private int[] applyNewFeatures(List<RankList> samples, int[] features) {
int totalFeatureCount = samples.get(0).get(0).getFeatureCount();
int[] newFeatures = new int[features.length + lcList.size()];
System.arraycopy(features, 0, newFeatures, 0, features.length);
// for(int i=0;i<features.length;i++)
// newFeatures[i] = features[i];
for (int k = 0; k < lcList.size(); k++)
newFeatures[features.length + k] = totalFeatureCount + k + 1;
float[] addedFeatures = new float[lcList.size()];
for (int i = 0; i < samples.size(); i++) {
RankList rl = samples.get(i);
for (int j = 0; j < rl.size(); j++) {
DataPoint p = rl.get(j);
for (int k = 0; k < lcList.size(); k++)
addedFeatures[k] = lcList.get(k).compute(p.getExternalFeatureVector());
p.addFeatures(addedFeatures);
}
}
int[] newFeatures2 = new int[lcList.size()];
for (int i = 0; i < lcList.size(); i++) newFeatures2[i] = newFeatures[i + features.length];
if (keepOrigFeatures) return newFeatures;
return newFeatures2;
}
/** @param args */
public static void main(String[] args) {
String[] rType =
new String[] {
"MART",
"RankNet",
"RankBoost",
"AdaRank",
"Coordinate Ascent",
"LambdaRank",
"LambdaMART",
"ListNet",
"Random Forests"
};
RANKER_TYPE[] rType2 =
new RANKER_TYPE[] {
RANKER_TYPE.MART,
RANKER_TYPE.RANKNET,
RANKER_TYPE.RANKBOOST,
RANKER_TYPE.ADARANK,
RANKER_TYPE.COOR_ASCENT,
RANKER_TYPE.LAMBDARANK,
RANKER_TYPE.LAMBDAMART,
RANKER_TYPE.LISTNET,
RANKER_TYPE.RANDOM_FOREST
};
String trainFile = "";
String featureDescriptionFile = "";
double ttSplit = 0.0; // train-test split
double tvSplit = 0.0; // train-validation split
int foldCV = -1;
String validationFile = "";
String testFile = "";
int rankerType = 4;
String trainMetric = "ERR@10";
String testMetric = "";
Evaluator.normalize = false;
String savedModelFile = "";
String rankFile = "";
boolean printIndividual = false;
// for my personal use
String indriRankingFile = "";
String scoreFile = "";
if (args.length < 2) {
System.out.println("Usage: java -jar RankLib.jar <Params>");
System.out.println("Params:");
System.out.println(" [+] Training (+ tuning and evaluation)");
System.out.println("\t-train <file>\t\tTraining data");
System.out.println("\t-ranker <type>\t\tSpecify which ranking algorithm to use");
System.out.println("\t\t\t\t0: MART (gradient boosted regression tree)");
System.out.println("\t\t\t\t1: RankNet");
System.out.println("\t\t\t\t2: RankBoost");
System.out.println("\t\t\t\t3: AdaRank");
System.out.println("\t\t\t\t4: Coordinate Ascent");
System.out.println("\t\t\t\t6: LambdaMART");
System.out.println("\t\t\t\t7: ListNet");
System.out.println("\t\t\t\t8: Random Forests");
System.out.println(
"\t[ -feature <file> ]\tFeature description file: list features to be considered by the learner, each on a separate line");
System.out.println("\t\t\t\tIf not specified, all features will be used.");
// System.out.println("\t[ -metric2t <metric> ]\tMetric to optimize on the training data.
