training command :
Code: Select all
stockfish.bmi2.halfkp_256x2-32-32.nnue-learn.2020-07-19.exe
uci
setoption name SkipLoadingEval value true
setoption name Threads value 36
isready
learn targetdir training loop 100 batchsize 1000000 eta 1 lambda 1 eval_limit 32000 nn_batch_size 1000 newbob_decay 0.5 eval_save_interval 500000000 loss_output_interval 125000000 mirror_percentage 50 validation_set_file_name validation\1m_d18.bin
NN2.BIN (evalsave6_rejected)
sfens : 3 375 000 000
test_cross_entropy : 0.252433
move accuracy : 33.1046%
loss : 0.0466118
Code: Select all
# PLAYER : RATING ERROR POINTS PLAYED (%) W D L D(%) OppAvg OppN
1 2000m_d14_nn2_evalsave6_rejected : 61 11 1168.5 2000 58.4 601 1135 264 56.8 0 1
2 2000m_d14_nn2_evalsave3_rejected : 46 11 1127.0 2000 56.4 532 1190 278 59.5 0 1
3 2000m_d14_nn2_evalsave5_rejected : 43 12 1118.5 2000 55.9 574 1089 337 54.5 0 1
4 2000m_d14_nn2_evalsave2 : 35 11 1098.5 2000 54.9 526 1145 329 57.3 0 1
5 2000m_d14_nn2_evalsave4 : 35 11 1097.5 2000 54.9 485 1225 290 61.3 0 1
6 2000m_d14_nn2_evalsave1 : 26 11 1072.0 2000 53.6 512 1120 368 56.0 0 1
7 2000m_d14_nn2_evalsave0 : 12 11 1033.5 2000 51.7 484 1099 417 55.0 0 1
8 stockfish 190720 no-nnue : 0 ---- 6284.5 14000 44.9 2283 8003 3714 57.2 37 7
White advantage = 55.13 +/- 2.18
Draw rate (equal opponents) = 50.00 % +/- 0.00
evalsave elo curves :
When i saw that the first evalsave was even stronger than the base engine, it smelt good...
evalsave loss cuves :
It's weird as the same loss values can produce so different strong nets. But the trend is here : when the loss values decrease, it smells good.
At the moment, i'm testing if the training can use more than 3.5B sfens thanks to the 1m d16 or 1m d19 as validation data :