 + BMI2 + VNNI	>= Intel Xeon (Cascade Lake)
x86-64-vnni512	Intel Advanced Vector Extensions 512 + BMI2 + VNNI	>= Intel Core i7/i5/i3 (Ice Lake)

filename	network
halfkp_256x2-32-32	halfkp_256x2-32-32

filename	usage
learn	training data generator + trainer with blas (training is fast)
profile-learn	training data generator + trainer with blas (training is fast) and PGO build