// Supported: MAP, NDCG@k, DCG@k, P@k, RR@k, BEST@k, ERR@k (default=" + trainMetric + ")");
System.out.println(
"\t[ -metric2t <metric> ]\tMetric to optimize on the training data. Supported: MAP, NDCG@k, DCG@k, P@k, RR@k, ERR@k (default="
+ trainMetric
+ ")");
System.out.println(
"\t[ -metric2T <metric> ]\tMetric to evaluate on the test data (default to the same as specified for -metric2t)");
System.out.println(
"\t[ -gmax <label> ]\tHighest judged relevance label. It affects the calculation of ERR (default="
+ (int) SimpleMath.logBase2(ERRScorer.MAX)
+ ", i.e. 5-point scale {0,1,2,3,4})");
// System.out.println("\t[ -qrel <file> ]\tTREC-style relevance judgment file. It only affects
// MAP and NDCG (default=unspecified)");
System.out.println(
"\t[ -test <file> ]\tSpecify if you want to evaluate the trained model on this data (default=unspecified)");
System.out.println(
"\t[ -validate <file> ]\tSpecify if you want to tune your system on the validation data (default=unspecified)");
System.out.println(
"\t\t\t\tIf specified, the final model will be the one that performs best on the validation data");
System.out.println("\t[ -tvs <x \\in [0..1]> ]\tSet train-validation split to be (x)(1.0-x)");
System.out.println(
"\t[ -tts <x \\in [0..1]> ]\tSet train-test split to be (x)(1.0-x). -tts will override -tvs");
System.out.println(
"\t[ -kcv <k> ]\t\tSpecify if you want to perform k-fold cross validation using ONLY the specified training data (default=NoCV)");
System.out.println(
"\t[ -norm <method>]\tNormalize feature vectors (default=no-normalization). Method can be:");
System.out.println("\t\t\t\tsum: normalize each feature by the sum of all its values");
System.out.println("\t\t\t\tzscore: normalize each feature by its mean/standard deviation");
System.out.println(
"\t[ -save <model> ]\tSave the learned model to the specified file (default=not-save)");
System.out.println(
"\t[ -silent ]\t\tDo not print progress messages (which are printed by default)");
System.out.println("");
System.out.println(" [-] RankNet-specific parameters");
System.out.println(
"\t[ -epoch <T> ]\t\tThe number of epochs to train (default=" + RankNet.nIteration + ")");
System.out.println(
"\t[ -layer <layer> ]\tThe number of hidden layers (default="
+ RankNet.nHiddenLayer
+ ")");
System.out.println(
"\t[ -node <node> ]\tThe number of hidden nodes per layer (default="
+ RankNet.nHiddenNodePerLayer
+ ")");
System.out.println(
"\t[ -lr <rate> ]\t\tLearning rate (default="
+ (new DecimalFormat("###.########")).format(RankNet.learningRate)
+ ")");
System.out.println("");
System.out.println(" [-] RankBoost-specific parameters");
System.out.println(
"\t[ -round <T> ]\t\tThe number of rounds to train (default="
+ RankBoost.nIteration
+ ")");
System.out.println(
"\t[ -tc <k> ]\t\tNumber of threshold candidates to search. -1 to use all feature values (default="
+ RankBoost.nThreshold
+ ")");
System.out.println("");
System.out.println(" [-] AdaRank-specific parameters");
System.out.println(
"\t[ -round <T> ]\t\tThe number of rounds to train (default=" + AdaRank.nIteration + ")");
System.out.println(
"\t[ -noeq ]\t\tTrain without enqueuing too-strong features (default=unspecified)");
System.out.println(
"\t[ -tolerance <t> ]\tTolerance between two consecutive rounds of learning (default="
+ AdaRank.tolerance
+ ")");
System.out.println(
"\t[ -max <times> ]\tThe maximum number of times can a feature be consecutively selected without changing performance (default="
