Auto-choose batchsize

Author: amten

src/amten/ml/NNParams.java

@@ -87,10 +87,11 @@ public NNParams() {
      * Number of examples to use in each mini-batch.
      * Batch-size 100 is a good choice for fully connected networks.
      * Batch-size 1 is a good choice for convolutional networks.
+     * If set to 0, a batch-size of 100 or 1 will be used depending on whether the network is fully connected or convolutional.
      *
-     * Default is 100.
+     * Default is 0.
      */
-    public int batchSize = 100;
+    public int batchSize = 0;
 
 
     /**

src/amten/ml/NeuralNetwork.java

@@ -113,6 +113,12 @@ public void train(Matrix x, Matrix y) throws Exception {
             myParams.inputWidth = inputSize;
         }
 
+        if (myParams.batchSize == 0) {
+            // Auto-choose batch-size.
+            // 100 for fully connected network and 1 for convolutional network.
+            myParams.batchSize = myLayerParams[1].isConvolutional() ? 1 : 100;
+        }
+
         initThetas();
         // If threads == 0, use the same number of threads as cores.
         myParams.numThreads = myParams.numThreads > 0 ? myParams.numThreads : Runtime.getRuntime().availableProcessors();

src/amten/ml/examples/NNClassificationExample.java

@@ -46,7 +46,6 @@ public static void runKaggleDigitsClassification(boolean useConvolution) throws
         params.numClasses = 10; // 10 digits to classify
         params.hiddenLayerParams = useConvolution ? new NNParams.NNLayerParams[]{ new NNParams.NNLayerParams(20, 5, 5, 2, 2) , new NNParams.NNLayerParams(100, 5, 5, 2, 2) } :
                                                     new NNParams.NNLayerParams[] { new NNParams.NNLayerParams(100) };
-        params.batchSize = useConvolution ? 1 : 100;
         params.maxIterations = useConvolution ? 10 : 200;
         params.learningRate = useConvolution ? 1E-2 : 0;
 

src/amten/ml/test/NeuralNetworkTest.java

@@ -41,7 +41,6 @@ public static void runKaggleDigitsClassification() throws Exception {
         params.numClasses = 10; // 10 digits to classify
         params.hiddenLayerParams = new NNParams.NNLayerParams[] { new NNParams.NNLayerParams(20, 5, 5, 2, 2) , new NNParams.NNLayerParams(100, 5, 5, 2, 2) };
         params.learningRate = 1E-2;
-        params.batchSize = 1;
         params.maxIterations = 10;
 
         long startTime = System.currentTimeMillis();

src/weka/classifiers/functions/NeuralNetwork.java

@@ -250,7 +250,7 @@ public void setBatchSize(int batchSize) {
         myParams.batchSize = batchSize;
     }
     public String batchSizeTipText() {
-        return "Number of training examples in each mini-batch (=1 recommended for convolutional networks) .";
+        return "Number of training examples in each mini-batch (0=Auto-choose) .";
     }
 
     public int getThreads() {