четверг, 4 сентября 2014 г.

namespace DropoutDemo

using System;
using System.Collections.Generic;

namespace DropoutDemo
{
  class DropoutProgram

  {
    static void Main(string[] args)
    {
      Console.WriteLine("\nBegin neural network dropout demo");
      Console.WriteLine("\nData is the famous Iris flower set.");
      Console.WriteLine("Input is sepal length, width, petal length, width");
      Console.WriteLine("Class to predict is species");
      Console.WriteLine("setosa = 0 0 1, versicolor = 0 1 0, virginica = 1 0 0 ");

      Console.WriteLine("\nRaw data has 150 total items like:\n");
      Console.WriteLine(" 5.1, 3.5, 1.4, 0.2, Iris setosa");
      Console.WriteLine(" 7.0, 3.2, 4.7, 1.4, Iris versicolor");
      Console.WriteLine(" 6.3, 3.3, 6.0, 2.5, Iris virginica");
      Console.WriteLine(" ......\n");

      Console.WriteLine("Loading 80-20% training-test data");

      double[][] trainData = new double[120][];
      trainData[0] = new double[] { 6.0, 3.4, 4.5, 1.6, 0, 1, 0 };
      trainData[1] = new double[] { 6.7, 2.5, 5.8, 1.8, 1, 0, 0 };
      trainData[2] = new double[] { 4.9, 3.1, 1.5, 0.1, 0, 0, 1 };
      trainData[3] = new double[] { 7.7, 2.8, 6.7, 2.0, 1, 0, 0 };
      trainData[4] = new double[] { 6.4, 3.2, 5.3, 2.3, 1, 0, 0 };
      trainData[5] = new double[] { 6.7, 3.1, 4.7, 1.5, 0, 1, 0 };
      trainData[6] = new double[] { 6.3, 3.4, 5.6, 2.4, 1, 0, 0 };
      trainData[7] = new double[] { 5.9, 3.2, 4.8, 1.8, 0, 1, 0 };
      trainData[8] = new double[] { 6.3, 2.5, 5.0, 1.9, 1, 0, 0 };
      trainData[9] = new double[] { 4.6, 3.2, 1.4, 0.2, 0, 0, 1 };
      trainData[10] = new double[] { 7.0, 3.2, 4.7, 1.4, 0, 1, 0 };
      trainData[11] = new double[] { 6.6, 3.0, 4.4, 1.4, 0, 1, 0 };
      trainData[12] = new double[] { 5.7, 2.8, 4.1, 1.3, 0, 1, 0 };
      trainData[13] = new double[] { 6.7, 3.0, 5.0, 1.7, 0, 1, 0 };
      trainData[14] = new double[] { 6.5, 3.0, 5.2, 2.0, 1, 0, 0 };
      trainData[15] = new double[] { 5.1, 3.8, 1.5, 0.3, 0, 0, 1 };
      trainData[16] = new double[] { 7.9, 3.8, 6.4, 2.0, 1, 0, 0 };
      trainData[17] = new double[] { 5.9, 3.0, 5.1, 1.8, 1, 0, 0 };
      trainData[18] = new double[] { 7.3, 2.9, 6.3, 1.8, 1, 0, 0 };
      trainData[19] = new double[] { 5.0, 2.0, 3.5, 1.0, 0, 1, 0 };
      trainData[20] = new double[] { 6.2, 2.8, 4.8, 1.8, 1, 0, 0 };
      trainData[21] = new double[] { 7.4, 2.8, 6.1, 1.9, 1, 0, 0 };
      trainData[22] = new double[] { 6.2, 3.4, 5.4, 2.3, 1, 0, 0 };
      trainData[23] = new double[] { 5.2, 3.5, 1.5, 0.2, 0, 0, 1 };
      trainData[24] = new double[] { 6.8, 3.0, 5.5, 2.1, 1, 0, 0 };
      trainData[25] = new double[] { 5.5, 2.6, 4.4, 1.2, 0, 1, 0 };
      trainData[26] = new double[] { 6.9, 3.1, 5.1, 2.3, 1, 0, 0 };
      trainData[27] = new double[] { 6.4, 2.7, 5.3, 1.9, 1, 0, 0 };
      trainData[28] = new double[] { 5.6, 2.7, 4.2, 1.3, 0, 1, 0 };
      trainData[29] = new double[] { 4.4, 3.0, 1.3, 0.2, 0, 0, 1 };
      trainData[30] = new double[] { 6.9, 3.1, 4.9, 1.5, 0, 1, 0 };