+ AdaRank.maxSelCount
+ ")");
System.out.println("");
System.out.println(" [-] Coordinate Ascent-specific parameters");
System.out.println(
"\t[ -r <k> ]\t\tThe number of random restarts (default=" + CoorAscent.nRestart + ")");
System.out.println(
"\t[ -i <iteration> ]\tThe number of iterations to search in each dimension (default="
+ CoorAscent.nMaxIteration
+ ")");
System.out.println(
"\t[ -tolerance <t> ]\tPerformance tolerance between two solutions (default="
+ CoorAscent.tolerance
+ ")");
System.out.println(
"\t[ -reg <slack> ]\tRegularization parameter (default=no-regularization)");
System.out.println("");
System.out.println(" [-] {MART, LambdaMART}-specific parameters");
System.out.println("\t[ -tree <t> ]\t\tNumber of trees (default=" + LambdaMART.nTrees + ")");
System.out.println(
"\t[ -leaf <l> ]\t\tNumber of leaves for each tree (default="
+ LambdaMART.nTreeLeaves
+ ")");
System.out.println(
"\t[ -shrinkage <factor> ]\tShrinkage, or learning rate (default="
+ LambdaMART.learningRate
+ ")");
System.out.println(
"\t[ -tc <k> ]\t\tNumber of threshold candidates for tree spliting. -1 to use all feature values (default="
+ LambdaMART.nThreshold
+ ")");
System.out.println(
"\t[ -mls <n> ]\t\tMin leaf support -- minimum #samples each leaf has to contain (default="
+ LambdaMART.minLeafSupport
+ ")");
System.out.println(
"\t[ -estop <e> ]\t\tStop early when no improvement is observed on validaton data in e consecutive rounds (default="
+ LambdaMART.nRoundToStopEarly
+ ")");
System.out.println("");
System.out.println(" [-] ListNet-specific parameters");
System.out.println(
"\t[ -epoch <T> ]\t\tThe number of epochs to train (default=" + ListNet.nIteration + ")");
System.out.println(
"\t[ -lr <rate> ]\t\tLearning rate (default="
+ (new DecimalFormat("###.########")).format(ListNet.learningRate)
+ ")");
System.out.println("");
System.out.println(" [-] Random Forests-specific parameters");
System.out.println("\t[ -bag <r> ]\t\tNumber of bags (default=" + RFRanker.nBag + ")");
System.out.println(
"\t[ -srate <r> ]\t\tSub-sampling rate (default=" + RFRanker.subSamplingRate + ")");
System.out.println(
"\t[ -frate <r> ]\t\tFeature sampling rate (default="
+ RFRanker.featureSamplingRate
+ ")");
int type = (RFRanker.rType.ordinal() - RANKER_TYPE.MART.ordinal());
System.out.println(
"\t[ -rtype <type> ]\tRanker to bag (default=" + type + ", i.e. " + rType[type] + ")");
System.out.println(
"\t[ -tree <t> ]\t\tNumber of trees in each bag (default=" + RFRanker.nTrees + ")");
System.out.println(
"\t[ -leaf <l> ]\t\tNumber of leaves for each tree (default="
+ RFRanker.nTreeLeaves
+ ")");
System.out.println(
"\t[ -shrinkage <factor> ]\tShrinkage, or learning rate (default="
+ RFRanker.learningRate
+ ")");
System.out.println(
"\t[ -tc <k> ]\t\tNumber of threshold candidates for tree spliting. -1 to use all feature values (default="
+ RFRanker.nThreshold
+ ")");
System.out.println(
"\t[ -mls <n> ]\t\tMin leaf support -- minimum #samples each leaf has to contain (default="
+ RFRanker.minLeafSupport
+ ")");
System.out.println("");
System.out.println(" [+] Testing previously saved models");
System.out.println("\t-load <model>\t\tThe model to load");
System.out.println(
"\t-test <file>\t\tTest data to evaluate the model (specify either this or -rank but not both)");
System.out.println(
"\t-rank <file>\t\tRank the samples in the specified file (specify either this or -test but not both)");
