Use Adagrad optimiser for Linear regression by default (opensearch-pr…

…oject#3291) * Use AdaGrad optimiser by default in Liner Resgression Signed-off-by: rithin-pullela-aws <rithinp@amazon.com> * Added issue link in the code comment as a reference. Signed-off-by: rithin-pullela-aws <rithinp@amazon.com> * Apply Spotless Signed-off-by: rithin-pullela-aws <rithinp@amazon.com> --------- Signed-off-by: rithin-pullela-aws <rithinp@amazon.com>
dhrubo-os · Dec 30, 2024 · f323141 · f323141
1 parent ff85b12
commit f323141
Showing 1 changed file with 7 additions and 6 deletions.
diff --git a/...rithms/src/main/java/org/opensearch/ml/engine/algorithms/regression/LinearRegression.java b/...rithms/src/main/java/org/opensearch/ml/engine/algorithms/regression/LinearRegression.java
@@ -52,7 +52,7 @@
 public class LinearRegression implements Trainable, Predictable {
     public static final String VERSION = "1.0.0";
     private static final LinearRegressionParams.ObjectiveType DEFAULT_OBJECTIVE_TYPE = LinearRegressionParams.ObjectiveType.SQUARED_LOSS;
-    private static final LinearRegressionParams.OptimizerType DEFAULT_OPTIMIZER_TYPE = LinearRegressionParams.OptimizerType.SIMPLE_SGD;
+    private static final LinearRegressionParams.OptimizerType DEFAULT_OPTIMIZER_TYPE = LinearRegressionParams.OptimizerType.ADA_GRAD;
     private static final double DEFAULT_LEARNING_RATE = 0.01;
     // Momentum
     private static final double DEFAULT_MOMENTUM_FACTOR = 0;
@@ -134,15 +134,15 @@ private void createOptimiser() {
                 break;
         }
         switch (optimizerType) {
+            case SIMPLE_SGD:
+                optimiser = SGD.getSimpleSGD(learningRate, momentumFactor, momentum);
+                break;
             case LINEAR_DECAY_SGD:
                 optimiser = SGD.getLinearDecaySGD(learningRate, momentumFactor, momentum);
                 break;
             case SQRT_DECAY_SGD:
                 optimiser = SGD.getSqrtDecaySGD(learningRate, momentumFactor, momentum);
                 break;
-            case ADA_GRAD:
-                optimiser = new AdaGrad(learningRate, epsilon);
-                break;
             case ADA_DELTA:
                 optimiser = new AdaDelta(momentumFactor, epsilon);
                 break;
@@ -153,8 +153,9 @@ private void createOptimiser() {
                 optimiser = new RMSProp(learningRate, momentumFactor, epsilon, decayRate);
                 break;
             default:
-                // Use default SGD with a constant learning rate.
-                optimiser = SGD.getSimpleSGD(learningRate, momentumFactor, momentum);
+                // Use AdaGrad by default, reference issue:
+                // https://github.com/opensearch-project/ml-commons/issues/3210#issuecomment-2556119802
+                optimiser = new AdaGrad(learningRate, epsilon);
                 break;
         }
     }