Examples with ScoreObject ml.shifu.shifu.container.ScoreObject used on opensource projects

Example 1 with ScoreObject

use of ml.shifu.shifu.container.ScoreObject in project shifu by ShifuML.

the class Scorer method scoreNsData.

public ScoreObject scoreNsData(MLDataPair inputPair, Map<NSColumn, String> rawNsDataMap) {
    if (inputPair == null && !this.alg.equalsIgnoreCase(NNConstants.NN_ALG_NAME)) {
        inputPair = NormalUtils.assembleNsDataPair(binCategoryMap, noVarSelect, modelConfig, selectedColumnConfigList, rawNsDataMap, cutoff, alg);
    }
    // clear cache
    this.cachedNormDataPair.clear();
    final MLDataPair pair = inputPair;
    List<MLData> modelResults = new ArrayList<MLData>();
    List<Callable<MLData>> tasks = null;
    if (this.multiThread) {
        tasks = new ArrayList<Callable<MLData>>();
    }
    for (final BasicML model : models) {
        // TODO, check if no need 'if' condition and refactor two if for loops please
        if (model instanceof BasicFloatNetwork || model instanceof NNModel) {
            final BasicFloatNetwork network = (model instanceof BasicFloatNetwork) ? (BasicFloatNetwork) model : ((NNModel) model).getIndependentNNModel().getBasicNetworks().get(0);
            String cacheKey = featureSetToString(network.getFeatureSet());
            MLDataPair dataPair = cachedNormDataPair.get(cacheKey);
            if (dataPair == null) {
                dataPair = NormalUtils.assembleNsDataPair(binCategoryMap, noVarSelect, modelConfig, selectedColumnConfigList, rawNsDataMap, cutoff, alg, network.getFeatureSet());
                cachedNormDataPair.put(cacheKey, dataPair);
            }
            final MLDataPair networkPair = dataPair;
            /*
                 * if(network.getFeatureSet().size() != networkPair.getInput().size()) {
                 * log.error("Network and input size mismatch: Network Size = " + network.getFeatureSet().size()
                 * + "; Input Size = " + networkPair.getInput().size());
                 * continue;
                 * }
                 */
            if (System.currentTimeMillis() % 1000 == 0L) {
                log.info("Network input count = {}, while input size = {}", network.getInputCount(), networkPair.getInput().size());
            }
            final int fnlOutputHiddenLayerIndex = outputHiddenLayerIndex;
            Callable<MLData> callable = new Callable<MLData>() {

                @Override
                public MLData call() {
                    MLData finalOutput = network.compute(networkPair.getInput());
                    if (fnlOutputHiddenLayerIndex == 0) {
                        return finalOutput;
                    }
                    // append output values in hidden layer
                    double[] hiddenOutputs = network.getLayerOutput(fnlOutputHiddenLayerIndex);
                    double[] outputs = new double[finalOutput.getData().length + hiddenOutputs.length];
                    System.arraycopy(finalOutput.getData(), 0, outputs, 0, finalOutput.getData().length);
                    System.arraycopy(hiddenOutputs, 0, outputs, finalOutput.getData().length, hiddenOutputs.length);
                    return new BasicMLData(outputs);
                }
            };
            if (multiThread) {
                tasks.add(callable);
            } else {
                try {
                    modelResults.add(callable.call());
                } catch (Exception e) {
                    log.error("error in model evaluation", e);
                }
            }
        } else if (model instanceof BasicNetwork) {
            final BasicNetwork network = (BasicNetwork) model;
            final MLDataPair networkPair = NormalUtils.assembleNsDataPair(binCategoryMap, noVarSelect, modelConfig, columnConfigList, rawNsDataMap, cutoff, alg, null);
            Callable<MLData> callable = new Callable<MLData>() {

                @Override
                public MLData call() {
                    return network.compute(networkPair.getInput());
                }
            };
            if (multiThread) {
                tasks.add(callable);
            } else {
                try {
                    modelResults.add(callable.call());
                } catch (Exception e) {
                    log.error("error in model evaluation", e);
                }
            }
        } else if (model instanceof SVM) {
            final SVM svm = (SVM) model;
            if (svm.getInputCount() != pair.getInput().size()) {
                log.error("SVM and input size mismatch: SVM Size = " + svm.getInputCount() + "; Input Size = " + pair.getInput().size());
                continue;
            }
            Callable<MLData> callable = new Callable<MLData>() {

                @Override
                public MLData call() {
                    return svm.compute(pair.getInput());
                }
            };
            if (multiThread) {
                tasks.add(callable);
            } else {
                try {
                    modelResults.add(callable.call());
                } catch (Exception e) {
                    log.error("error in model evaluation", e);
                }
            }
        } else if (model instanceof LR) {
            final LR lr = (LR) model;
            if (lr.getInputCount() != pair.getInput().size()) {
                log.error("LR and input size mismatch: LR Size = " + lr.getInputCount() + "; Input Size = " + pair.getInput().size());
                continue;
            }
            Callable<MLData> callable = new Callable<MLData>() {