System.out.println(
"\t[ -metric2T <metric> ]\tMetric to evaluate on the test data (default="
+ trainMetric
+ ")");
System.out.println(
"\t[ -gmax <label> ]\tHighest judged relevance label. It affects the calculation of ERR (default="
+ (int) SimpleMath.logBase2(ERRScorer.MAX)
+ ", i.e. 5-point scale {0,1,2,3,4})");
System.out.println(
"\t[ -score <file>]\tStore ranker's score for each object being ranked (has to be used with -rank)");
// System.out.println("\t[ -qrel <file> ]\tTREC-style relevance judgment file. It only affects
// MAP and NDCG (default=unspecified)");
System.out.println(
"\t[ -idv ]\t\tPrint model performance (in test metric) on individual ranked lists (has to be used with -test)");
System.out.println(
"\t[ -norm ]\t\tNormalize feature vectors (similar to -norm for training/tuning)");
/* System.out.println("");
System.out.println(" +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++");
System.out.println(" + NOTE: ALWAYS include -letor if you're doing experiments on LETOR 4.0 dataset. +");
System.out.println(" + The reason is a relevance degree of 2 in the dataset is actually counted as 3 +");
System.out.println(" + (this is based on the evaluation script they provided). To be consistent +");
System.out.println(" + with their numbers, this program will change 2 to 3 when it loads the data +");
System.out.println(" + into memory if the -letor flag is specified. +");
System.out.println(" +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++");
*/
System.out.println("");
return;
}
MyThreadPool.init(Runtime.getRuntime().availableProcessors());
// MyThreadPool.init(2);
for (int i = 0; i < args.length; i++) {
if (args[i].compareTo("-train") == 0) trainFile = args[++i];
else if (args[i].compareTo("-ranker") == 0) rankerType = Integer.parseInt(args[++i]);
else if (args[i].compareTo("-feature") == 0) featureDescriptionFile = args[++i];
else if (args[i].compareTo("-metric2t") == 0) trainMetric = args[++i];
else if (args[i].compareTo("-metric2T") == 0) testMetric = args[++i];
else if (args[i].compareTo("-gmax") == 0)
ERRScorer.MAX = Math.pow(2, Double.parseDouble(args[++i]));
else if (args[i].compareTo("-qrel") == 0) qrelFile = args[++i];
else if (args[i].compareTo("-tts") == 0) ttSplit = Double.parseDouble(args[++i]);
else if (args[i].compareTo("-tvs") == 0) tvSplit = Double.parseDouble(args[++i]);
else if (args[i].compareTo("-kcv") == 0) foldCV = Integer.parseInt(args[++i]);
else if (args[i].compareTo("-validate") == 0) validationFile = args[++i];
else if (args[i].compareTo("-test") == 0) testFile = args[++i];
else if (args[i].compareTo("-norm") == 0) {
Evaluator.normalize = true;
String n = args[++i];
if (n.compareTo("sum") == 0) Evaluator.nml = new SumNormalizor();
else if (n.compareTo("zscore") == 0) Evaluator.nml = new ZScoreNormalizor();
else {
System.out.println("Unknown normalizor: " + n);
System.out.println("System will now exit.");
System.exit(1);
}
} else if (args[i].compareTo("-save") == 0) Evaluator.modelFile = args[++i];
else if (args[i].compareTo("-silent") == 0) Ranker.verbose = false;
else if (args[i].compareTo("-load") == 0) {
savedModelFile = args[++i];
modelToLoad = args[i];
} else if (args[i].compareTo("-idv") == 0) printIndividual = true;
else if (args[i].compareTo("-rank") == 0) rankFile = args[++i];
else if (args[i].compareTo("-score") == 0) scoreFile = args[++i];
// Ranker-specific parameters
