void CStringEx::PutFloatS(float f, int len)
{
// Code From
// Copyright SOCABIM 1999
// Pierre Caussin
// Jean-Paul Retru
switch (_fpclass(f))
{
case _FPCLASS_SNAN:
case _FPCLASS_QNAN:
Format("%s", m_csNAN); // "NaN"
return;
case _FPCLASS_NINF:
Format("%s", m_csNINF);// "-Inf"
return;
case _FPCLASS_PINF:
Format("%s", m_csPINF);// "+Inf"
return;
}
PutDoubleS((double) f, len);
int acc = 8;
if(len< acc+1) return;
Accuracy(acc);
return;
}
void OsuScore::computeSpeedValue(const Beatmap& beatmap)
{
_speedValue = pow(5.0f * std::max(1.0f, beatmap.DifficultyAttribute(_mods, Beatmap::Speed) / 0.0675f) - 4.0f, 3.0f) / 100000.0f;
int numTotalHits = TotalHits();
f32 approachRate = beatmap.DifficultyAttribute(_mods, Beatmap::AR);
f32 approachRateFactor = 1.0f;
if (approachRate > 10.33f)
approachRateFactor += 0.3f * (approachRate - 10.33f);
_speedValue *= approachRateFactor;
// Longer maps are worth more
_speedValue *=
0.95f + 0.4f * std::min(1.0f, static_cast<f32>(numTotalHits) / 2000.0f) +
(numTotalHits > 2000 ? log10(static_cast<f32>(numTotalHits) / 2000.0f) * 0.5f : 0.0f);
// Penalize misses exponentially. This mainly fixes tag4 maps and the likes until a per-hitobject solution is available
_speedValue *= pow(0.97f, _numMiss);
// Combo scaling
float maxCombo = beatmap.DifficultyAttribute(_mods, Beatmap::MaxCombo);
if (maxCombo > 0)
_speedValue *= std::min(static_cast<f32>(pow(_maxCombo, 0.8f) / pow(maxCombo, 0.8f)), 1.0f);
// We want to give more reward for lower AR when it comes to speed and HD. This nerfs high AR and buffs lower AR.
if ((_mods & EMods::Hidden) > 0)
_speedValue *= 1.0f + 0.04f * (12.0f - approachRate);
// Scale the speed value with accuracy _slightly_
_speedValue *= 0.02f + Accuracy();
// It is important to also consider accuracy difficulty when doing that
_speedValue *= 0.96f + (pow(beatmap.DifficultyAttribute(_mods, Beatmap::OD), 2) / 1600);
}
double LetterClassifier::Test(const cv::Mat &test_dataset, const cv::Mat &test_labels)
{
if (mean.empty())
LoadPCAcoeffs();
cv::Mat src, avg, trans, pred_labels;
// cv::repeat(mean, 1, test_dataset.cols, avg);
// src = test_dataset - avg;
trans = coeffs.t()*test_dataset;
cvsvm->predict(trans.t(), pred_labels);
pred_labels.convertTo(pred_labels, test_labels.type());
return Accuracy(pred_labels, test_labels);
}
void OsuScore::computeAimValue(const Beatmap& beatmap)
{
f32 rawAim = beatmap.DifficultyAttribute(_mods, Beatmap::Aim);
if ((_mods & EMods::TouchDevice) > 0)
rawAim = pow(rawAim, 0.8f);
_aimValue = pow(5.0f * std::max(1.0f, rawAim / 0.0675f) - 4.0f, 3.0f) / 100000.0f;
int numTotalHits = TotalHits();
// Longer maps are worth more
f32 LengthBonus = 0.95f + 0.4f * std::min(1.0f, static_cast<f32>(numTotalHits) / 2000.0f) +
(numTotalHits > 2000 ? log10(static_cast<f32>(numTotalHits) / 2000.0f) * 0.5f : 0.0f);
_aimValue *= LengthBonus;
// Penalize misses exponentially. This mainly fixes tag4 maps and the likes until a per-hitobject solution is available
_aimValue *= pow(0.97f, _numMiss);
// Combo scaling
float maxCombo = beatmap.DifficultyAttribute(_mods, Beatmap::MaxCombo);
if (maxCombo > 0)
_aimValue *= std::min(static_cast<f32>(pow(_maxCombo, 0.8f) / pow(maxCombo, 0.8f)), 1.0f);
f32 approachRate = beatmap.DifficultyAttribute(_mods, Beatmap::AR);
f32 approachRateFactor = 1.0f;
if (approachRate > 10.33f)
approachRateFactor += 0.3f * (approachRate - 10.33f);
else if (approachRate < 8.0f)
{
approachRateFactor += 0.01f * (8.0f - approachRate);
}
_aimValue *= approachRateFactor;
// We want to give more reward for lower AR when it comes to aim and HD. This nerfs high AR and buffs lower AR.
