//can the network predict y > x^2?
public static void RunBasic()
{
FeedFowardNetwork n = new FeedFowardNetwork(2, 10, 1)
{
LearningConstant = .3,
totalTrainingSize = 2000,
miniBatchSize = 20
};
Random rng = new Random();
var allInputs = new List<List<double>>();
var allOutputs = new List<List<double>>();
for (int i = 0; i < n.totalTrainingSize; i++)
{
var inputs = new List<double>();
double x = rng.Next(0, 20);
double y = rng.Next(0, 266);
inputs.Add(x);
inputs.Add(y);
allInputs.Add(inputs);
var correctOutputs = new List<double>(new double[] { (Math.Pow(x, 2) < y) ? 1.0 : 0.0 });
allOutputs.Add(correctOutputs);
}
double currentPercentCorrect = 0;
int currentEpoch = 1;
while (currentPercentCorrect < .9)
{
for (int i = 0; i < allInputs.Count; i++)
{
var inputs = allInputs[i];
var correctOutputs = allOutputs[i];
n.bp(n.ff(inputs), correctOutputs);
if (i % n.miniBatchSize == 0 && i != 0)
{
n.averageAndCorrect();
}
}
currentPercentCorrect = RunBasicValidation(n);
currentEpoch++;
}
Console.ReadKey();
Console.ReadKey();
}
public static double RunMNISTValidation(FeedFowardNetwork n)
{
var validationSet = MNISTProc.getImageData(@"..\..\data\t10k-images.idx3-ubyte", @"..\..\data\t10k-labels.idx1-ubyte", false);
double numCorrect = 0;
for (int i = 0; i < 10000; i++)
{
var image = validationSet[i];
var netOutput = n.ff(image.Item1);
var correctOutput = image.Item2;
var netGuess = netOutput.IndexOf(netOutput.Max());
var correctNumber = correctOutput.ToList().IndexOf(1);
if (netGuess.Equals(correctNumber))
{
numCorrect++;
}
}
return numCorrect / 10000;
}
public static double RunBasicValidation(FeedFowardNetwork n)
{
Random rng = new Random();
double percentCorrect = 0;
for (int i = 0; i < 15; i++)
{
var inputs = new List<double>();
double x = rng.Next(0, 20);
double y = rng.Next(0, 266);
inputs.Add(x);
inputs.Add(y);
var correctOutput = (Math.Pow(x, 2) < y) ? 1.0 : 0.0;
var netResponse = n.ff(inputs)[0] > .5 ? 1.0 : 0.0;
if (correctOutput == netResponse)
{
percentCorrect++;
}
}
percentCorrect /= 15;
return percentCorrect;
}
static void Main(string[] args)
{
FeedFowardNetwork n = new FeedFowardNetwork(784, 40, 10)
{
LearningConstant = .0005,
totalTrainingSize = 60000,
miniBatchSize = 5
};
//Training
var dataSet = MNISTProc.getImageData(@"..\..\data\train-images.idx3-ubyte", @"..\..\data\train-labels.idx1-ubyte", true);
double currentPercentCorrect = 0;
int epoch = 1;
while (currentPercentCorrect < 1)
{
for (int inputSet = 0; inputSet < dataSet.Count; inputSet++)
{
var image = dataSet[inputSet];
n.bp(n.ff(image.Item1), new List<double>(image.Item2));
if (inputSet % n.miniBatchSize == 0 && inputSet != 0)
{
n.averageAndCorrect();
}
if (inputSet % 1000 == 0) Console.WriteLine(inputSet);
}
currentPercentCorrect = RunMNISTValidation(n);
Console.WriteLine("Epoch " + epoch + ": " + currentPercentCorrect);
epoch++;
}
Console.ReadKey();
Console.ReadKey();
}