// RankNet
else if (args[i].compareTo("-epoch") == 0) {
RankNet.nIteration = Integer.parseInt(args[++i]);
ListNet.nIteration = Integer.parseInt(args[i]);
} else if (args[i].compareTo("-layer") == 0)
RankNet.nHiddenLayer = Integer.parseInt(args[++i]);
else if (args[i].compareTo("-node") == 0)
RankNet.nHiddenNodePerLayer = Integer.parseInt(args[++i]);
else if (args[i].compareTo("-lr") == 0) {
RankNet.learningRate = Double.parseDouble(args[++i]);
ListNet.learningRate = Neuron.learningRate;
}
// RankBoost
else if (args[i].compareTo("-tc") == 0) {
RankBoost.nThreshold = Integer.parseInt(args[++i]);
LambdaMART.nThreshold = Integer.parseInt(args[i]);
}
// AdaRank
else if (args[i].compareTo("-noeq") == 0) AdaRank.trainWithEnqueue = false;
else if (args[i].compareTo("-max") == 0) AdaRank.maxSelCount = Integer.parseInt(args[++i]);
// COORDINATE ASCENT
else if (args[i].compareTo("-r") == 0) CoorAscent.nRestart = Integer.parseInt(args[++i]);
else if (args[i].compareTo("-i") == 0) CoorAscent.nMaxIteration = Integer.parseInt(args[++i]);
// ranker-shared parameters
else if (args[i].compareTo("-round") == 0) {
RankBoost.nIteration = Integer.parseInt(args[++i]);
AdaRank.nIteration = Integer.parseInt(args[i]);
} else if (args[i].compareTo("-reg") == 0) {
CoorAscent.slack = Double.parseDouble(args[++i]);
CoorAscent.regularized = true;
} else if (args[i].compareTo("-tolerance") == 0) {
AdaRank.tolerance = Double.parseDouble(args[++i]);
CoorAscent.tolerance = Double.parseDouble(args[i]);
}
// MART / LambdaMART / Random forest
else if (args[i].compareTo("-tree") == 0) {
LambdaMART.nTrees = Integer.parseInt(args[++i]);
RFRanker.nTrees = Integer.parseInt(args[i]);
} else if (args[i].compareTo("-leaf") == 0) {
LambdaMART.nTreeLeaves = Integer.parseInt(args[++i]);
RFRanker.nTreeLeaves = Integer.parseInt(args[i]);
} else if (args[i].compareTo("-shrinkage") == 0) {
LambdaMART.learningRate = Float.parseFloat(args[++i]);
RFRanker.learningRate = Float.parseFloat(args[i]);
} else if (args[i].compareTo("-mls") == 0) {
LambdaMART.minLeafSupport = Integer.parseInt(args[++i]);
RFRanker.minLeafSupport = Integer.parseInt(args[i]);
} else if (args[i].compareTo("-estop") == 0)
LambdaMART.nRoundToStopEarly = Integer.parseInt(args[++i]);
// Random forest
else if (args[i].compareTo("-bag") == 0) RFRanker.nBag = Integer.parseInt(args[++i]);
else if (args[i].compareTo("-srate") == 0)
RFRanker.subSamplingRate = Float.parseFloat(args[++i]);
else if (args[i].compareTo("-frate") == 0)
RFRanker.featureSamplingRate = Float.parseFloat(args[++i]);
else if (args[i].compareTo("-letor") == 0) letor = true;
/////////////////////////////////////////////////////
// These parameters are *ONLY* for my personal use
/////////////////////////////////////////////////////
else if (args[i].compareTo("-nf") == 0) newFeatureFile = args[++i];
else if (args[i].compareTo("-keep") == 0) keepOrigFeatures = true;
else if (args[i].compareTo("-t") == 0) topNew = Integer.parseInt(args[++i]);
else if (args[i].compareTo("-indri") == 0) indriRankingFile = args[++i];
else if (args[i].compareTo("-hr") == 0) mustHaveRelDoc = true;
else {
System.out.println("Unknown command-line parameter: " + args[i]);
System.out.println("System will now exit.");
System.exit(1);
}
}
if (testMetric.compareTo("") == 0) testMetric = trainMetric;
System.out.println("");
// System.out.println((keepOrigFeatures)?"Keep orig. features":"Discard orig. features");
