static void AssignRandomWeights(int learnSampleCount,
TLearnContext* ctx,
TFold* fold) {
TVector<float> sampleWeights;
sampleWeights.yresize(learnSampleCount);
const ui64 randSeed = ctx->Rand.GenRand();
NPar::TLocalExecutor::TExecRangeParams blockParams(0, learnSampleCount);
blockParams.SetBlockSize(10000);
ctx->LocalExecutor.ExecRange([&](int blockIdx) {
TFastRng64 rand(randSeed + blockIdx);
rand.Advance(10); // reduce correlation between RNGs in different threads
const float baggingTemperature = ctx->Params.ObliviousTreeOptions->BootstrapConfig->GetBaggingTemperature();
float* sampleWeightsData = sampleWeights.data();
NPar::TLocalExecutor::BlockedLoopBody(blockParams, [&rand, sampleWeightsData, baggingTemperature](int i) {
const float w = -FastLogf(rand.GenRandReal1() + 1e-100);
sampleWeightsData[i] = powf(w, baggingTemperature);
})(blockIdx);
}, 0, blockParams.GetBlockCount(), NPar::TLocalExecutor::WAIT_COMPLETE);
TFold& ff = *fold;
ff.AssignPermuted(sampleWeights, &ff.SampleWeights);
if (!ff.LearnWeights.empty()) {
for (int i = 0; i < learnSampleCount; ++i) {
ff.SampleWeights[i] *= ff.LearnWeights[i];
}
}
const int approxDimension = ff.GetApproxDimension();
for (TFold::TBodyTail& bt : ff.BodyTailArr) {
for (int dim = 0; dim < approxDimension; ++dim) {
double* weightedDerData = bt.WeightedDer[dim].data();
const double* derData = bt.Derivatives[dim].data();
const float* sampleWeightsData = ff.SampleWeights.data();
ctx->LocalExecutor.ExecRange([=](int z) {
weightedDerData[z] = derData[z] * sampleWeightsData[z];
}, NPar::TLocalExecutor::TExecRangeParams(bt.BodyFinish, bt.TailFinish).SetBlockSize(4000)
, NPar::TLocalExecutor::WAIT_COMPLETE);
}
}
}
static inline void BinarizeFloatFeature(int featureIdx,
const TDocumentStorage& docStorage,
const TDocSelector& docSelector,
const TVector<float>& borders,
ENanMode nanMode,
NPar::TLocalExecutor& localExecutor,
int floatFeatureIdx,
TAllFeatures* features,
bool* seenNans) {
size_t docCount = docSelector.GetDocCount();
const TVector<float>& src = docStorage.Factors[featureIdx];
TVector<ui8>& hist = features->FloatHistograms[floatFeatureIdx];
hist.resize(docCount);
ui8* histData = hist.data();
const float* featureBorderData = borders.data();
const int featureBorderSize = borders.ysize();
localExecutor.ExecRange([&] (int i) {
const auto& featureVal = src[docSelector(i)];
if (IsNan(featureVal)) {
*seenNans = true;
histData[i] = nanMode == ENanMode::Min ? 0 : featureBorderSize;
} else {
int j = 0;
while (j < featureBorderSize && featureVal > featureBorderData[j]) {
++histData[i];
++j;
}
// histData[i] = LowerBound(featureBorderData, featureBorderData + featureBorderSize, featureVal) - featureBorderData;
}
}
, NPar::TLocalExecutor::TExecRangeParams(0, docCount).SetBlockSize(1000)
, NPar::TLocalExecutor::WAIT_COMPLETE);
}
void dump_as_array(const int fd) final {
osmium::io::detail::reliable_write(fd, reinterpret_cast<const char*>(m_vector.data()), byte_size());
}
