feat(#43): Convolutional Neural Networks

example/src/main/java/de/example/cnn/CnnExampleOnMNIST.java

@@ -0,0 +1,105 @@
+package de.example.cnn;
+
+import de.edux.ml.api.ExecutionMode;
+import de.edux.ml.mlp.core.network.NetworkBuilder;
+import de.edux.ml.mlp.core.network.layers.ConvolutionalLayer;
+import de.edux.ml.mlp.core.network.layers.DenseLayer;
+import de.edux.ml.mlp.core.network.layers.FlattenLayer;
+import de.edux.ml.mlp.core.network.layers.PoolingLayer;
+import de.edux.ml.mlp.core.network.layers.ReLuLayer;
+import de.edux.ml.mlp.core.network.layers.SoftmaxLayer;
+import de.edux.ml.mlp.core.network.loader.Loader;
+import de.edux.ml.mlp.core.network.loader.MetaData;
+import de.edux.ml.mlp.core.network.loader.mnist.MnistLoader;
+
+import java.io.File;
+
+/**
+ * @author Samuel Abramov
+ */
+public class CnnExampleOnMNIST {
+    public static void main(String[] args) {
+        String trainImages =
+                "example"
+                        + File.separator
+                        + "datasets"
+                        + File.separator
+                        + "mnist"
+                        + File.separator
+                        + "train-images.idx3-ubyte";
+        String trainLabels =
+                "example"
+                        + File.separator
+                        + "datasets"
+                        + File.separator
+                        + "mnist"
+                        + File.separator
+                        + "train-labels.idx1-ubyte";
+        String testImages =
+                "example"
+                        + File.separator
+                        + "datasets"
+                        + File.separator
+                        + "mnist"
+                        + File.separator
+                        + "t10k-images.idx3-ubyte";
+        String testLabels =
+                "example"
+                        + File.separator
+                        + "datasets"
+                        + File.separator
+                        + "mnist"
+                        + File.separator
+                        + "t10k-labels.idx1-ubyte";
+
+        int batchSize = 100;
+        int epochs = 10;
+        float initialLearningRate = 0.1f;
+        float finalLearningRate = 0.0001f;
+
+        Loader trainLoader = new MnistLoader(trainImages, trainLabels, batchSize);
+        Loader testLoader = new MnistLoader(testImages, testLabels, batchSize);
+        MetaData trainMetaData = trainLoader.open();
+        int inputSize = trainMetaData.getInputSize();
+        int numberOfOutputClasses = trainMetaData.getNumberOfClasses();
+        trainLoader.close();
+
+        long startTime = System.currentTimeMillis();
+        new NetworkBuilder()
+                .addLayer(new ConvolutionalLayer(8, 3, 28, 28, 1)) // 8 Filter, 3x3, input 28x28, 1 grayscale channel
+                .addLayer(new ReLuLayer())
+                .addLayer(new PoolingLayer(8, 26, 26, 2, 2, 2)) // Pooling layer (2x2, stride 2)
+
+                .addLayer(new ConvolutionalLayer(16, 3, 13, 13, 8)) // 16 Filter, 3x3, input 13x13, 8 channels
+                .addLayer(new ReLuLayer())
+                .addLayer(new PoolingLayer(16, 11, 11, 2, 2, 2)) // Pooling layer (2x2, stride 2)
+
+                .addLayer(new FlattenLayer(16, 5, 5)) // Updated dimensions: 16 channels, 5x5 output
+                .addLayer(new DenseLayer(16 * 5 * 5, 128)) // Dense layer input from flattened convolution output
+                .addLayer(new ReLuLayer())
+                .addLayer(new DenseLayer(128, numberOfOutputClasses)) // Final dense layer with number of classes as output
+                .addLayer(new SoftmaxLayer())
+
+                // Hyperparameter configuration
+                .withBatchSize(batchSize)
+                .withLearningRates(initialLearningRate, finalLearningRate)
+                .withExecutionMode(ExecutionMode.SINGLE_THREAD)
+                .withEpochs(epochs)
+
+                // Build network
+                .build()
+                .printArchitecture()
+                .fit(trainLoader, testLoader)
+                .saveModel("cnn_mnist_trained.edux");
+
+        long endTime = System.currentTimeMillis();
+        System.out.println("Training took: " + (endTime - startTime) / 1000 + " seconds");
+
+        new NetworkBuilder()
+                .withExecutionMode(ExecutionMode.SINGLE_THREAD)
+                .withEpochs(2)
+                .withLearningRates(0.001f, 0.001f)
+                .loadModel("cnn_mnist_trained.edux")
+                .fit(trainLoader, testLoader);
+    }
+}

example/src/main/java/de/example/mlp/MlpExampleOnMNIST.java

@@ -47,7 +47,7 @@ public static void main(String[] args) {
 
     int batchSize = 100;
     ExecutionMode singleThread = ExecutionMode.SINGLE_THREAD;
-    int epochs = 100;
+    int epochs = 5;
     float initialLearningRate = 0.1f;
     float finalLearningRate = 0.0001f;
 
@@ -81,7 +81,7 @@ public static void main(String[] args) {
     // Loading a trained model
     new NetworkBuilder()
         .withExecutionMode(singleThread)
-        .withEpochs(5)
+        .withEpochs(2)
         .withLearningRates(0.001f, 0.001f)
         .loadModel("mnist_trained.edux")
         .fit(trainLoader, testLoader);

lib/src/main/java/de/edux/ml/api/ExecutionMode.java

@@ -16,7 +16,7 @@ public enum ExecutionMode {
    * Single-thread execution mode. In this mode, all batches are processed sequentially in a single
    * thread.
    */
-  SINGLE_THREAD(1);
+  SINGLE_THREAD(1), MULTI_THREAD(6);
 
   int threads = 1;