List of usage examples for org.apache.mahout.classifier.sgd OnlineLogisticRegression classifyFull
public Vector classifyFull(Vector r, Vector instance)
From source file:com.cloudera.knittingboar.records.TestTwentyNewsgroupsCustomRecordParseOLRRun.java
License:Apache License
@Test public void testRecordFactoryOnDatasetShard() throws Exception { // TODO a test with assertions is not a test // p.270 ----- metrics to track lucene's parsing mechanics, progress, // performance of OLR ------------ double averageLL = 0.0; double averageCorrect = 0.0; int k = 0;//from ww w . j a va2s . c o m double step = 0.0; int[] bumps = new int[] { 1, 2, 5 }; TwentyNewsgroupsRecordFactory rec_factory = new TwentyNewsgroupsRecordFactory("\t"); // rec_factory.setClassSplitString("\t"); JobConf job = new JobConf(defaultConf); long block_size = localFs.getDefaultBlockSize(workDir); LOG.info("default block size: " + (block_size / 1024 / 1024) + "MB"); // matches the OLR setup on p.269 --------------- // stepOffset, decay, and alpha --- describe how the learning rate decreases // lambda: amount of regularization // learningRate: amount of initial learning rate @SuppressWarnings("resource") OnlineLogisticRegression learningAlgorithm = new OnlineLogisticRegression(20, FEATURES, new L1()).alpha(1) .stepOffset(1000).decayExponent(0.9).lambda(3.0e-5).learningRate(20); FileInputFormat.setInputPaths(job, workDir); // try splitting the file in a variety of sizes TextInputFormat format = new TextInputFormat(); format.configure(job); Text value = new Text(); int numSplits = 1; InputSplit[] splits = format.getSplits(job, numSplits); LOG.info("requested " + numSplits + " splits, splitting: got = " + splits.length); LOG.info("---- debug splits --------- "); rec_factory.Debug(); int total_read = 0; for (int x = 0; x < splits.length; x++) { LOG.info("> Split [" + x + "]: " + splits[x].getLength()); int count = 0; InputRecordsSplit custom_reader = new InputRecordsSplit(job, splits[x]); while (custom_reader.next(value)) { Vector v = new RandomAccessSparseVector(TwentyNewsgroupsRecordFactory.FEATURES); int actual = rec_factory.processLine(value.toString(), v); String ng = rec_factory.GetNewsgroupNameByID(actual); // calc stats --------- double mu = Math.min(k + 1, 200); double ll = learningAlgorithm.logLikelihood(actual, v); averageLL = averageLL + (ll - averageLL) / mu; Vector p = new DenseVector(20); learningAlgorithm.classifyFull(p, v); int estimated = p.maxValueIndex(); int correct = (estimated == actual ? 1 : 0); averageCorrect = averageCorrect + (correct - averageCorrect) / mu; learningAlgorithm.train(actual, v); k++; int bump = bumps[(int) Math.floor(step) % bumps.length]; int scale = (int) Math.pow(10, Math.floor(step / bumps.length)); if (k % (bump * scale) == 0) { step += 0.25; LOG.info(String.format("%10d %10.3f %10.3f %10.2f %s %s", k, ll, averageLL, averageCorrect * 100, ng, rec_factory.GetNewsgroupNameByID(estimated))); } learningAlgorithm.close(); count++; } LOG.info("read: " + count + " records for split " + x); total_read += count; } // for each split LOG.info("total read across all splits: " + total_read); rec_factory.Debug(); }
From source file:com.cloudera.knittingboar.sgd.olr.TestBaseOLRTest20Newsgroups.java
License:Apache License
public void testResults() throws Exception { OnlineLogisticRegression classifier = ModelSerializer .readBinary(new FileInputStream(model20News.toString()), OnlineLogisticRegression.class); Text value = new Text(); long batch_vec_factory_time = 0; int k = 0;//from ww w.j a v a 2 s . co m int num_correct = 0; // ---- this all needs to be done in JobConf job = new JobConf(defaultConf); // TODO: work on this, splits are generating for everything in dir // InputSplit[] splits = generateDebugSplits(inputDir, job); //fullRCV1Dir InputSplit[] splits = generateDebugSplits(testData20News, job); System.out.println("split count: " + splits.length); InputRecordsSplit