      trainData[31] = new double[] { 5.4, 3.0, 4.5, 1.5, 0, 1, 0 };
      trainData[32] = new double[] { 5.8, 2.7, 4.1, 1.0, 0, 1, 0 };
      trainData[33] = new double[] { 4.6, 3.6, 1.0, 0.2, 0, 0, 1 };
      trainData[34] = new double[] { 5.1, 3.5, 1.4, 0.2, 0, 0, 1 };
      trainData[35] = new double[] { 4.9, 3.0, 1.4, 0.2, 0, 0, 1 };
      trainData[36] = new double[] { 5.1, 3.4, 1.5, 0.2, 0, 0, 1 };
      trainData[37] = new double[] { 5.5, 2.4, 3.8, 1.1, 0, 1, 0 };
      trainData[38] = new double[] { 6.8, 2.8, 4.8, 1.4, 0, 1, 0 };
      trainData[39] = new double[] { 6.7, 3.0, 5.2, 2.3, 1, 0, 0 };
      trainData[40] = new double[] { 5.7, 3.0, 4.2, 1.2, 0, 1, 0 };
      trainData[41] = new double[] { 6.0, 2.2, 5.0, 1.5, 1, 0, 0 };
      trainData[42] = new double[] { 6.5, 2.8, 4.6, 1.5, 0, 1, 0 };
      trainData[43] = new double[] { 6.3, 2.5, 4.9, 1.5, 0, 1, 0 };
      trainData[44] = new double[] { 6.7, 3.1, 5.6, 2.4, 1, 0, 0 };
      trainData[45] = new double[] { 6.4, 2.8, 5.6, 2.1, 1, 0, 0 };
      trainData[46] = new double[] { 5.5, 2.4, 3.7, 1.0, 0, 1, 0 };
      trainData[47] = new double[] { 5.2, 3.4, 1.4, 0.2, 0, 0, 1 };
      trainData[48] = new double[] { 6.0, 2.2, 4.0, 1.0, 0, 1, 0 };
      trainData[49] = new double[] { 6.1, 2.8, 4.0, 1.3, 0, 1, 0 };
      trainData[50] = new double[] { 6.1, 3.0, 4.6, 1.4, 0, 1, 0 };
      trainData[51] = new double[] { 5.0, 3.2, 1.2, 0.2, 0, 0, 1 };
      trainData[52] = new double[] { 4.8, 3.4, 1.9, 0.2, 0, 0, 1 };
      trainData[53] = new double[] { 6.3, 3.3, 6.0, 2.5, 1, 0, 0 };
      trainData[54] = new double[] { 5.0, 3.5, 1.6, 0.6, 0, 0, 1 };
      trainData[55] = new double[] { 6.0, 3.0, 4.8, 1.8, 1, 0, 0 };
      trainData[56] = new double[] { 6.3, 2.8, 5.1, 1.5, 1, 0, 0 };
      trainData[57] = new double[] { 7.2, 3.2, 6.0, 1.8, 1, 0, 0 };
      trainData[58] = new double[] { 4.6, 3.4, 1.4, 0.3, 0, 0, 1 };
      trainData[59] = new double[] { 6.9, 3.2, 5.7, 2.3, 1, 0, 0 };
      trainData[60] = new double[] { 6.5, 3.0, 5.5, 1.8, 1, 0, 0 };
      trainData[61] = new double[] { 4.9, 3.1, 1.5, 0.1, 0, 0, 1 };
      trainData[62] = new double[] { 5.1, 3.8, 1.9, 0.4, 0, 0, 1 };
      trainData[63] = new double[] { 5.3, 3.7, 1.5, 0.2, 0, 0, 1 };
      trainData[64] = new double[] { 5.0, 3.3, 1.4, 0.2, 0, 0, 1 };
      trainData[65] = new double[] { 6.6, 2.9, 4.6, 1.3, 0, 1, 0 };
      trainData[66] = new double[] { 5.9, 3.0, 4.2, 1.5, 0, 1, 0 };
      trainData[67] = new double[] { 5.2, 2.7, 3.9, 1.4, 0, 1, 0 };
      trainData[68] = new double[] { 7.7, 3.0, 6.1, 2.3, 1, 0, 0 };
      trainData[69] = new double[] { 5.4, 3.9, 1.7, 0.4, 0, 0, 1 };
      trainData[70] = new double[] { 5.1, 3.5, 1.4, 0.3, 0, 0, 1 };
      trainData[71] = new double[] { 7.2, 3.6, 6.1, 2.5, 1, 0, 0 };
      trainData[72] = new double[] { 6.4, 3.2, 4.5, 1.5, 0, 1, 0 };
      trainData[73] = new double[] { 4.8, 3.0, 1.4, 0.3, 0, 0, 1 };
      trainData[74] = new double[] { 6.2, 2.2, 4.5, 1.5, 0, 1, 0 };
      trainData[75] = new double[] { 5.8, 2.7, 3.9, 1.2, 0, 1, 0 };