System.out.println("[+] General Parameters:");
System.out.println("LETOR 4.0 dataset: " + (letor ? "Yes" : "No"));
Evaluator e = new Evaluator(rType2[rankerType], trainMetric, testMetric);
if (trainFile.compareTo("") != 0) {
System.out.println("Training data:\t" + trainFile);
if (foldCV != -1) {
System.out.println("Cross validation: " + foldCV + " folds.");
} else {
if (testFile.compareTo("") != 0) System.out.println("Test data:\t" + testFile);
else if (ttSplit > 0.0) // choose to split training data into train and test
System.out.println("Train-Test split: " + ttSplit);
if (validationFile.compareTo("") != 0) // the user has specified the validation set
System.out.println("Validation data:\t" + validationFile);
else if (ttSplit <= 0.0 && tvSplit > 0.0)
System.out.println("Train-Validation split: " + tvSplit);
}
System.out.println("Ranking method:\t" + rType[rankerType]);
if (featureDescriptionFile.compareTo("") != 0)
System.out.println("Feature description file:\t" + featureDescriptionFile);
else System.out.println("Feature description file:\tUnspecified. All features will be used.");
System.out.println("Train metric:\t" + trainMetric);
System.out.println("Test metric:\t" + testMetric);
if (trainMetric.toUpperCase().startsWith("ERR") || testMetric.toUpperCase().startsWith("ERR"))
System.out.println(
"Highest relevance label (to compute ERR): "
+ (int) SimpleMath.logBase2(ERRScorer.MAX));
if (qrelFile.compareTo("") != 0)
System.out.println(
"TREC-format relevance judgment (only affects MAP and NDCG scores): " + qrelFile);
System.out.println(
"Feature normalization: " + ((Evaluator.normalize) ? Evaluator.nml.name() : "No"));
if (modelFile.compareTo("") != 0) System.out.println("Model file: " + modelFile);
System.out.println("");
System.out.println("[+] " + rType[rankerType] + "'s Parameters:");
RankerFactory rf = new RankerFactory();
rf.createRanker(rType2[rankerType]).printParameters();
System.out.println("");
// starting to do some work
if (foldCV != -1) e.evaluate(trainFile, featureDescriptionFile, foldCV);
else {
if (ttSplit > 0.0) // we should use a held-out portion of the training data for testing?
e.evaluate(trainFile, validationFile, featureDescriptionFile, ttSplit);
else if (tvSplit > 0.0) // should we use a portion of the training data for validation?
e.evaluate(trainFile, tvSplit, testFile, featureDescriptionFile);
else e.evaluate(trainFile, validationFile, testFile, featureDescriptionFile);
}
} else // scenario: test a saved model
{
System.out.println("Model file:\t" + savedModelFile);
System.out.println(
"Feature normalization: " + ((Evaluator.normalize) ? Evaluator.nml.name() : "No"));
if (rankFile.compareTo("") != 0) {
if (scoreFile.compareTo("") != 0) e.score(savedModelFile, rankFile, scoreFile);
else if (indriRankingFile.compareTo("") != 0)
e.rank(savedModelFile, rankFile, indriRankingFile);
else e.rank(savedModelFile, rankFile);
} else {
System.out.println("Test metric:\t" + testMetric);
if (testMetric.startsWith("ERR"))
System.out.println(
"Highest relevance label (to compute ERR): "
+ (int) SimpleMath.logBase2(ERRScorer.MAX));
if (savedModelFile.compareTo("") != 0) e.test(savedModelFile, testFile, printIndividual);
// This is *ONLY* for my personal use. It is *NOT* exposed via cmd-line
// It will evaluate the input ranking (without being reranked by any model) using any
// measure specified via metric2T
else e.test(testFile);
}
}
MyThreadPool.getInstance().shutdown();
}