custom_reader_0 = new InputRecordsSplit(job, splits[0]); TwentyNewsgroupsRecordFactory VectorFactory = new TwentyNewsgroupsRecordFactory("\t"); for (int x = 0; x < 8000; x++) { if (custom_reader_0.next(value)) { long startTime = System.currentTimeMillis(); Vector v = new RandomAccessSparseVector(FEATURES); int actual = VectorFactory.processLine(value.toString(), v); long endTime = System.currentTimeMillis(); //System.out.println("That took " + (endTime - startTime) + " milliseconds"); batch_vec_factory_time += (endTime - startTime); String ng = VectorFactory.GetClassnameByID(actual); //.GetNewsgroupNameByID( actual ); // calc stats --------- double mu = Math.min(k + 1, 200); double ll = classifier.logLikelihood(actual, v); //averageLL = averageLL + (ll - averageLL) / mu; metrics.AvgLogLikelihood = metrics.AvgLogLikelihood + (ll - metrics.AvgLogLikelihood) / mu; Vector p = new DenseVector(20); classifier.classifyFull(p, v); int estimated = p.maxValueIndex(); int correct = (estimated == actual ? 1 : 0); if (estimated == actual) { num_correct++; } //averageCorrect = averageCorrect + (correct - averageCorrect) / mu; metrics.AvgCorrect = metrics.AvgCorrect + (correct - metrics.AvgCorrect) / mu; //this.polr.train(actual, v); k++; // if (x == this.BatchSize - 1) { int bump = bumps[(int) Math.floor(step) % bumps.length]; int scale = (int) Math.pow(10, Math.floor(step / bumps.length)); if (k % (bump * scale) == 0) { step += 0.25; System.out.printf( "Worker %s:\t Tested Recs: %10d, numCorrect: %d, AvgLL: %10.3f, Percent Correct: %10.2f, VF: %d\n", "OLR-standard-test", k, num_correct, metrics.AvgLogLikelihood, metrics.AvgCorrect * 100, batch_vec_factory_time); } classifier.close(); } else { // nothing else to process in split! break; } // if } // for the number of passes in the run }
From source file:com.cloudera.knittingboar.sgd.olr.TestBaseOLR_Train20Newsgroups.java
License:Apache License
public void testTrainNewsGroups() throws IOException { File base = new File("/Users/jpatterson/Downloads/datasets/20news-bydate/20news-bydate-train/"); overallCounts = HashMultiset.create(); long startTime = System.currentTimeMillis(); // p.269 --------------------------------------------------------- Map<String, Set<Integer>> traceDictionary = new TreeMap<String, Set<Integer>>(); // encodes the text content in both the subject and the body of the email FeatureVectorEncoder encoder = new StaticWordValueEncoder("body"); encoder.setProbes(2);//from w w w .j a va 2s.c o m encoder.setTraceDictionary(traceDictionary); // provides a constant offset that the model can use to encode the average frequency // of each class FeatureVectorEncoder bias = new ConstantValueEncoder("Intercept"); bias.setTraceDictionary(traceDictionary); // used to encode the number of lines in a message FeatureVectorEncoder lines = new ConstantValueEncoder("Lines"); lines.setTraceDictionary(traceDictionary); FeatureVectorEncoder logLines = new ConstantValueEncoder("LogLines"); logLines.setTraceDictionary(traceDictionary); Dictionary newsGroups = new Dictionary(); // matches the OLR setup on p.269 --------------- // stepOffset, decay, and alpha --- describe how the learning rate decreases // lambda: amount of regularization // learningRate: amount of initial learning rate OnlineLogisticRegression learningAlgorithm = new OnlineLogisticRegression(20, FEATURES, new L1()).alpha(1) .stepOffset(1000).decayExponent(0.9).lambda(3.0e-5).learningRate(20); // bottom of p.269 ------------------------------ // because OLR expects to get integer class IDs for the target variable during training // we need a dictionary to convert the target variable (the newsgroup name) // to an integer, which is the newsGroup object List<File> files = new ArrayList<File>(); for (File newsgroup : base.listFiles()) { newsGroups.intern(newsgroup.getName()); files.addAll(Arrays.asList(newsgroup.listFiles())); } // mix up the files, helps training in