      trainData[76] = new double[] { 7.2, 3.0, 5.8, 1.6, 1, 0, 0 };
      trainData[77] = new double[] { 4.7, 3.2, 1.6, 0.2, 0, 0, 1 };
      trainData[78] = new double[] { 6.3, 2.3, 4.4, 1.3, 0, 1, 0 };
      trainData[79] = new double[] { 6.8, 3.2, 5.9, 2.3, 1, 0, 0 };
      trainData[80] = new double[] { 5.0, 2.3, 3.3, 1.0, 0, 1, 0 };
      trainData[81] = new double[] { 5.7, 2.5, 5.0, 2.0, 1, 0, 0 };
      trainData[82] = new double[] { 7.7, 2.6, 6.9, 2.3, 1, 0, 0 };
      trainData[83] = new double[] { 6.5, 3.0, 5.8, 2.2, 1, 0, 0 };
      trainData[84] = new double[] { 6.1, 2.8, 4.7, 1.2, 0, 1, 0 };
      trainData[85] = new double[] { 4.7, 3.2, 1.3, 0.2, 0, 0, 1 };
      trainData[86] = new double[] { 6.9, 3.1, 5.4, 2.1, 1, 0, 0 };
      trainData[87] = new double[] { 6.4, 3.1, 5.5, 1.8, 1, 0, 0 };
      trainData[88] = new double[] { 6.0, 2.9, 4.5, 1.5, 0, 1, 0 };
      trainData[89] = new double[] { 6.4, 2.9, 4.3, 1.3, 0, 1, 0 };
      trainData[90] = new double[] { 4.4, 2.9, 1.4, 0.2, 0, 0, 1 };
      trainData[91] = new double[] { 5.0, 3.6, 1.4, 0.2, 0, 0, 1 };
      trainData[92] = new double[] { 4.4, 3.2, 1.3, 0.2, 0, 0, 1 };
      trainData[93] = new double[] { 5.1, 3.7, 1.5, 0.4, 0, 0, 1 };
      trainData[94] = new double[] { 4.8, 3.1, 1.6, 0.2, 0, 0, 1 };
      trainData[95] = new double[] { 6.5, 3.2, 5.1, 2.0, 1, 0, 0 };
      trainData[96] = new double[] { 6.1, 2.9, 4.7, 1.4, 0, 1, 0 };
      trainData[97] = new double[] { 5.4, 3.7, 1.5, 0.2, 0, 0, 1 };
      trainData[98] = new double[] { 5.7, 3.8, 1.7, 0.3, 0, 0, 1 };
      trainData[99] = new double[] { 7.1, 3.0, 5.9, 2.1, 1, 0, 0 };
      trainData[100] = new double[] { 5.4, 3.9, 1.3, 0.4, 0, 0, 1 };
      trainData[101] = new double[] { 6.1, 2.6, 5.6, 1.4, 1, 0, 0 };
      trainData[102] = new double[] { 6.4, 2.8, 5.6, 2.2, 1, 0, 0 };
      trainData[103] = new double[] { 5.0, 3.0, 1.6, 0.2, 0, 0, 1 };
      trainData[104] = new double[] { 5.8, 2.8, 5.1, 2.4, 1, 0, 0 };
      trainData[105] = new double[] { 6.3, 2.9, 5.6, 1.8, 1, 0, 0 };
      trainData[106] = new double[] { 6.2, 2.9, 4.3, 1.3, 0, 1, 0 };
      trainData[107] = new double[] { 5.5, 3.5, 1.3, 0.2, 0, 0, 1 };
      trainData[108] = new double[] { 6.7, 3.1, 4.4, 1.4, 0, 1, 0 };
      trainData[109] = new double[] { 4.9, 3.1, 1.5, 0.1, 0, 0, 1 };
      trainData[110] = new double[] { 4.6, 3.1, 1.5, 0.2, 0, 0, 1 };
      trainData[111] = new double[] { 5.0, 3.5, 1.3, 0.3, 0, 0, 1 };
      trainData[112] = new double[] { 5.5, 2.5, 4.0, 1.3, 0, 1, 0 };
      trainData[113] = new double[] { 5.5, 4.2, 1.4, 0.2, 0, 0, 1 };
      trainData[114] = new double[] { 5.5, 2.3, 4.0, 1.3, 0, 1, 0 };
      trainData[115] = new double[] { 5.2, 4.1, 1.5, 0.1, 0, 0, 1 };
      trainData[116] = new double[] { 5.6, 2.5, 3.9, 1.1, 0, 1, 0 };
      trainData[117] = new double[] { 5.6, 2.9, 3.6, 1.3, 0, 1, 0 };
      trainData[118] = new double[] { 4.9, 2.4, 3.3, 1.0, 0, 1, 0 };
      trainData[119] = new double[] { 5.7, 2.8, 4.5, 1.3, 0, 1, 0 };

