/// <summary>Creates an instance of the <code>ConvolutionalNN</code> class.</summary>
/// <param name="firstPart">The first part of the convolutional neural network.</param>
/// <param name="fnn">The second part of the convolutional neural network.</param>
/// <param name="createIO">Whether the input image and the output array of the network are to be created.</param>
public ConvolutionalNN(PurelyConvolutionalNN firstPart, FeedForwardNN fnn, bool createIO = true)
{
this.firstPart = firstPart;
this.i2a = new ImageToArray(firstPart.OutputDepth, firstPart.OutputWidth, firstPart.OutputHeight, false);
this.fnn = fnn;
this.errorArray = Backbone.CreateArray <float>(firstPart.OutputDepth * firstPart.OutputWidth * firstPart.OutputHeight);
this.errorImage = new Image(this.firstPart.OutputDepth, this.firstPart.OutputWidth, this.firstPart.OutputHeight);
if (createIO)
{
this.SetInputGetOutput(new Image(firstPart.InputDepth, firstPart.InputWidth, firstPart.InputHeight));
}
this.siameseID = new object();
}
/// <summary>Creates an instance of the <code>DeConvolutionalNN</code> class.</summary>
/// <param name="firstPart">First part of the network.</param>
/// <param name="cnn">Second part of the network.</param>
/// <param name="createIO">Whether the input array and the output image of the netwok are to be created.</param>
public DeConvolutionalNN(FeedForwardNN firstPart, PurelyConvolutionalNN cnn, bool createIO = true)
{
this.firstPart = firstPart;
this.a2i = new ArrayToImage(cnn.InputDepth, cnn.InputWidth, cnn.InputHeight, false);
this.cnn = cnn;
this.errorArray = Backbone.CreateArray <float>(cnn.InputDepth * cnn.InputWidth * cnn.InputHeight);
this.errorImage = new Image(cnn.InputDepth, cnn.InputWidth, cnn.InputHeight);
this.layersConnected = false;
if (createIO)
{
this.SetInputGetOutput(Backbone.CreateArray <float>(firstPart.InputSize));
}
this.siameseID = new object();
}
public PopulationController(FeedForwardNN nn, int PopulationCount)
{
if (PopulationCount < 2)
{
Console.ForegroundColor = ConsoleColor.Red;
Console.WriteLine("Population count must be more than 2");
Console.ForegroundColor = ConsoleColor.White;
}
this.PopulationCount = PopulationCount;
this.descriptor = nn.Descriptor;
Population = new PopulationUnit[PopulationCount];
for (int i = 0; i < PopulationCount; i++)
{
Population[i] = new PopulationUnit(nn);
}
}
/// <summary>Either creates a siamese of the given <code>ConvolutionalNN</code> instance or clones it.</summary>
/// <param name="original">The original instance to be created a siamese of or cloned.</param>
/// <param name="siamese"><code>true</code> if a siamese is to be created, <code>false</code> otherwise.</param>
protected ConvolutionalNN(ConvolutionalNN original, bool siamese)
{
this.errorArray = Backbone.CreateArray <float>(firstPart.OutputDepth * firstPart.OutputWidth * firstPart.OutputHeight);
this.errorImage = new Image(this.firstPart.OutputDepth, this.firstPart.OutputWidth, this.firstPart.OutputHeight);
this.layersConnected = false;
if (siamese)
{
this.firstPart = (PurelyConvolutionalNN)original.CreateSiamese();
this.i2a = (ImageToArray)original.i2a.CreateSiamese();
this.fnn = (FeedForwardNN)original.fnn.CreateSiamese();
this.siameseID = original.SiameseID;
}
else
{
this.firstPart = (PurelyConvolutionalNN)original.firstPart.Clone();
this.i2a = (ImageToArray)original.i2a.Clone();
this.fnn = (FeedForwardNN)original.fnn.Clone();
this.siameseID = new object();
}
}
/// <summary>Either creates a siamese or clones the given <code>DeConvolutionalNN</code> instance.</summary>
/// <param name="original">The original instance to be created a siamese of or cloned.</param>
/// <param name="siamese"><code>true</code> if a siamese is to be created, <code>false</code> if a clone is.</param>
protected DeConvolutionalNN(DeConvolutionalNN original, bool siamese)
{
this.errorArray = Backbone.CreateArray <float>(original.cnn.InputDepth * original.cnn.InputWidth * original.cnn.InputHeight);
this.errorImage = new Image(original.cnn.InputDepth, original.cnn.InputWidth, original.cnn.InputHeight);
this.layersConnected = false;
if (siamese)
{
this.firstPart = (FeedForwardNN)original.firstPart.CreateSiamese();
this.a2i = (ArrayToImage)original.a2i.CreateSiamese();
this.cnn = (PurelyConvolutionalNN)original.cnn.CreateSiamese();
this.siameseID = original.SiameseID;
}
else
{
this.firstPart = (FeedForwardNN)original.firstPart.Clone();
this.a2i = (ArrayToImage)original.a2i.Clone();
this.cnn = (PurelyConvolutionalNN)original.cnn.Clone();
this.siameseID = new object();
}
}
static void Main(string[] args)
{
FeedForwardNN nn = new FeedForwardNN(new FeedForwardNNDescriptor());
nn.ReadWeights("mnist.xml");
float[][] testData = new float[10000][];
float[][] testAnswers = new float[10000][];
int index = 0;
using (StreamReader sr = new StreamReader(@"C:\Users\Dima\source\repos\NeuralNet1\NeuralNetRun\bin\Debug\mnist_test.csv", System.Text.Encoding.Default))
{
string line;
while ((line = sr.ReadLine()) != null)
{
string[] array = line.Split(',');
float[] data = new float[784];
int label = Convert.ToInt32(array[0]);
for (int i = 0; i < 784; i++)
{
data[i] = Normalize.Minimax(Convert.ToInt32(array[i + 1]), 0, 255);
}
float[] answer = new float[10];
answer[label] = 1;
testData[index] = data;
testAnswers[index] = answer;
index++;
}
}
Trainer trainer = new Trainer(ref nn);
trainer.TrainBackPropogation(1, 1, 0.0001f, 0.3f, testData, new float[][] { }, testAnswers, new float[][] { });
float[][] o = new float[3][];
o[0] = nn.Run(testData[0]);
o[1] = nn.Run(testData[1]);
o[2] = nn.Run(testData[2]);
int answer0;
int answer1;
int answer2;
float max = 0;
for (int i = 0; i < o[0].Length; i++)
{
if (o[0][i] > max)
{
answer0 = i;
max = o[0][i];
}
}
max = 0;
for (int i = 0; i < o[0].Length; i++)
{
if (o[1][i] > max)
{
answer1 = i;
max = o[1][i];
}
}
max = 0;
for (int i = 0; i < o[0].Length; i++)
{
if (o[2][i] > max)
{
answer2 = i;
max = o[2][i];
}
}
}
public PopulationUnit(FeedForwardNN NN)
{
this.NN = NN;
Outputs = new float[NN.Descriptor.LayersData[NN.Descriptor.LayersData.Length - 1]];
}
public Trainer(ref FeedForwardNN NN)
{
NeuralNet = NN;
}