OLR Collections.shuffle(files); System.out.printf("%d training files\n", files.size()); // p.270 ----- metrics to track lucene's parsing mechanics, progress, performance of OLR ------------ double averageLL = 0.0; double averageCorrect = 0.0; double averageLineCount = 0.0; int k = 0; double step = 0.0; int[] bumps = new int[] { 1, 2, 5 }; double lineCount = 0; // last line on p.269 Analyzer analyzer = new StandardAnalyzer(Version.LUCENE_31); Splitter onColon = Splitter.on(":").trimResults(); int input_file_count = 0; // ----- p.270 ------------ "reading and tokenzing the data" --------- for (File file : files) { BufferedReader reader = new BufferedReader(new FileReader(file)); input_file_count++; // identify newsgroup ---------------- // convert newsgroup name to unique id // ----------------------------------- String ng = file.getParentFile().getName(); int actual = newsGroups.intern(ng); Multiset<String> words = ConcurrentHashMultiset.create(); // check for line count header ------- String line = reader.readLine(); while (line != null && line.length() > 0) { // if this is a line that has a line count, let's pull that value out ------ if (line.startsWith("Lines:")) { String count = Iterables.get(onColon.split(line), 1); try { lineCount = Integer.parseInt(count); averageLineCount += (lineCount - averageLineCount) / Math.min(k + 1, 1000); } catch (NumberFormatException e) { // if anything goes wrong in parse: just use the avg count lineCount = averageLineCount; } } boolean countHeader = (line.startsWith("From:") || line.startsWith("Subject:") || line.startsWith("Keywords:") || line.startsWith("Summary:")); // loop through the lines in the file, while the line starts with: " " do { // get a reader for this specific string ------ StringReader in = new StringReader(line); // ---- count words in header --------- if (countHeader) { countWords(analyzer, words, in); } // iterate to the next string ---- line = reader.readLine(); } while (line.startsWith(" ")); } // while (lines in header) { // -------- count words in body ---------- countWords(analyzer, words, reader); reader.close(); // ----- p.271 ----------- Vector v = new RandomAccessSparseVector(FEATURES); // original value does nothing in a ContantValueEncoder bias.addToVector("", 1, v); // original value does nothing in a ContantValueEncoder lines.addToVector("", lineCount / 30, v); // original value does nothing in a ContantValueEncoder logLines.addToVector("", Math.log(lineCount + 1), v); // now scan through all the words and add them for (String word : words.elementSet()) { encoder.addToVector(word, Math.log(1 + words.count(word)), v); } //Utils.PrintVectorNonZero(v); // calc stats --------- double mu = Math.min(k + 1, 200); double ll = learningAlgorithm.logLikelihood(actual, v); averageLL = averageLL + (ll - averageLL) / mu; Vector p = new DenseVector(20); learningAlgorithm.classifyFull(p, v); int estimated = p.maxValueIndex(); int correct = (estimated == actual ? 1 : 0); averageCorrect = averageCorrect + (correct - averageCorrect) / mu; learningAlgorithm.train(actual, v); k++; int bump = bumps[(int) Math.floor(step) % bumps.length]; int scale = (int) Math.pow(10, Math.floor(step / bumps.length)); if (k % (bump * scale) == 0) { step += 0.25; System.out.printf("%10d %10.3f %10.3f %10.2f %s %s\n", k, ll, averageLL, averageCorrect * 100, ng, newsGroups.values().get(estimated)); } learningAlgorithm.close(); /* if (k>4) { break; } */ } Utils.PrintVectorSection(learningAlgorithm.getBeta().viewRow(0), 3); long endTime = System.currentTimeMillis(); //System.out.println("That took " + (endTime - startTime) + " milliseconds"); long duration = (endTime - startTime); System.out.println("Processed Input Files: " + input_file_count + ", time: " + duration + "ms"); ModelSerializer.writeBinary("/tmp/olr-news-group.model", learningAlgorithm); // learningAlgorithm.getBest().getPayload().getLearner().getModels().get(0)); }