      double[][] testData = new double[30][];
      testData[0] = new double[] { 6.0, 2.7, 5.1, 1.6, 0, 1, 0 };
      testData[1] = new double[] { 5.1, 3.3, 1.7, 0.5, 0, 0, 1 };
      testData[2] = new double[] { 6.7, 3.3, 5.7, 2.1, 1, 0, 0 };
      testData[3] = new double[] { 5.1, 2.5, 3.0, 1.1, 0, 1, 0 };
      testData[4] = new double[] { 5.6, 2.8, 4.9, 2.0, 1, 0, 0 };
      testData[5] = new double[] { 5.8, 2.7, 5.1, 1.9, 1, 0, 0 };
      testData[6] = new double[] { 5.0, 3.4, 1.5, 0.2, 0, 0, 1 };
      testData[7] = new double[] { 5.4, 3.4, 1.7, 0.2, 0, 0, 1 };
      testData[8] = new double[] { 4.9, 2.5, 4.5, 1.7, 1, 0, 0 };
      testData[9] = new double[] { 5.7, 4.4, 1.5, 0.4, 0, 0, 1 };
      testData[10] = new double[] { 7.7, 3.8, 6.7, 2.2, 1, 0, 0 };
      testData[11] = new double[] { 5.7, 2.9, 4.2, 1.3, 0, 1, 0 };
      testData[12] = new double[] { 5.0, 3.4, 1.6, 0.4, 0, 0, 1 };
      testData[13] = new double[] { 6.3, 3.3, 4.7, 1.6, 0, 1, 0 };
      testData[14] = new double[] { 4.5, 2.3, 1.3, 0.3, 0, 0, 1 };
      testData[15] = new double[] { 4.8, 3.4, 1.6, 0.2, 0, 0, 1 };
      testData[16] = new double[] { 5.8, 4.0, 1.2, 0.2, 0, 0, 1 };
      testData[17] = new double[] { 6.7, 3.3, 5.7, 2.5, 1, 0, 0 };
      testData[18] = new double[] { 4.3, 3.0, 1.1, 0.1, 0, 0, 1 };
      testData[19] = new double[] { 5.4, 3.4, 1.5, 0.4, 0, 0, 1 };
      testData[20] = new double[] { 5.6, 3.0, 4.1, 1.3, 0, 1, 0 };
      testData[21] = new double[] { 6.1, 3.0, 4.9, 1.8, 1, 0, 0 };
      testData[22] = new double[] { 5.7, 2.6, 3.5, 1.0, 0, 1, 0 };
      testData[23] = new double[] { 5.8, 2.7, 5.1, 1.9, 1, 0, 0 };
      testData[24] = new double[] { 5.6, 3.0, 4.5, 1.5, 0, 1, 0 };
      testData[25] = new double[] { 4.8, 3.0, 1.4, 0.1, 0, 0, 1 };
      testData[26] = new double[] { 5.1, 3.8, 1.6, 0.2, 0, 0, 1 };
      testData[27] = new double[] { 7.6, 3.0, 6.6, 2.1, 1, 0, 0 };
      testData[28] = new double[] { 6.3, 2.7, 4.9, 1.8, 1, 0, 0 };
      testData[29] = new double[] { 5.8, 2.6, 4.0, 1.2, 0, 1, 0 };

