HCha/Stockfish
1
0
Fork 0
mirror of https://github.com/HChaZZY/Stockfish.git synced 2025-12-25 03:26:24 +08:00
Code Issues Packages Projects Releases Wiki Activity

Removed SGD_UPDATE macro.

Browse Source
This commit is contained in:
nodchip
2020-09-09 00:07:09 +09:00
parent 0271d70775
commit f3a158725d
2 changed files with 2 additions and 61 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
src/learn/learn.h
View File

@@ -5,14 +5,6 @@
#include <vector>
// ----------------------
// update formula
// ----------------------
// SGD looking only at the sign of the gradient. It requires less memory, but the accuracy is...
// #define SGD_UPDATE
// ----------------------
// Select the objective function
// ----------------------
@@ -134,10 +126,6 @@ typedef float LearnFloatType;
#define LOSS_FUNCTION_IS_ELMO_METHOD
#if defined(SGD_UPDATE)
#define LEARN_UPDATE "SGD"
#endif
#if defined(LOSS_FUNCTION_IS_WINNING_PERCENTAGE)
#define LOSS_FUNCTION "WINNING_PERCENTAGE"
#elif defined(LOSS_FUNCTION_IS_CROSS_ENTOROPY)

51
src/learn/learning_tools.h
View File

@@ -4,13 +4,12 @@
// A set of machine learning tools related to the weight array used for machine learning of evaluation functions
#include "learn.h"
#if defined (EVAL_LEARN)
#include <array>
#if defined(SGD_UPDATE) || defined(USE_KPPP_MIRROR_WRITE)
#include "../misc.h" // PRNG , my_insertion_sort
#endif
#include <array>
#include <cmath> // std::sqrt()
namespace EvalLearningTools
@@ -29,14 +28,6 @@ namespace EvalLearningTools
// cumulative value of one mini-batch gradient
LearnFloatType g = LearnFloatType(0);
// When ADA_GRAD_UPDATE. LearnFloatType == float,
// total 4*2 + 4*2 + 1*2 = 18 bytes
// It suffices to secure a Weight array that is 4.5 times the size of the evaluation function parameter of 1GB.
// However, sizeof(Weight)==20 code is generated if the structure alignment is in 4-byte units, so
// Specify pragma pack(2).
// For SGD_UPDATE, this structure is reduced by 10 bytes to 8 bytes.
// Learning rate η(eta) such as AdaGrad.
// It is assumed that eta1,2,3,eta1_epoch,eta2_epoch have been set by the time updateFV() is called.
// The epoch of update_weights() gradually changes from eta1 to eta2 until eta1_epoch.
@@ -76,44 +67,6 @@ namespace EvalLearningTools
template <typename T> void updateFV(T& v) { updateFV(v, 1.0); }
#if defined(SGD_UPDATE)
// See only the sign of the gradient Update with SGD
// When executing this function, the value of g and the member do not change
// Guaranteed by the caller. It does not have to be an atomic operation.
template <typename T>
void updateFV(T & v , double k)
{
if (g == 0)
return;
// See only the sign of g and update.
// If g <0, add v a little.
// If g> 0, subtract v slightly.
// Since we only add integers, no decimal part is required.
// It's a good idea to move around 0-5.
// It is better to have a Gaussian distribution, so generate a 5-bit random number (each bit has a 1/2 probability of 1),
// Pop_count() it. At this time, it has a binomial distribution.
//int16_t diff = (int16_t)POPCNT32((u32)prng.rand(31));
// → If I do this with 80 threads, this AsyncPRNG::rand() locks, so I slowed down. This implementation is not good.
int16_t diff = 1;
double V = v;
if (g > 0.0)
V-= diff;
else
V+= diff;
V = (std::min)((double)(std::numeric_limits<T>::max)(), V);
V = (std::max)((double)(std::numeric_limits<T>::min)(), V);
v = (T)V;
}
#endif
// grad setting
template <typename T> void set_grad(const T& g_) { g = g_; }