Adds a multi-label linear sgd classifier

Classification/Core/src/main/java/org/tribuo/classification/ImmutableLabelInfo.java

@@ -158,6 +158,11 @@ public String toReadableString() {
         return builder.toString();
     }
 
+    @Override
+    public String toString() {
+        return toReadableString();
+    }
+
     @Override
     public Iterator<Pair<Integer, Label>> iterator() {
         return new ImmutableInfoIterator(idLabelMap);

Classification/SGD/pom.xml

@@ -47,6 +47,11 @@
             <artifactId>tribuo-math</artifactId>
             <version>${project.version}</version>
         </dependency>
+        <dependency>
+            <groupId>${project.groupId}</groupId>
+            <artifactId>tribuo-common-sgd</artifactId>
+            <version>${project.version}</version>
+        </dependency>
         <!-- test time dependencies -->
         <dependency>
             <groupId>${project.groupId}</groupId>

Classification/SGD/src/main/java/org/tribuo/classification/sgd/LabelObjective.java

@@ -16,10 +16,8 @@
 
 package org.tribuo.classification.sgd;
 
-import com.oracle.labs.mlrg.olcut.config.Configurable;
-import com.oracle.labs.mlrg.olcut.provenance.ConfiguredObjectProvenance;
-import com.oracle.labs.mlrg.olcut.provenance.Provenancable;
 import com.oracle.labs.mlrg.olcut.util.Pair;
+import org.tribuo.common.sgd.SGDObjective;
 import org.tribuo.math.la.SGDVector;
 import org.tribuo.math.util.VectorNormalizer;
 
@@ -29,15 +27,23 @@
  * An objective knows if it generates a probabilistic model or not,
  * and what kind of normalization needs to be applied to produce probability values.
  */
-public interface LabelObjective extends Configurable, Provenancable<ConfiguredObjectProvenance> {
+public interface LabelObjective extends SGDObjective<Integer> {
 
     /**
      * Scores a prediction, returning the loss and a vector of per label gradients.
-     * @param truth The true label id.
+     *
+     * @deprecated In 4.1, to migrate to the new name {@link #lossAndGradient}.
+     * @param truth      The true label id.
      * @param prediction The prediction for each label id.
      * @return The score and per label gradient.
      */
-    public Pair<Double,SGDVector> valueAndGradient(int truth, SGDVector prediction);
+    @Deprecated
+    Pair<Double, SGDVector> valueAndGradient(int truth, SGDVector prediction);
+
+    @Override
+    default Pair<Double, SGDVector> lossAndGradient(Integer truth, SGDVector prediction) {
+        return valueAndGradient(truth, prediction);
+    }
 
     /**
      * Generates a new {@link VectorNormalizer} which normalizes the predictions into [0,1].

Classification/SGD/src/main/java/org/tribuo/classification/sgd/linear/LinearSGDModel.java

@@ -16,31 +16,21 @@
 
 package org.tribuo.classification.sgd.linear;
 
-import com.oracle.labs.mlrg.olcut.util.Pair;
 import org.tribuo.Example;
-import org.tribuo.Excuse;
-import org.tribuo.Feature;
 import org.tribuo.ImmutableFeatureMap;
 import org.tribuo.ImmutableOutputInfo;
-import org.tribuo.Model;
 import org.tribuo.Prediction;
 import org.tribuo.classification.Label;
+import org.tribuo.common.sgd.AbstractLinearSGDModel;
 import org.tribuo.math.LinearParameters;
 import org.tribuo.math.la.DenseMatrix;
 import org.tribuo.math.la.DenseVector;
-import org.tribuo.math.la.SparseVector;
 import org.tribuo.math.util.VectorNormalizer;
 import org.tribuo.provenance.ModelProvenance;
 
-import java.util.ArrayList;
-import java.util.Collections;
-import java.util.Comparator;
-import java.util.HashMap;
+import java.io.IOException;
 import java.util.LinkedHashMap;
-import java.util.List;
 import java.util.Map;
-import java.util.Optional;
-import java.util.PriorityQueue;
 