if ((_mods & EMods::Hidden) > 0)
_aimValue *= 1.0f + 0.04f * (12.0f - approachRate);
if ((_mods & EMods::Flashlight) > 0)
// Apply object-based bonus for flashlight.
_aimValue *= 1.0f + 0.35f * std::min(1.0f, static_cast<f32>(numTotalHits) / 200.0f) +
(numTotalHits > 200 ? 0.3f * std::min(1.0f, static_cast<f32>(numTotalHits - 200) / 300.0f) +
(numTotalHits > 500 ? static_cast<f32>(numTotalHits - 500) / 1200.0f : 0.0f) : 0.0f);
// Scale the aim value with accuracy _slightly_
_aimValue *= 0.5f + Accuracy() / 2.0f;
// It is important to also consider accuracy difficulty when doing that
_aimValue *= 0.98f + (pow(beatmap.DifficultyAttribute(_mods, Beatmap::OD), 2) / 2500);
}
void CStringEx::PutDoubleS(double value, int nsignificands)
{
// Code From
// Copyright SOCABIM 1999
// Pierre Caussin
// Jean-Paul Retru
CString fmt;
int nk;
switch (_fpclass(value))
{
case _FPCLASS_SNAN:
case _FPCLASS_QNAN:
Format("%s", m_csNAN); // "NaN"
return;
case _FPCLASS_NINF:
Format("%s", m_csNINF);// "-Inf"
return;
case _FPCLASS_PINF:
Format("%s", m_csPINF);// "+Inf"
return;
case _FPCLASS_NN:
case _FPCLASS_PN:
nk = (int)log10(fabs(value));
if (fabs(value)<1) nk--;
if (nk > nsignificands+1 || nk < 2-nsignificands)
{
fmt.Format("%%.%dfE%d",nsignificands-1,nk);
Format(fmt,value/pow(10.,nk));
}
else
{
fmt.Format("%%.%df",nsignificands>nk+1?nsignificands-nk-1:0);
Format(fmt,value);
}
return;
default:
fmt.Format("%%.%df",nsignificands);
Format(fmt,0.);
}
int acc = 16;
Accuracy(acc);
return;
}
void OsuScore::computeAccValue(const Beatmap& beatmap)
{
// This percentage only considers HitCircles of any value - in this part of the calculation we focus on hitting the timing hit window
f32 betterAccuracyPercentage;
s32 numHitObjectsWithAccuracy;
if (beatmap.ScoreVersion() == Beatmap::EScoreVersion::ScoreV2)
{
numHitObjectsWithAccuracy = TotalHits();
betterAccuracyPercentage = Accuracy();
}
// Either ScoreV1 or some unknown value. Let's default to previous behavior.
else
{
numHitObjectsWithAccuracy = beatmap.NumHitCircles();
if (numHitObjectsWithAccuracy > 0)
betterAccuracyPercentage = static_cast<f32>((_num300 - (TotalHits() - numHitObjectsWithAccuracy)) * 6 + _num100 * 2 + _num50) / (numHitObjectsWithAccuracy * 6);
else
betterAccuracyPercentage = 0;
// It is possible to reach a negative accuracy with this formula. Cap it at zero - zero points
if (betterAccuracyPercentage < 0)
betterAccuracyPercentage = 0;
}
// Lots of arbitrary values from testing.