      Console.WriteLine("\nFirst 5 rows of training data:");
      ShowMatrix(trainData, 5, 1, true);
      Console.WriteLine("First 3 rows of test data:");
      ShowMatrix(testData, 3, 1, true);

      Console.WriteLine("\nCreating a 4-input, 9-hidden, 3-output dropout neural network");
      Console.WriteLine("Using tanh (hidden) and softmax (output) activations");
      const int numInput = 4;
      const int numHidden = 9; 
      const int numOutput = 3;
      NeuralNetwork nn = new NeuralNetwork(numInput, numHidden, numOutput);

      int maxEpochs = 1000;
      double learnRate = 0.05;
      Console.WriteLine("Setting maxEpochs = 1000, learnRate = 0.05");
      Console.WriteLine("No momentum or weight decay");
      Console.WriteLine("No early exit on low error condition");

      Console.WriteLine("\nBeginning training using back-propagation with dropout\n");
      nn.Train(trainData, maxEpochs, learnRate);
      Console.WriteLine("Training complete\n");

      double[] weights = nn.GetWeights();
      Console.WriteLine("Final neural network weights and bias values:");
      ShowVector(weights, 10, 3, true);

      double trainAcc = nn.Accuracy(trainData);
      Console.WriteLine("\nAccuracy on training data = " + trainAcc.ToString("F4"));

      double testAcc = nn.Accuracy(testData);
      Console.WriteLine("\nAccuracy on test data = " + testAcc.ToString("F4"));

      Console.WriteLine("\nEnd dropout demo\n");
      Console.ReadLine();
    } // Main

    static void ShowVector(double[] vector, int valsPerRow, int decimals, bool newLine)
    {
      for (int i = 0; i < vector.Length; ++i)
      {
        if (i % valsPerRow == 0) Console.WriteLine("");
        Console.Write(vector[i].ToString("F" + decimals).PadLeft(decimals + 4) + " ");
      }
      if (newLine == true) Console.WriteLine("");
    }

    static void ShowMatrix(double[][] matrix, int numRows, int decimals, bool newLine)
    {
      for (int i = 0; i < numRows; ++i)
      {
        Console.Write(i.ToString().PadLeft(3) + ": ");
        for (int j = 0; j < matrix[i].Length; ++j)
        {
          if (matrix[i][j] >= 0.0) Console.Write(" "); else Console.Write("-"); ;
          Console.Write(Math.Abs(matrix[i][j]).ToString("F" + decimals) + " ");
        }
        Console.WriteLine("");
      }
      if (newLine == true) Console.WriteLine("");
    }

    static double[] MySoftmax(double[] oSums) // does all output nodes at once so scale doesn't have to be re-computed each time
    {
      // determine max output sum
      double max = oSums[0];
      for (int i = 0; i < oSums.Length; ++i)
        if (oSums[i] > max) max = oSums[i];

      // determine scaling factor -- sum of exp(each val - max)
      double scale = 0.0;
      for (int i = 0; i < oSums.Length; ++i)
        scale += Math.Exp(oSums[i] - max);

      double[] result = new double[oSums.Length];
      for (int i = 0; i < oSums.Length; ++i)
        result[i] = Math.Exp(oSums[i] - max) / scale;

      return result; // now scaled so that xi sum to 1.0
    }

  } // class Program

  public class NeuralNetwork
  {
    private static Random rnd;

    private int numInput;
    private int numHidden;
    private int numOutput;

    private double[] inputs;
    private double[][] ihWeights; // input-hidden
    private double[] hBiases;
    private double[] hOutputs;

    private double[][] hoWeights; // hidden-output
    private double[] oBiases;
    private double[] outputs;

    public NeuralNetwork(int numInput, int numHidden, int numOutput)
    {
      rnd = new Random(0); // multi-purpose, inc. dropout

      this.numInput = numInput;
      this.numHidden = numHidden;
      this.numOutput = numOutput;

      this.inputs = new double[numInput];
      this.ihWeights = MakeMatrix(numInput, numHidden);
      this.hBiases = new double[numHidden];
      this.hOutputs = new double[numHidden];

      this.hoWeights = MakeMatrix(numHidden, numOutput);
      this.oBiases = new double[numOutput];
      this.outputs = new double[numOutput];

      InitializeWeights(); // set weights and biases to small random values
    } // ctor

    private static double[][] MakeMatrix(int rows, int cols) // helper for ctor
    {
      double[][] result = new double[rows][];
      for (int r = 0; r < result.Length; ++r)
        result[r] = new double[cols];
      return result;
    }

    // public override string ToString() . .

    // ----------------------------------------------------------------------------------------

    public void SetWeights(double[] weights)
    {
      // copy weights and biases in weights[] array to i-h weights, i-h biases, h-o weights, h-o biases
      int numWeights = (numInput * numHidden) + (numHidden * numOutput) + numHidden + numOutput;
      if (weights.Length != numWeights)
        throw new Exception("Bad weights array length: ");

      int k = 0; // points into weights param

      for (int i = 0; i < numInput; ++i)
        for (int j = 0; j < numHidden; ++j)
          ihWeights[i][j] = weights[k++];
      for (int i = 0; i < numHidden; ++i)
        hBiases[i] = weights[k++];
      for (int i = 0; i < numHidden; ++i)
        for (int j = 0; j < numOutput; ++j)
          hoWeights[i][j] = weights[k++];
      for (int i = 0; i < numOutput; ++i)
        oBiases[i] = weights[k++];
    }