                @Override
                public MLData call() {
                    return lr.compute(pair.getInput());
                }
            };
            if (multiThread) {
                tasks.add(callable);
            } else {
                try {
                    modelResults.add(callable.call());
                } catch (Exception e) {
                    log.error("error in model evaluation", e);
                }
            }
        } else if (model instanceof TreeModel) {
            final TreeModel tm = (TreeModel) model;
            if (tm.getInputCount() != pair.getInput().size()) {
                throw new RuntimeException("GBDT and input size mismatch: tm input Size = " + tm.getInputCount() + "; data input Size = " + pair.getInput().size());
            }
            Callable<MLData> callable = new Callable<MLData>() {

                @Override
                public MLData call() {
                    MLData result = tm.compute(pair.getInput());
                    return result;
                }
            };
            if (multiThread) {
                tasks.add(callable);
            } else {
                try {
                    modelResults.add(callable.call());
                } catch (Exception e) {
                    log.error("error in model evaluation", e);
                }
            }
        } else if (model instanceof GenericModel) {
            Callable<MLData> callable = new Callable<MLData>() {

                @Override
                public MLData call() {
                    return ((GenericModel) model).compute(pair.getInput());
                }
            };
            if (multiThread) {
                tasks.add(callable);
            } else {
                try {
                    modelResults.add(callable.call());
                } catch (Exception e) {
                    log.error("error in model evaluation", e);
                }
            }
        } else {
            throw new RuntimeException("unsupport models");
        }
    }
    List<Double> scores = new ArrayList<Double>();
    List<Integer> rfTreeSizeList = new ArrayList<Integer>();
    SortedMap<String, Double> hiddenOutputs = null;
    if (CollectionUtils.isNotEmpty(modelResults) || CollectionUtils.isNotEmpty(tasks)) {
        int modelSize = modelResults.size() > 0 ? modelResults.size() : tasks.size();
        if (modelSize != this.models.size()) {
            log.error("Get model results size doesn't match with models size.");
            return null;
        }
        if (multiThread) {
            modelResults = this.executorManager.submitTasksAndWaitResults(tasks);
        } else {
        // not multi-thread, modelResults is directly being populated in callable.call
        }
        if (this.outputHiddenLayerIndex != 0) {
            hiddenOutputs = new TreeMap<String, Double>(new Comparator<String>() {

                @Override
                public int compare(String o1, String o2) {
                    String[] split1 = o1.split("_");
                    String[] split2 = o2.split("_");
                    int model1Index = Integer.parseInt(split1[1]);
                    int model2Index = Integer.parseInt(split2[1]);
                    if (model1Index > model2Index) {
                        return 1;
                    } else if (model1Index < model2Index) {
                        return -1;
                    } else {
                        int hidden1Index = Integer.parseInt(split1[2]);
                        int hidden2Index = Integer.parseInt(split2[2]);
                        if (hidden1Index > hidden2Index) {
                            return 1;
                        } else if (hidden1Index < hidden2Index) {
                            return -1;
                        } else {
                            int hidden11Index = Integer.parseInt(split1[3]);
                            int hidden22Index = Integer.parseInt(split2[3]);
                            return Integer.valueOf(hidden11Index).compareTo(Integer.valueOf(hidden22Index));
                        }
                    }
                }
            });
        }
        for (int i = 0; i < this.models.size(); i++) {
            BasicML model = this.models.get(i);
            MLData score = modelResults.get(i);
            if (model instanceof BasicNetwork || model instanceof NNModel) {
                if (modelConfig != null && modelConfig.isRegression()) {
                    scores.add(toScore(score.getData(0)));
                    if (this.outputHiddenLayerIndex != 0) {
                        for (int j = 1; j < score.getData().length; j++) {
                            hiddenOutputs.put("model_" + i + "_" + this.outputHiddenLayerIndex + "_" + (j - 1), score.getData()[j]);
                        }
                    }
                } else if (modelConfig != null && modelConfig.isClassification() && modelConfig.getTrain().isOneVsAll()) {
                    // if one vs all classification
                    scores.add(toScore(score.getData(0)));
                } else {
                    double[] outputs = score.getData();
                    for (double d : outputs) {
                        scores.add(toScore(d));
                    }
                }
            } else if (model instanceof SVM) {
                scores.add(toScore(score.getData(0)));
            } else if (model instanceof LR) {
                scores.add(toScore(score.getData(0)));
            } else if (model instanceof TreeModel) {
                if (modelConfig.isClassification() && !modelConfig.getTrain().isOneVsAll()) {
                    double[] scoreArray = score.getData();
                    for (double sc : scoreArray) {
                        scores.add(sc);
                    }
                } else {
                    // if one vs all multiple classification or regression
                    scores.add(toScore(score.getData(0)));
                }
                final TreeModel tm = (TreeModel) model;
                // regression for RF
                if (!tm.isClassfication() && !tm.isGBDT()) {
                    rfTreeSizeList.add(tm.getTrees().size());
                }
            } else if (model instanceof GenericModel) {
                scores.add(toScore(score.getData(0)));
            } else {
                throw new RuntimeException("unsupport models");
            }
        }
    }
    Integer tag = Constants.DEFAULT_IDEAL_VALUE;
    if (scores.size() == 0 && System.currentTimeMillis() % 100 == 0) {
        log.warn("No Scores Calculated...");
    }
    return new ScoreObject(scores, tag, rfTreeSizeList, hiddenOutputs);
}

Example 2 with ScoreObject

use of ml.shifu.shifu.container.ScoreObject in project shifu by ShifuML.

the class ScorerTest method scoreTest.