// Considering to use derivation from perfect accuracy in a probabilistic manner - assume normal distribution
_accValue =
pow(1.52163f, beatmap.DifficultyAttribute(_mods, Beatmap::OD)) * pow(betterAccuracyPercentage, 24) *
2.83f;
// Bonus for many hitcircles - it's harder to keep good accuracy up for longer
_accValue *= std::min(1.15f, static_cast<f32>(pow(numHitObjectsWithAccuracy / 1000.0f, 0.3f)));
if ((_mods & EMods::Hidden) > 0)
_accValue *= 1.08f;
if ((_mods & EMods::Flashlight) > 0)
_accValue *= 1.02f;
}
void CStringEx::PutDouble(double f, int len)
{
switch (_fpclass(f))
{
case _FPCLASS_SNAN:
case _FPCLASS_QNAN:
Format("%s", m_csNAN); // "NaN"
return;
case _FPCLASS_NINF:
Format("%s", m_csNINF);// "-Inf"
return;
case _FPCLASS_PINF:
Format("%s", m_csPINF);// "+Inf"
return;
}
dtoan(GetBuffer(len),f,len);
ReleaseBuffer(len);
// have we exceeded our acuracy?
int acc = 16;
if(len< acc+1) return;
Accuracy(acc);
}
//================================================================================================================================
//function [Accuracy, Criterion] = Mcl_Accuracy(X, Cat, Ncats, CatWeight)
//--------------------------------------------------------------------------------------------------------------------------------
// Finds the accuracy and decision criterion for a 1-criterion univariate classifier. The user supplies data as X and Cat and
// the category for which to compute the univariate criterion as idCat. If there are many columns of X, a separate classifier
// is computed for each column. Each row of X is a data sample associated with a category label supplied by Cat. The samples
// of each categories are permitted to have differential influence on the solution via the CatWeight parameter which is useful
// for enforcing priors.
//--------------------------------------------------------------------------------------------------------------------------------
// NOMENCLATURE
//----------------------------------------------------
// nSamp (int32 vector: nCats)
// The number of samples in each category. For each iCat category, nSamp(1+iCat) is equal to sum(Cat==int32(iCat)).
// ntSamp (int32 scalar)
// The total number of samples (in all categories). Equal to sum(nSamp) or numel(Cat).
// nCols
// The number of columns of X.
//--------------------------------------------------------------------------------------------------------------------------------
// INPUT (values are not altered by mex function)
//----------------------------------------------------
// X (double 2D array: ntSamp * nCols)
// Each row of X gives a the coordinate of a sample of data. Each row is a separate sample, each column is a spatial dimension.
// The univariate classifiers are applied separately to each column of X.
// Cat (int32 vector: ntSamp)
// The 0-based category ID. The elements of this vector must be integers from the set {0,...,Ncats-1}.
// Ncats (int32 scalar)
// The number of categories.
// CatWeight (optional double vector: nCats)
// For category iCat, the weighted accuracy is computed using weights for each category's samples, CatWeight[iCat]; Default = ones(nCats,1);
//--------------------------------------------------------------------------------------------------------------------------------
// OUTPUT (values altered by mex function)
//----------------------------------------------------
// Accuracy (double 2D array: Ncats * nCols)
// The absolute value of Accuracy(iCat,iCol) gives the weighted accuracy for category iCat in X(:,iCol). The sign of accuracy
// disambiguates the decision rule used.
// Accuracy(iCat,iCol) is negative when the decision rule is: Assign iCat if X(:,iCol)<Criterion(iCat,iCol).
// Accuracy(iCat,iCol) is positive when the decision rule is: Assign iCat if X(:,iCol)>Criterion(iCat,iCol).
// Criterion (double 2D array: Ncats * nCols)
//================================================================================================================================
void Mcl_Accuracy(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
{
// Basic error checking of arguments doesn't take very long and helps insure against major disasters.