    private void InitializeWeights()
    {
      // initialize weights and biases to small random values
      int numWeights = (numInput * numHidden) + (numHidden * numOutput) + numHidden + numOutput;
      double[] initialWeights = new double[numWeights];
      double lo = -0.01;
      double hi = 0.01;
      for (int i = 0; i < initialWeights.Length; ++i)
        initialWeights[i] = (hi - lo) * rnd.NextDouble() + lo;
      this.SetWeights(initialWeights);
    }

    public double[] GetWeights()
    {
      // returns the current set of weights, presumably after dropout-training
      int numWeights = (numInput * numHidden) + (numHidden * numOutput) + numHidden + numOutput;
      double[] result = new double[numWeights];
      int k = 0;
      for (int i = 0; i < numInput; ++i)
        for (int j = 0; j < numHidden; ++j)
          result[k++] = ihWeights[i][j];
      for (int i = 0; i < numHidden; ++i)
        result[k++] = hBiases[i];
      for (int i = 0; i < numHidden; ++i)
        for (int j = 0; j < numOutput; ++j)
          result[k++] = hoWeights[i][j];
      for (int i = 0; i < numOutput; ++i)
        result[k++] = oBiases[i];

      return result;
    }

    // ----------------------------------------------------------------------------------------

    private int[] MakeDropNodes()
    {
      List<int> resultList = new List<int>();
      for (int i = 0; i < this.numHidden; ++i)
      {
        double p = rnd.NextDouble();
        if (p < 0.50)
          resultList.Add(i);
      }

      if (resultList.Count == 0)
        resultList.Add(rnd.Next(0, numHidden));
      else if (resultList.Count == numHidden)
        resultList.RemoveAt(rnd.Next(0, numHidden));

      return resultList.ToArray();
    }

    private bool IsDropNode(int node, int[] dropNodes)
    {
      if (dropNodes == null)
        return false;

      if (Array.BinarySearch(dropNodes, node) >= 0)
        return true;
      else
        return false;
    }

    private double[] ComputeOutputs(double[] xValues, int[] dropNodes)
    {
      // skips hidden nodes int dropNodes[]
      // if dropNodes[] is null, no nodes are dropped
      // i = input index, j = hidden, k = output
      if (xValues.Length != numInput)
        throw new Exception("Bad xValues array length");

      double[] hSums = new double[numHidden]; // hidden nodes sums scratch array
      double[] oSums = new double[numOutput]; // output nodes sums

      for (int i = 0; i < xValues.Length; ++i) // copy x-values to inputs
        this.inputs[i] = xValues[i];

      for (int j = 0; j < numHidden; ++j)  // each hidden node
      {
        if (IsDropNode(j, dropNodes) == true) continue; // skip
        for (int i = 0; i < numInput; ++i)
          hSums[j] += this.inputs[i] * this.ihWeights[i][j]; // accumulate sum (note +=)

        hSums[j] += this.hBiases[j]; // add bias

        this.hOutputs[j] = HyperTanFunction(hSums[j]); // apply activation
      }

      for (int k = 0; k < numOutput; ++k)   // each output node
      {
        for (int j = 0; j < numHidden; ++j)
        {
          if (IsDropNode(j, dropNodes) == true) continue; // skip
          oSums[k] += hOutputs[j] * hoWeights[j][k];
        }

        oSums[k] += oBiases[k]; // add bias

        double[] softOut = Softmax(oSums); // softmax activation does all outputs at once for efficiency
        Array.Copy(softOut, outputs, softOut.Length);
      }

      // copy this.outputs to return for calling convenience
      double[] retResult = new double[numOutput]; // could define a GetOutputs method instead
      Array.Copy(this.outputs, retResult, retResult.Length);
      return retResult;

    } // ComputeOutputs

    private static double HyperTanFunction(double x)
    {
      if (x < -20.0) return -1.0; // approximation is correct to 30 decimals
      else if (x > 20.0) return 1.0;
      else return Math.Tanh(x);
    }

    private static double[] Softmax(double[] oSums) // does all output nodes at once so scale doesn't have to be re-computed each time
    {
      // determine max output sum
      double max = oSums[0];
      for (int i = 0; i < oSums.Length; ++i)
        if (oSums[i] > max) max = oSums[i];

      // determine scaling factor -- sum of exp(each val - max)
      double scale = 0.0;
      for (int i = 0; i < oSums.Length; ++i)
        scale += Math.Exp(oSums[i] - max);

      double[] result = new double[oSums.Length];
      for (int i = 0; i < oSums.Length; ++i)
        result[i] = Math.Exp(oSums[i] - max) / scale;

      return result; // now scaled so that xi sum to 1.0
    }

    // ----------------------------------------------------------------------------------------

    private void UpdateWeights(double[] tValues, double learnRate, int[] dropNodes)
    {
      // update the weights and biases using back-propagation
      // assumes that SetWeights and ComputeOutputs have been called
      if (tValues.Length != numOutput)
        throw new Exception("target values not same Length as output in UpdateWeights");

