Small performance improvement

xEcho1337 · Dec 9, 2024 · b6eb524 · b6eb524
1 parent dfe2de3
commit b6eb524
Show file tree

Hide file tree

Showing 2 changed files with 11 additions and 9 deletions.
diff --git a/src/main/java/net/echo/brain4j/training/optimizers/impl/Adam.java b/src/main/java/net/echo/brain4j/training/optimizers/impl/Adam.java
@@ -67,15 +67,16 @@ public void postIteration(List<Layer> layers) {
         this.beta1Timestep = Math.pow(beta1, timestep);
         this.beta2Timestep = Math.pow(beta2, timestep);
 
-        for (Layer layer : layers) {
-            // 30% improvement using parallel stream. TODO: Implement GPU support for better parallelization
-            layer.getSynapses().parallelStream().forEach(this::update);
+        layers.parallelStream().forEach(layer -> {
+            for (Synapse synapse : layer.getSynapses()) {
+                update(synapse);
+            }
 
             for (Neuron neuron : layer.getNeurons()) {
                 double deltaBias = learningRate * neuron.getDelta();
                 neuron.setBias(neuron.getBias() + deltaBias);
             }
-        }
+        });
     }
 
     public double getBeta1() {

diff --git a/src/test/java/XorTest.java b/src/test/java/XorTest.java
@@ -13,9 +13,9 @@ public class XorTest {
     public static void main(String[] args) {
         Model network = new Model(
                 new DenseLayer(2, Activations.LINEAR),
-                new DenseLayer(32, Activations.RELU),
-                new DenseLayer(32, Activations.RELU),
-                new DenseLayer(32, Activations.RELU),
+                new DenseLayer(256, Activations.RELU),
+                new DenseLayer(256, Activations.RELU),
+                new DenseLayer(256, Activations.RELU),
                 new DenseLayer(1, Activations.SIGMOID)
         );
 
@@ -32,14 +32,15 @@ public static void main(String[] args) {
 
         long start = System.nanoTime();
 
-        for (int i = 0; i < 1000; i++) {
+        int epoches = 500;
+        for (int i = 0; i < epoches; i++) {
             network.fit(training, 1);
         }
 
         double error = network.evaluate(training);
         double took = (System.nanoTime() - start) / 1e6;
 
-        System.out.println("Took " + took + " ms with an average of " + (took / 1000) + " ms per epoch and error " + error);
+        System.out.println("Took " + took + " ms with an average of " + (took / epoches) + " ms per epoch and error " + error);
 
         for (DataRow row : training.getDataRows()) {
             Vector output = network.predict(row.inputs());