if( nrhs<3 )
mexErrMsgTxt("Not enough input arguments.");
if( nlhs==0 )
mexErrMsgTxt("Not enough output argument.");
if( nlhs>2 )
mexErrMsgTxt("Too many output arguments.");
//----------------------------------------------------------------------------------------------
// INPUT
//----------------------------------------------------------------------------------------------
double *X = null;
if( !mxIsDouble(prhs[0]) )
mexErrMsgTxt("X must be type double.");
X = (double*)mxGetData(prhs[0]);
int ntSamp = (int)mxGetNumberOfElements(prhs[0]);
int nCols = 1;
if( (int)mxGetNumberOfDimensions(prhs[0])>1 )
{
ntSamp = (int)mxGetM(prhs[0]);
nCols = (int)mxGetN(prhs[0]);
}
//----------------------------------------------------------------------------------------------
int *Cat = null;
if( !mxIsInt32(prhs[1]) )
mexErrMsgTxt("Cat must be type int32.");
if( (int)mxGetNumberOfElements(prhs[1]) < ntSamp )
mexErrMsgTxt("Cat must have ntSamp rows. Usually, ntSamp = size(X,1).");
Cat = (int*)mxGetData(prhs[1]);
//----------------------------------------------------------------------------------------------
int Ncats = 1;
if( !mxIsInt32(prhs[2]) )
mexErrMsgTxt("Ncats must be type int32.");
if( mxGetNumberOfElements(prhs[2])!=1 )
mexErrMsgTxt("Ncats must be a scalar.");
Ncats = ((int*)mxGetData(prhs[2]))[0];
//----------------------------------------------------------------------------------------------
double *CatWeight = null;
if( nrhs>=4 )
{
if( !mxIsDouble(prhs[3]) || (int)mxGetNumberOfElements(prhs[3])<Ncats )
mexErrMsgTxt("CatWeight must be a double vector of length Ncats.");
CatWeight = (double*)mxGetData(prhs[3]);
}
//----------------------------------------------------------------------------------------------
// OUTPUT
//----------------------------------------------------------------------------------------------
double* Acc = null;
plhs[0] = mxCreateDoubleMatrix(Ncats,nCols,mxREAL);
Acc = (double*)mxGetData(plhs[0]);
//----------------------------------------------------------------------------------------------
double* Crit = null;
if( nlhs>=2 )
{
plhs[1] = mxCreateDoubleMatrix(Ncats,nCols,mxREAL);
Crit = (double*)mxGetData(plhs[1]);
}
//----------------------------------------------------------------------------------------------
// STATIC
//----------------------------------------------------------------------------------------------
// Register the function that must delete static memory
if( nMcl_Accuracy_Idx==0 )
mexAtExit(Mcl_Accuracy_DeleteStaticMemory);
// Determine whether static memory must be initialized
if( nMcl_Accuracy_Idx<ntSamp )
{
delete Mcl_Accuracy_Idx;
nMcl_Accuracy_Idx = 2*ntSamp;
Mcl_Accuracy_Idx = new int[nMcl_Accuracy_Idx];
}
//----------------------------------------------------------------------------------------------
//mexPrintf("Mcl_Accuracy: [ntSamp, nCols, Ncats] = [%i, %i, %i]\n", ntSamp, nCols, Ncats);
int idxCrit = 0;
int iCol, iSamp, iCat, iOut;
double *Xcol;
// For each column of data
for(iCol=0; iCol<nCols; iCol++)
{
// Initialize the index
for(iSamp=0; iSamp<ntSamp; iSamp++)
Mcl_Accuracy_Idx[iSamp] = iSamp;
//mexPrintf("Mcl_Accuracy 02: [iCol] = [%i]\n", iCol);
// Sort
Xcol = X+iCol*ntSamp;
QuickSort(Mcl_Accuracy_Idx, Xcol, 0, ntSamp-1);
//mexPrintf("Mcl_Accuracy 03: [iCol] = [%i]\n", iCol);
// For each category
for(iCat=0; iCat<Ncats; iCat++)
{
// Accuracy
iOut = iCat+Ncats*iCol;
//mexPrintf("-------------------------------------------------------------------------\n");
