training doesn't properly reset state #207
Description
doing
setoption name SkipLoadingEval value true
setoption name Threads value 4
setoption name Use NNUE value pure
learn targetdir data_x epochs 1 batchsize 100000 use_draw_in_training 1 use_draw_in_validation 1 lr 1 lambda 1 eval_limit 32000 nn_batch_size 1000 newbob_decay 0.99 eval_save_interval 1000000 loss_output_interval 100000 set_recommended_uci_options
(using epochs based training from https://github.com/Sopel97/Stockfish/tree/cyclic_reader to stop learning after reasonable time)
twice in a row results in weird activations.
INFO: observed 23356 (out of 43979) features
INFO: (min, max) of pre-activations = -0.0883445, 1.06719 (limit = 258.008)
INFO: largest min activation = 3.40282e+38, smallest max activation = -3.40282e+
38
INFO: largest min activation = 0.316678, smallest max activation = 0.675339
INFO: largest min activation = 0.149582, smallest max activation = 0.653783
it may be a display only issue, but I'm not sure.