HCha/Stockfish
1
0
Fork 0
mirror of https://github.com/HChaZZY/Stockfish.git synced 2025-12-25 11:36:51 +08:00
Code Issues Packages Projects Releases Wiki Activity

Cross entropy loss.

Browse Source
This commit is contained in:
Tomasz Sobczyk
2020-11-29 17:08:22 +01:00
committed by nodchip
parent 539bd2d1c8
commit aa55692b97
2 changed files with 51 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
src/learn/learn.cpp
View File

@@ -200,11 +200,14 @@ namespace Learner
static ValueWithGrad<double> get_loss(Value shallow, Value teacher_signal, int result, int ply)
{
using namespace Learner::Autograd::UnivariateStatic;
/*
auto q_ = sigmoid(VariableParameter<double, 0>{} * winning_probability_coefficient);
auto p_ = sigmoid(ConstantParameter<double, 1>{} * winning_probability_coefficient);
auto t_ = (ConstantParameter<double, 2>{} + 1.0) * 0.5;
auto lambda_ = ConstantParameter<double, 3>{};
auto loss_ = pow(lambda_ * (q_ - p_) + (1.0 - lambda_) * (q_ - t_), 2.0);
*/
/*
auto q_ = VariableParameter<double, 0>{};
@@ -212,6 +215,20 @@ namespace Learner
auto loss_ = pow(q_ - p_, 2.0) * (1.0 / (2400.0 * 2.0 * 600.0));
*/
const double epsilon = 1e-12;
auto q_ = sigmoid(VariableParameter<double, 0>{} * winning_probability_coefficient);
auto p_ = sigmoid(ConstantParameter<double, 1>{} * winning_probability_coefficient);
auto t_ = (ConstantParameter<double, 2>{} + 1.0) * 0.5;
auto lambda_ = ConstantParameter<double, 3>{};
auto teacher_entropy_ = -(p_ * log(p_ + epsilon) + (1.0 - p_) * log(1.0 - p_ + epsilon));
auto outcome_entropy_ = -(t_ * log(t_ + epsilon) + (1.0 - t_) * log(1.0 - t_ + epsilon));
auto teacher_loss_ = -(p_ * log(q_) + (1.0 - p_) * log(1.0 - q_));
auto outcome_loss_ = -(t_ * log(q_) + (1.0 - t_) * log(1.0 - q_));
auto result_ = lambda_ * teacher_loss_ + (1.0 - lambda_) * outcome_loss_;
auto entropy_ = lambda_ * teacher_entropy_ + (1.0 - lambda_) * outcome_entropy_;
auto loss_ = result_ - entropy_;
auto args = std::tuple((double)shallow, (double)teacher_signal, (double)result, calculate_lambda(teacher_signal));
return loss_.eval(args);
}