HCha/Stockfish
1
0
Fork 0
mirror of https://github.com/HChaZZY/Stockfish.git synced 2025-12-24 02:57:11 +08:00
Code Issues Packages Projects Releases Wiki Activity

Use plain nnue eval for validation loss calculation instead of first performing qsearch

Browse Source
This commit is contained in:
Tomasz Sobczyk
2020-12-21 20:35:23 +01:00
committed by nodchip
parent 6853b4aac2
commit 8ca82646a9
2 changed files with 2 additions and 31 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
src/learn/learn.cpp
View File

@@ -611,8 +611,6 @@ namespace Learner
atomic<int>& move_accord_count atomic<int>& move_accord_count
); );
Value get_shallow_value(Position& pos);
bool check_progress(); bool check_progress();
// save merit function parameters to a file // save merit function parameters to a file
@@ -996,7 +994,7 @@ namespace Learner
continue; continue;
} }
const Value shallow_value = get_shallow_value(pos); const Value shallow_value = Eval::evaluate(pos);
// Evaluation value of deep search // Evaluation value of deep search
const auto deep_value = (Value)ps.score; const auto deep_value = (Value)ps.score;
@@ -1018,33 +1016,6 @@ namespace Learner
test_loss_sum += local_loss_sum; test_loss_sum += local_loss_sum;
} }
Value LearnerThink::get_shallow_value(Position& pos)
{
// Evaluation value for shallow search
// The value of evaluate() may be used, but when calculating loss, learn_cross_entropy and
// Use qsearch() because it is difficult to compare the values.
// EvalHash has been disabled in advance. (If not, the same value will be returned every time)
const auto [_, pv] = Search::qsearch(pos);
const auto rootColor = pos.side_to_move();
std::vector<StateInfo, AlignedAllocator<StateInfo>> states(pv.size());
for (size_t i = 0; i < pv.size(); ++i)
{
pos.do_move(pv[i], states[i]);
}
const Value shallow_value =
(rootColor == pos.side_to_move())
? Eval::evaluate(pos)
: -Eval::evaluate(pos);
for (auto it = pv.rbegin(); it != pv.rend(); ++it)
pos.undo_move(*it);
return shallow_value;
}
bool LearnerThink::check_progress() bool LearnerThink::check_progress()
{ {
auto out = sync_region_cout.new_region(); auto out = sync_region_cout.new_region();

2
tests/instrumented_learn.sh
View File

@@ -124,7 +124,7 @@ cat << EOF > learn01.exp
send "uci\n" send "uci\n"
send "setoption name SkipLoadingEval value true\n" send "setoption name SkipLoadingEval value true\n"
send "setoption name Use NNUE value true\n" send "setoption name Use NNUE value pure\n"
send "setoption name Threads value $threads\n" send "setoption name Threads value $threads\n"
send "isready\n" send "isready\n"
send "learn targetdir training_data epochs 1 sfen_read_size 100 thread_buffer_size 10 batchsize 100 use_draw_in_training 1 use_draw_in_validation 1 lr 1 eval_limit 32000 nn_batch_size 30 newbob_decay 0.5 eval_save_interval 30 loss_output_interval 10 validation_set_file_name validation_data/validation_data.bin\n" send "learn targetdir training_data epochs 1 sfen_read_size 100 thread_buffer_size 10 batchsize 100 use_draw_in_training 1 use_draw_in_validation 1 lr 1 eval_limit 32000 nn_batch_size 30 newbob_decay 0.5 eval_save_interval 30 loss_output_interval 10 validation_set_file_name validation_data/validation_data.bin\n"