//mexPrintf("Mcl_Accuracy 04: [iCol, iCat, iOut] = [%i, %i, %i]\n", iCol, iCat, iOut);
//mexPrintf("-----------------\n");
Acc[iOut] = Accuracy(&idxCrit, Mcl_Accuracy_Idx, Xcol, Cat, CatWeight, ntSamp, Ncats, iCat);
if( idxCrit<0 || idxCrit > ntSamp )
Acc[iOut] = -Acc[iOut];
//mexPrintf("Mcl_Accuracy 05: [iCol, iCat, iOut, Acc[iOut], idxCrit] = [%i, %i, %i, %f, %i]\n", iCol, iCat, iOut, Acc[iOut], idxCrit);
// Get the crit
if(Crit!=null)
{
if(idxCrit<-ntSamp || idxCrit>ntSamp)
Crit[iOut] = Xcol[Mcl_Accuracy_Idx[ntSamp-1]]+1.0;
else if(idxCrit==0 )
Crit[iOut] = Xcol[Mcl_Accuracy_Idx[0]]-1.0;
else if( idxCrit>0 )
Crit[iOut] = 0.5 * (Xcol[Mcl_Accuracy_Idx[ idxCrit]] + Xcol[Mcl_Accuracy_Idx[ idxCrit-1]]);
else
Crit[iOut] = 0.5 * (Xcol[Mcl_Accuracy_Idx[-idxCrit]] + Xcol[Mcl_Accuracy_Idx[-idxCrit-1]]);
}
//mexPrintf("Mcl_Accuracy 06: [iCol, iCat, iOut, Acc[iOut], Crit[iOut]] = [%i, %i, %i, %f, %f]\n", iCol, iCat, iOut, Acc[iOut], Crit[iOut]);
}
}
}
PP_NAMESPACE_BEGIN
CatchScore::CatchScore(
s64 scoreId,
EGamemode mode,
s64 userId,
s32 beatmapId,
s32 score,
s32 maxCombo,
s32 num300,
s32 num100,
s32 num50,
s32 numMiss,
s32 numGeki,
s32 numKatu,
EMods mods,
const Beatmap& beatmap
) : Score{scoreId, mode, userId, beatmapId, score, maxCombo, num300, num100, num50, numMiss, numGeki, numKatu, mods}
{
// Don't count scores made with supposedly unranked mods
if ((_mods & EMods::Relax) > 0 ||
(_mods & EMods::Relax2) > 0 ||
(_mods & EMods::Autoplay) > 0)
{
_value = 0;
return;
}
// We are heavily relying on aim in catch the beat
_value = pow(5.0f * std::max(1.0f, beatmap.DifficultyAttribute(_mods, Beatmap::Aim) / 0.0049f) - 4.0f, 2.0f) / 100000.0f;
// Longer maps are worth more. "Longer" means how many hits there are which can contribute to combo
int numTotalHits = totalComboHits();
// Longer maps are worth more
f32 lengthBonus =
0.95f + 0.4f * std::min<f32>(1.0f, static_cast<f32>(numTotalHits) / 3000.0f) +
(numTotalHits > 3000 ? log10(static_cast<f32>(numTotalHits) / 3000.0f) * 0.5f : 0.0f);
// Longer maps are worth more
_value *= lengthBonus;
// Penalize misses exponentially. This mainly fixes tag4 maps and the likes until a per-hitobject solution is available
_value *= pow(0.97f, _numMiss);
// Combo scaling
float beatmapMaxCombo = beatmap.DifficultyAttribute(_mods, Beatmap::MaxCombo);
if (beatmapMaxCombo > 0)
_value *= std::min<f32>(pow(static_cast<f32>(_maxCombo), 0.8f) / pow(beatmapMaxCombo, 0.8f), 1.0f);
f32 approachRate = beatmap.DifficultyAttribute(_mods, Beatmap::AR);
f32 approachRateFactor = 1.0f;
if (approachRate > 9.0f)
approachRateFactor += 0.1f * (approachRate - 9.0f); // 10% for each AR above 9
else if (approachRate < 8.0f)
approachRateFactor += 0.025f * (8.0f - approachRate); // 2.5% for each AR below 8
_value *= approachRateFactor;
if ((_mods & EMods::Hidden) > 0)
// Hiddens gives nothing on max approach rate, and more the lower it is
_value *= 1.05f + 0.075f * (10.0f - std::min(10.0f, approachRate)); // 7.5% for each AR below 10
if ((_mods & EMods::Flashlight) > 0)
// Apply length bonus again if flashlight is on simply because it becomes a lot harder on longer maps.
_value *= 1.35f * lengthBonus;
// Scale the aim value with accuracy _slightly_
_value *= pow(Accuracy(), 5.5f);
// Custom multipliers for NoFail and SpunOut.
if ((_mods & EMods::NoFail) > 0)
_value *= 0.90f;
if ((_mods & EMods::SpunOut) > 0)
_value *= 0.95f;
}