      // back-prop related arrays. could be class members to avoid millions of allocations
      double[] hGrads = new double[numHidden];
      double[] oGrads = new double[numOutput];

      // 1. compute output gradients
      for (int k = 0; k < numOutput; ++k)
      {
        // implicit MSE
        double derivative = (1 - outputs[k]) * outputs[k]; // derivative of softmax = (1 - y) * y (same as log-sigmoid)
        oGrads[k] = derivative * (tValues[k] - outputs[k]); // 'mean squared error version' includes (1-y)(y) derivative
      }

      // 2. compute hidden gradients
      for (int j = 0; j < numHidden; ++j)
      {
        if (IsDropNode(j, dropNodes) == true) continue;
        double derivative = (1 - hOutputs[j]) * (1 + hOutputs[j]); // derivative of tanh = (1 - y) * (1 + y)
        double sum = 0.0;
        for (int k = 0; k < numOutput; ++k) // each hidden delta is the sum of numOutput terms
        {
          double x = oGrads[k] * hoWeights[j][k];
          sum += x;
        }
        hGrads[j] = derivative * sum;
      }

      //// 3a. update input-hidden weights (gradients must be computed right-to-left but weights can be updated in any order)
      //for (int i = 0; i < numInput; ++i) // 0..2 (3)
      //{
      //  for (int j = 0; j < numHidden; ++j) // 0..3 (4)
      //  {
      //    if (IsDropNode(j, dropNodes) == true) continue;
      //    double delta = learnRate * hGrads[j] * inputs[i]; // compute the new delta
      //    ihWeights[i][j] += delta; // update. note we use '+' instead of '-'. this can be very tricky.
      //  }
      //}

      //// 3b. update hidden biases
      //for (int j = 0; j < numHidden; ++j)
      //{
      //  if (IsDropNode(j, dropNodes) == true) continue;
      //  double delta = learnRate * hGrads[j] * 1.0; // the 1.0 is the constant input for any bias; could leave out
      //  hBiases[j] += delta;
      //}

      // 3. update input-hidden weights and hidden biases
      // combined for processing efficiency at expense of clarity
      for (int j = 0; j < numHidden; ++j)
      {
        if (IsDropNode(j, dropNodes) == true) continue;
        for (int i = 0; i < numInput; ++i)
        {
          double delta = learnRate * hGrads[j] * inputs[i]; // compute the new delta
          ihWeights[i][j] += delta; // update. note we use '+' instead of '-'. this can be very tricky.
        }
        double biasDelta = learnRate * hGrads[j] * 1.0; // the 1.0 is the constant input for any bias; could leave out
        hBiases[j] += biasDelta;
      }

      //// 4. update hidden-output weights
      //for (int j = 0; j < numHidden; ++j)
      //{
      //  if (IsDropNode(j, dropNodes) == true) continue;
      //  for (int k = 0; k < numOutput; ++k)
      //  {
      //    double delta = learnRate * oGrads[k] * hOutputs[j];  // see above: hOutputs are inputs to the nn outputs
      //    hoWeights[j][k] += delta;
      //  }
      //}

      //// 4b. update output biases
      //for (int k = 0; k < numOutput; ++k)
      //{
      //  double delta = learnRate * oGrads[k] * 1.0;
      //  oBiases[k] += delta;
      //}

      // 4. update hidden-output weights and output biases
      // combined for processing efficiency at expense of clarity
      for (int k = 0; k < numOutput; ++k)
      {
        for (int j = 0; j < numHidden; ++j)
        {
          if (IsDropNode(j, dropNodes) == true) continue;
          double delta = learnRate * oGrads[k] * hOutputs[j];  // see above: hOutputs are inputs to the nn outputs
          hoWeights[j][k] += delta;
        }
        double biasDelta = learnRate * oGrads[k] * 1.0;
        oBiases[k] += biasDelta;
      }

    } // UpdateWeights

    // ----------------------------------------------------------------------------------------

    public void Train(double[][] trainData, int maxEpochs, double learnRate)
    {
      // train a back-prop style NN classifier using dropout
      // no momentum or weight decay
      int epoch = 0;
      double[] xValues = new double[numInput]; // inputs
      double[] tValues = new double[numOutput]; // target values

      int[] sequence = new int[trainData.Length];
      for (int i = 0; i < sequence.Length; ++i)
        sequence[i] = i;

      while (epoch < maxEpochs)
      {
        // MSE early exit
        //double mse = MeanSquaredError(trainData); // expensive! consider only every k epochs
        //if (mse < 0.010) break; // consider passing value in as parameter

