diff options
Diffstat (limited to 'ml/Dimension.cc')
-rw-r--r-- | ml/Dimension.cc | 76 |
1 files changed, 57 insertions, 19 deletions
diff --git a/ml/Dimension.cc b/ml/Dimension.cc index 0fe07530c..bf34abb72 100644 --- a/ml/Dimension.cc +++ b/ml/Dimension.cc @@ -6,8 +6,13 @@ using namespace ml; -std::pair<CalculatedNumber *, size_t> -TrainableDimension::getCalculatedNumbers() { +bool Dimension::isActive() const { + bool SetObsolete = rrdset_flag_check(RD->rrdset, RRDSET_FLAG_OBSOLETE); + bool DimObsolete = rrddim_flag_check(RD, RRDDIM_FLAG_OBSOLETE); + return !SetObsolete && !DimObsolete; +} + +std::pair<CalculatedNumber *, size_t> Dimension::getCalculatedNumbers() { size_t MinN = Cfg.MinTrainSamples; size_t MaxN = Cfg.MaxTrainSamples; @@ -68,7 +73,7 @@ TrainableDimension::getCalculatedNumbers() { return { CNs, TotalValues }; } -MLResult TrainableDimension::trainModel() { +MLResult Dimension::trainModel() { auto P = getCalculatedNumbers(); CalculatedNumber *CNs = P.first; unsigned N = P.second; @@ -81,7 +86,15 @@ MLResult TrainableDimension::trainModel() { SamplesBuffer SB = SamplesBuffer(CNs, N, 1, Cfg.DiffN, Cfg.SmoothN, Cfg.LagN, SamplingRatio, Cfg.RandomNums); - KM.train(SB, Cfg.MaxKMeansIters); + std::vector<DSample> Samples = SB.preprocess(); + + KMeans KM; + KM.train(Samples, Cfg.MaxKMeansIters); + + { + std::lock_guard<std::mutex> Lock(Mutex); + Models[0] = KM; + } Trained = true; ConstantModel = true; @@ -90,16 +103,25 @@ MLResult TrainableDimension::trainModel() { return MLResult::Success; } -void PredictableDimension::addValue(CalculatedNumber Value, bool Exists) { +bool Dimension::shouldTrain(const TimePoint &TP) const { + if (ConstantModel) + return false; + + return (LastTrainedAt + Seconds(Cfg.TrainEvery * updateEvery())) < TP; +} + +bool Dimension::predict(CalculatedNumber Value, bool Exists) { if (!Exists) { CNs.clear(); - return; + AnomalyBit = false; + return false; } unsigned N = Cfg.DiffN + Cfg.SmoothN + Cfg.LagN; if (CNs.size() < N) { CNs.push_back(Value); - return; + AnomalyBit = false; + return false; } std::rotate(std::begin(CNs), std::begin(CNs) + 1, std::end(CNs)); @@ -108,28 +130,44 @@ void PredictableDimension::addValue(CalculatedNumber Value, bool Exists) { ConstantModel = false; CNs[N - 1] = Value; -} -std::pair<MLResult, bool> PredictableDimension::predict() { - unsigned N = Cfg.DiffN + Cfg.SmoothN + Cfg.LagN; - if (CNs.size() != N) { + if (!isTrained() || ConstantModel) { AnomalyBit = false; - return { MLResult::MissingData, AnomalyBit }; + return false; } CalculatedNumber *TmpCNs = new CalculatedNumber[N * (Cfg.LagN + 1)](); std::memcpy(TmpCNs, CNs.data(), N * sizeof(CalculatedNumber)); - - SamplesBuffer SB = SamplesBuffer(TmpCNs, N, 1, Cfg.DiffN, Cfg.SmoothN, Cfg.LagN, + SamplesBuffer SB = SamplesBuffer(TmpCNs, N, 1, + Cfg.DiffN, Cfg.SmoothN, Cfg.LagN, 1.0, Cfg.RandomNums); - AnomalyScore = computeAnomalyScore(SB); + const DSample Sample = SB.preprocess().back(); delete[] TmpCNs; - if (AnomalyScore == std::numeric_limits<CalculatedNumber>::quiet_NaN()) { + std::unique_lock<std::mutex> Lock(Mutex, std::defer_lock); + if (!Lock.try_lock()) { AnomalyBit = false; - return { MLResult::NaN, AnomalyBit }; + return false; } - AnomalyBit = AnomalyScore >= (100 * Cfg.DimensionAnomalyScoreThreshold); - return { MLResult::Success, AnomalyBit }; + for (const auto &KM : Models) { + double AnomalyScore = KM.anomalyScore(Sample); + if (AnomalyScore == std::numeric_limits<CalculatedNumber>::quiet_NaN()) { + AnomalyBit = false; + continue; + } + + if (AnomalyScore < (100 * Cfg.DimensionAnomalyScoreThreshold)) { + AnomalyBit = false; + return false; + } + } + + AnomalyBit = true; + return true; +} + +std::array<KMeans, 1> Dimension::getModels() { + std::unique_lock<std::mutex> Lock(Mutex); + return Models; } |