Various improvements

Added possibility of having multiple input channels for convolution
(e.g. color images).
Load data on demand for training data sets that are too large to fit in
memory.
Added convergence test for auto-stopping training.
Threaded getPredictions().

Author: amten

src/amten/ml/Convolutions.java

@@ -62,7 +62,7 @@ public static Matrix antiPoolDelta(Matrix delta, Matrix prePoolRowIndexes, int n
         return result;
     }
 
-    public static Matrix generatePatchesFromInputLayer(Matrix inputs, int inputWidth, int patchWidth, int patchHeight) {
+    public static Matrix generatePatchesFromInputLayer(Matrix inputs, int inputWidth, int inputHeight, int patchWidth, int patchHeight) {
         // Input data has one row per example.
         // Input data has one column per pixel.
         //      Assumes pixel-numbers are generated row-wise.
@@ -71,20 +71,22 @@ public static Matrix generatePatchesFromInputLayer(Matrix inputs, int inputWidth
         // Output data have one row per example/patch
         // Output data have one column per patchPixel.
 
-        int inputHeight = inputs.numColumns() / inputWidth;
+        int numChannels = inputs.numColumns() / (inputWidth*inputHeight);
         int numPatchesPerExample = (inputWidth-patchWidth+1)*(inputHeight-patchHeight+1);
         int numExamples = inputs.numRows();
-        Matrix output = new Matrix(numExamples*numPatchesPerExample, patchWidth*patchHeight);
+        Matrix output = new Matrix(numExamples*numPatchesPerExample, numChannels*patchWidth*patchHeight);
         for (int example = 0; example < numExamples; example++) {
             int patchNum = 0;
             for (int inputStartY = 0; inputStartY < inputHeight-patchHeight+1; inputStartY++) {
                 for (int inputStartX = 0; inputStartX < inputWidth-patchWidth+1; inputStartX++) {
-                    for (int patchPixelY = 0; patchPixelY < patchHeight; patchPixelY++) {
-                        for (int patchPixelX = 0; patchPixelX < patchWidth; patchPixelX++) {
-                            int inputY = inputStartY + patchPixelY;
-                            int inputX = inputStartX + patchPixelX;
-                            double value = inputs.get(example, inputY*inputWidth + inputX);
-                            output.set(example*numPatchesPerExample + patchNum, patchPixelY*patchWidth + patchPixelX, value);
+                    for (int channel = 0; channel < numChannels; channel++) {
+                        for (int patchPixelY = 0; patchPixelY < patchHeight; patchPixelY++) {
+                            for (int patchPixelX = 0; patchPixelX < patchWidth; patchPixelX++) {
+                                int inputY = inputStartY + patchPixelY;
+                                int inputX = inputStartX + patchPixelX;
+                                double value = inputs.get(example, channel*inputHeight*inputWidth + inputY*inputWidth + inputX);
+                                output.set(example*numPatchesPerExample + patchNum, channel*patchHeight*patchWidth + patchPixelY*patchWidth + patchPixelX, value);
+                            }
                         }
                     }
                     patchNum++;

src/amten/ml/DataLoader.java

@@ -0,0 +1,29 @@
+package amten.ml;
+
+import amten.ml.matrix.Matrix;
+
+import java.io.IOException;
+
+/**
+ * Interface to be implemented by a class that loads training data from disk on-demand.
+ * Implement when training dataset is too large to fit in memory.
+ *
+ * @author Johannes Amtén
+ */
+public interface DataLoader {
+
+    /**
+     *
+     * @param xIDs Some type of id numbers for the datapoints to load.
+     * @return Training data for the specified ids
+     * @throws IOException
+     */
+    public Matrix loadData(Matrix xIDs) throws IOException;
+
+    /**
+     *
+     * @return Number of values for each datapoint
+     */
+    public int getDataSize();
+
+}