        Shuffle(sequence); // visit each training data in random order
        for (int i = 0; i < trainData.Length; ++i)
        {
          int idx = sequence[i];
          Array.Copy(trainData[idx], xValues, numInput); // more flexible might be a 'GetInputsAndTargets()'
          Array.Copy(trainData[idx], numInput, tValues, 0, numOutput);
          int[] dropNodes = MakeDropNodes();
          ComputeOutputs(xValues, dropNodes); // copy xValues in, compute outputs (and store them internally)
          UpdateWeights(tValues, learnRate, dropNodes);
        } // each training tuple
        ++epoch;
      }

      // divide hidden-output weights by 2.0 to account for dropout
      for (int j = 0; j < numHidden; ++j)
        for (int k = 0; k < numOutput; ++k)
          hoWeights[j][k] /= 2.0;
    } // Train

    private static void Shuffle(int[] sequence)
    {
      for (int i = 0; i < sequence.Length; ++i)
      {
        int r = rnd.Next(i, sequence.Length);
        int tmp = sequence[r];
        sequence[r] = sequence[i];
        sequence[i] = tmp;
      }
    }

    private double MeanSquaredError(double[][] trainData) // used as a training stopping condition
    {
      // average squared error per training tuple
      double sumSquaredError = 0.0;
      double[] xValues = new double[numInput]; // first numInput values in trainData
      double[] tValues = new double[numOutput]; // last numOutput values

      for (int i = 0; i < trainData.Length; ++i) // looks like (6.9 3.2 5.7 2.3) (0 0 1)  (no parens)
      {
        Array.Copy(trainData[i], xValues, numInput); // get xValues. assumes in first columns!
        Array.Copy(trainData[i], numInput, tValues, 0, numOutput); // get target values

        double[] yValues = this.ComputeOutputs(xValues, null); // using current weights (no drop-nodes)
        for (int j = 0; j < numOutput; ++j)
        {
          double err = tValues[j] - yValues[j];
          sumSquaredError += err * err;
        }
      }
      return sumSquaredError / trainData.Length;
    }

    //private double MeanCrossEntropyError(double[][] trainData)
    //{
    //  double sumError = 0.0;
    //  double[] xValues = new double[numInput]; // first numInput values in trainData
    //  double[] tValues = new double[numOutput]; // last numOutput values

    //  for (int i = 0; i < trainData.Length; ++i) // training data: (6.9 3.2 5.7 2.3) (0 0 1) parens not there
    //  {
    //    Array.Copy(trainData[i], xValues, numInput); // get xValues.
    //    Array.Copy(trainData[i], numInput, tValues, 0, numOutput); // get target values
    //    double[] yValues = this.ComputeOutputs(xValues); // compute output using current weights
    //    for (int j = 0; j < numOutput; ++j)
    //    {
    //      sumError += Math.Log(yValues[j]) * tValues[j]; // CE error for one training data
    //    }
    //  }
    //  return -1.0 * sumError / trainData.Length;
    //}

    // ----------------------------------------------------------------------------------------

    public double Accuracy(double[][] testData)
    {
      // percentage correct using winner-takes all
      int numCorrect = 0;
      int numWrong = 0;
      double[] xValues = new double[numInput]; // inputs
      double[] tValues = new double[numOutput]; // targets
      double[] yValues; // computed Y

      for (int i = 0; i < testData.Length; ++i)
      {
        Array.Copy(testData[i], xValues, numInput); // parse test data into x-values and t-values
        Array.Copy(testData[i], numInput, tValues, 0, numOutput);
        yValues = this.ComputeOutputs(xValues, null);  // null == don't use any drop-nodes
        int maxIndex = MaxIndex(yValues); // which cell in yValues has largest value?

        if (tValues[maxIndex] == 1.0) // ugly. consider AreEqual(double x, double y)
          ++numCorrect;
        else
          ++numWrong;
      }
      return (numCorrect * 1.0) / (numCorrect + numWrong); // ugly 2 - check for divide by zero
    }

    private static int MaxIndex(double[] vector) // helper for Accuracy()
    {
      // index of largest value
      int bigIndex = 0;
      double biggestVal = vector[0];
      for (int i = 0; i < vector.Length; ++i)
      {
        if (vector[i] > biggestVal)
        {
          biggestVal = vector[i]; bigIndex = i;
        }
      }
      return bigIndex;
    }

  } // class NeuralNetwork

} // ns

Комментариев нет:

Отправить комментарий