// @Test
public void scoreTest() {
    List<ColumnConfig> list = new ArrayList<ColumnConfig>();
    ColumnConfig col = new ColumnConfig();
    col.setColumnType(ColumnType.N);
    col.setColumnName("A");
    col.setColumnNum(0);
    col.setFinalSelect(true);
    list.add(col);
    col = new ColumnConfig();
    col.setColumnType(ColumnType.N);
    col.setColumnName("B");
    col.setColumnNum(1);
    col.setFinalSelect(true);
    list.add(col);
    Scorer s = new Scorer(models, list, "NN", modelConfig);
    double[] input = { 0., 0. };
    double[] ideal = { 1. };
    MLDataPair pair = new BasicMLDataPair(new BasicMLData(input), new BasicMLData(ideal));
    ScoreObject o = s.score(pair, null);
    List<Double> scores = o.getScores();
    Assert.assertTrue(scores.get(0) > 400);
    Assert.assertTrue(scores.get(1) == 1000);
}

Example 3 with ScoreObject

use of ml.shifu.shifu.container.ScoreObject in project shifu by ShifuML.

the class ModelRunner method computeNsData.

/**
 * Run model to compute score for input NS Data map
 *
 * @param rawDataNsMap
 *            - the original input, but key is wrapped by NSColumn
 * @return CaseScoreResult - model score
 */
public CaseScoreResult computeNsData(final Map<NSColumn, String> rawDataNsMap) {
    if (MapUtils.isEmpty(rawDataNsMap)) {
        return null;
    }
    CaseScoreResult scoreResult = new CaseScoreResult();
    if (this.scorer != null) {
        ScoreObject so = scorer.scoreNsData(rawDataNsMap);
        if (so == null) {
            return null;
        }
        scoreResult.setScores(so.getScores());
        scoreResult.setMaxScore(so.getMaxScore());
        scoreResult.setMinScore(so.getMinScore());
        scoreResult.setAvgScore(so.getMeanScore());
        scoreResult.setMedianScore(so.getMedianScore());
        scoreResult.setHiddenLayerScores(so.getHiddenLayerScores());
    }
    if (MapUtils.isNotEmpty(this.subScorers)) {
        if (this.isMultiThread && this.subScorers.size() > 1 && this.executorManager == null) {
            int threadPoolSize = Math.min(Runtime.getRuntime().availableProcessors(), this.subScorers.size());
            this.executorManager = new ExecutorManager<Pair<String, ScoreObject>>(threadPoolSize);
            // add a shutdown hook as a safe guard if some one not call close
            Runtime.getRuntime().addShutdownHook(new Thread(new Runnable() {

                @Override
                public void run() {
                    ModelRunner.this.executorManager.forceShutDown();
                }
            }));
            log.info("MultiThread is enabled in ModelRunner, threadPoolSize = " + threadPoolSize);
        }
        List<Callable<Pair<String, ScoreObject>>> tasks = new ArrayList<Callable<Pair<String, ScoreObject>>>(this.subScorers.size());
        Iterator<Map.Entry<String, Scorer>> iterator = this.subScorers.entrySet().iterator();
        while (iterator.hasNext()) {
            final Map.Entry<String, Scorer> entry = iterator.next();
            Callable<Pair<String, ScoreObject>> callable = new Callable<Pair<String, ScoreObject>>() {

                @Override
                public Pair<String, ScoreObject> call() {
                    String modelName = entry.getKey();
                    Scorer subScorer = entry.getValue();
                    ScoreObject so = subScorer.scoreNsData(rawDataNsMap);
                    if (so != null) {
                        return Pair.of(modelName, so);
                    } else {
                        return null;
                    }
                }
            };
            tasks.add(callable);
        }
        if (this.isMultiThread && this.subScorers.size() > 1) {
            List<Pair<String, ScoreObject>> results = this.executorManager.submitTasksAndWaitResults(tasks);
            for (Pair<String, ScoreObject> result : results) {
                if (result != null) {
                    scoreResult.addSubModelScore(result.getLeft(), result.getRight());
                }
            }
        } else {
            for (Callable<Pair<String, ScoreObject>> task : tasks) {
                Pair<String, ScoreObject> result = null;
                try {
                    result = task.call();
                } catch (Exception e) {
                // do nothing
                }
                if (result != null) {
                    scoreResult.addSubModelScore(result.getLeft(), result.getRight());
                }
            }
        }
    }
    return scoreResult;
}