 /**
  * The inference time version of a linear model trained using SGD.
@@ -52,122 +42,89 @@
  * Proceedings of COMPSTAT, 2010.
  * </pre>
  */
-public class LinearSGDModel extends Model<Label> {
+public class LinearSGDModel extends AbstractLinearSGDModel<Label> {
     private static final long serialVersionUID = 2L;
 
-    private final DenseMatrix weights;
     private final VectorNormalizer normalizer;
 
-    LinearSGDModel(String name, ModelProvenance description,
-                   ImmutableFeatureMap featureIDMap, ImmutableOutputInfo<Label> labelIDMap,
+    // Unused as the weights now live in AbstractLinearSGDModel
+    // It remains for serialization compatibility with Tribuo 4.0
+    @Deprecated
+    private DenseMatrix weights = null;
+
+    /**
+     * Constructs a linear classification model trained via SGD.
+     * @param name The model name.
+     * @param provenance The model provenance.
+     * @param featureIDMap The feature domain.
+     * @param outputIDInfo The output domain.
+     * @param parameters The model parameters (i.e., the weight matrix).
+     * @param normalizer The normalization function.
+     * @param generatesProbabilities Does this model generate probabilities?
+     */
+    LinearSGDModel(String name, ModelProvenance provenance,
+                   ImmutableFeatureMap featureIDMap, ImmutableOutputInfo<Label> outputIDInfo,
                    LinearParameters parameters, VectorNormalizer normalizer, boolean generatesProbabilities) {
-        super(name, description, featureIDMap, labelIDMap, generatesProbabilities);
-        this.weights = parameters.getWeightMatrix();
+        super(name, provenance, featureIDMap, outputIDInfo, parameters.getWeightMatrix(), generatesProbabilities);
         this.normalizer = normalizer;
     }
 
-    private LinearSGDModel(String name, ModelProvenance description,
-                          ImmutableFeatureMap featureIDMap, ImmutableOutputInfo<Label> labelIDMap,
+    /**
+     * Constructs a linear classification model trained via SGD.
+     * @param name The model name.
+     * @param provenance The model provenance.
+     * @param featureIDMap The feature domain.
+     * @param outputIDInfo The output domain.
+     * @param weights The model parameters (i.e., the weight matrix).
+     * @param normalizer The normalization function.
+     * @param generatesProbabilities Does this model generate probabilities?
+     */
+    private LinearSGDModel(String name, ModelProvenance provenance,
+                          ImmutableFeatureMap featureIDMap, ImmutableOutputInfo<Label> outputIDInfo,
                           DenseMatrix weights, VectorNormalizer normalizer, boolean generatesProbabilities) {
-        super(name, description, featureIDMap, labelIDMap, generatesProbabilities);
-        this.weights = weights;
+        super(name, provenance, featureIDMap, outputIDInfo, weights, generatesProbabilities);
         this.normalizer = normalizer;
     }
 
     @Override
     public Prediction<Label> predict(Example<Label> example) {
-        SparseVector features = SparseVector.createSparseVector(example,featureIDMap,true);
-        // Due to bias feature
-        if (features.numActiveElements() == 1) {
-            throw new IllegalArgumentException("No features found in Example " + example.toString());
-        }
-        DenseVector prediction = weights.leftMultiply(features);
+        PredAndActive predTuple = predictSingle(example);
+        DenseVector prediction = predTuple.prediction;
         prediction.normalize(normalizer);
 
         double maxScore = Double.NEGATIVE_INFINITY;
         Label maxLabel = null;
         Map<String,Label> predMap = new LinkedHashMap<>();
         for (int i = 0; i < prediction.size(); i++) {
             String labelName = outputIDInfo.getOutput(i).getLabel();
-            Label label = new Label(labelName, prediction.get(i));
+            double score = prediction.get(i);
+            Label label = new Label(labelName, score);
             predMap.put(labelName,label);
-            if (label.getScore() > maxScore) {
-                maxScore = label.getScore();
+            if (score > maxScore) {
+                maxScore = score;
                 maxLabel = label;
             }
         }
-        return new Prediction<>(maxLabel, predMap, features.numActiveElements()-1, example, generatesProbabilities);
+        return new Prediction<>(maxLabel, predMap, predTuple.numActiveFeatures-1, example, generatesProbabilities);
     }
 
     @Override
-    public Map<String, List<Pair<String, Double>>> getTopFeatures(int n) {
-        int maxFeatures = n < 0 ? featureIDMap.size() + 1 : n;
-
-        Comparator<Pair<String,Double>> comparator = Comparator.comparingDouble(p -> Math.abs(p.getB()));
-
-        //
-        // Use a priority queue to find the top N features.
-        int numClasses = weights.getDimension1Size();
-        int numFeatures = weights.getDimension2Size()-1; //Removing the bias feature.
-        Map<String, List<Pair<String,Double>>> map = new HashMap<>();
-        for (int i = 0; i < numClasses; i++) {
-            PriorityQueue<Pair<String,Double>> q = new PriorityQueue<>(maxFeatures, comparator);
-
-            for (int j = 0; j < numFeatures; j++) {
-                Pair<String,Double> curr = new Pair<>(featureIDMap.get(j).getName(), weights.get(i,j));
-
-                if (q.size() < maxFeatures) {
-                    q.offer(curr);
-                } else if (comparator.compare(curr, q.peek()) > 0) {
-                    q.poll();
-                    q.offer(curr);
-                }
-            }
-            Pair<String,Double> curr = new Pair<>(BIAS_FEATURE, weights.get(i,numFeatures));
-
-            if (q.size() < maxFeatures) {
-                q.offer(curr);
-            } else if (comparator.compare(curr, q.peek()) > 0) {
-                q.poll();
-                q.offer(curr);
-            }
-            ArrayList<Pair<String,Double>> b = new ArrayList<>();
-            while (q.size() > 0) {
-                b.add(q.poll());
-            }
-
-            Collections.reverse(b);
-            map.put(outputIDInfo.getOutput(i).getLabel(), b);
-        }
-        return map;
+    protected LinearSGDModel copy(String newName, ModelProvenance newProvenance) {
+        return new LinearSGDModel(newName,newProvenance,featureIDMap,outputIDInfo,new DenseMatrix(baseWeights),normalizer,generatesProbabilities);
     }
 
     @Override
-    public Optional<Excuse<Label>> getExcuse(Example<Label> example) {
-        Prediction<Label> prediction = predict(example);
-        Map<String, List<Pair<String, Double>>> weightMap = new HashMap<>();
-        int numClasses = weights.getDimension1Size();
-        int numFeatures = weights.getDimension2Size()-1;
-
-        for (int i = 0; i < numClasses; i++) {
-            List<Pair<String, Double>> classScores = new ArrayList<>();
-            for (Feature f : example) {
-                int id = featureIDMap.getID(f.getName());
-                if (id > -1) {
-                    double score = weights.get(i,id) * f.getValue();
-                    classScores.add(new Pair<>(f.getName(), score));
-                }
-            }
-            classScores.add(new Pair<>(Model.BIAS_FEATURE,weights.get(i,numFeatures)));
-            classScores.sort((Pair<String, Double> o1, Pair<String, Double> o2) -> o2.getB().compareTo(o1.getB()));
-            weightMap.put(outputIDInfo.getOutput(i).getLabel(), classScores);
-        }
-
-        return Optional.of(new Excuse<>(example, prediction, weightMap));
+    protected String getDimensionName(int index) {
+        return outputIDInfo.getOutput(index).getLabel();
     }
 
-    @Override
-    protected LinearSGDModel copy(String newName, ModelProvenance newProvenance) {
-        return new LinearSGDModel(newName,newProvenance,featureIDMap,outputIDInfo,new DenseMatrix(weights),normalizer,generatesProbabilities);
+    private void readObject(java.io.ObjectInputStream in) throws IOException, ClassNotFoundException {
+        in.defaultReadObject();
+
+        // Bounce old 4.0 style models into the new 4.1 style models
+        if (weights != null && baseWeights == null) {
+            baseWeights = weights;
+            weights = null;
+        }
     }
 }