public NetTrainer(IModelService modelService)
{
Network net = new Network();
net.AddLayer(new BasicLayer(new ActivationTANH(), true, DefaultFingerprintSize));
net.AddLayer(new BasicLayer(new ActivationTANH(), true, DefaultHiddenNeuronsCount));
net.AddLayer(new BasicLayer(new ActivationTANH(), false, OutPutNeurons));
net.Structure.FinalizeStructure();
net.Reset();
this.modelService = modelService;
pauseSem = new Semaphore(0, 1, "PauseSemaphore");
}
public void BuildNetwork()
{
_network = new Network(_node);
_network.AddLayer(4); //Hidden layer with 2 neurons
_network.AddLayer(1); //Output layer with 1 neuron
_network.BindInputLayer(_input); //Bind Input Data
_network.BindTraining(_desired); //Bind desired output data
_network.AutoLinkFeedforward(); //Create synapses between the layers for typical feedforward networks.
}
public void AddLayers()
{
Network.AddLayer(InputNeuronCount, "Input"); // input layer
for (int i = 0; i < HiddenLayerCount; ++i)
{
Network.AddLayer(HiddenNeuronCount, String.Format("Hidden{0}", i)); // hidden layers
}
Network.AddLayer(OutputNeuronCount, "Output"); // output layer
}
private void финСетьToolStripMenuItem_Click(object sender, EventArgs e)
{
Network network = new Network();
network.AddLayer(new FullyConnLayar(new Relu(), new NeuralNetwork.Size(1, 1, 400)));
network.AddLayer(new MaxPool2D(new SimpleFunc(), 2, 1));
network.AddLayer(new FullyConnLayar(new Relu(), new NeuralNetwork.Size(1, 1, 100)));
network.AddLayer(new MaxPool2D(new SimpleFunc(), 2, 1));
network.AddLayer(new FullyConnLayar(new Relu(), new NeuralNetwork.Size(1, 1, 4)));
network.Compile(new NeuralNetwork.Size(10, 1, 4));
//network.Normalization();
//network.Normalization();
NetworkData.network = network;
WriteNetwork();
}
/// <summary>
/// Connect the network.
/// </summary>
public virtual void ConnectNetwork()
{
Layer input = _network.InitUnboundInputLayer(_numberOfInputs);
Layer hidden;
if (_config.All2AllEnable.Value)
{
if (_config.HorizontalLinesEnable.Value && _config.VerticalLinesEnable.Value)
{
hidden = _network.AddLayer((int)Math.Ceiling(_numberOfInputs / 4d));
}
else
{
hidden = _network.AddLayer((int)Math.Ceiling(_numberOfInputs / 2d));
}
input.CrossLinkForward(); //link input to existing hidden neurons
}
else
{
hidden = _network.AddLayer(0);
}
_network.AddLayer(_numberOfOutputs);
if (_config.VerticalLinesEnable.Value)
{
hidden.AddRange(AddVerticalLines(input, hidden));
}
if (_config.HorizontalLinesEnable.Value)
{
hidden.AddRange(AddHorizontalLines(input, hidden));
}
if (_config.RingsEnable.Value)
{
hidden.AddRange(AddRings(input, hidden));
}
if (_config.LittleSquaresEnable.Value)
{
hidden.AddRange(AddSquares(input, hidden));
}
hidden.CrossLinkForward(); //link hidden to output layer
}
static bool CatTrain()
{
BitmapCatEnumerator enums = new BitmapCatEnumerator("Sorted", new System.Drawing.Size(50, 25));
Network network = new Network();
network.AddLayer(new Conv2D(new Relu(), 7, 7, 32));
network.AddLayer(new MaxPool2D(new Relu(), 2, 2));
network.AddLayer(new Conv2D(new Relu(), 5, 5, 64));
network.AddLayer(new MaxPool2D(new Relu(), 2, 2));
network.AddLayer(new FullyConnLayar(new Relu(), new Size(1, 1, 256)));
network.AddLayer(new FullyConnLayar(new Sigmoid(), new Size(1, 1, 2)));
network.Compile(new Size(3, 25, 50), true);
network.Normalization();
var pair = enums.GetRandom(ref network);
OneEnumerator one = new OneEnumerator();
one.input = pair.Key;
one.output = pair.Value;
MomentumParallel sgd = new MomentumParallel(network, 0.9, 1e-6);
double[] errors = sgd.TrainBatch(enums, 32, 1000);
return(errors[0] > errors.Last());
}
static bool SaveLoad()
{
BitmapCatEnumerator enums = new BitmapCatEnumerator("Sorted", new System.Drawing.Size(50, 25));
Network network = new Network();
network.AddLayer(new Conv2D(new Relu(), 7, 7, 32));
network.AddLayer(new MaxPool2D(new Relu(), 2, 2));
network.AddLayer(new Conv2D(new Relu(), 5, 5, 64));
network.AddLayer(new MaxPool2D(new Relu(), 2, 2));
network.AddLayer(new FullyConnLayar(new Relu(), new Size(1, 1, 256)));
network.AddLayer(new FullyConnLayar(new Sigmoid(), new Size(1, 1, 2)));
network.Compile(new Size(3, 25, 50), true);
network.Normalization();
network.Normalization();
string data = network.SaveJSON();
Network network2 = new Network();
network2.LoadJSON(data);
var pair = enums.GetRandom(ref network);
var out1 = network.GetOutput(pair.Key);
var out2 = network.GetOutput(pair.Key);
return(out1[0, 0, 0] == out2[0, 0, 0]);
}
public MainWindow()
{
InitializeComponent();
TestInput Input = new TestInput(48, 48, 1);
Network net = new Network();
net.AddLayer(new Conv(190, 5, 5, 48, 48, true));
net.AddLayer(new Conv(260, 5, 5, 44, 44, true));
net.Process(Input.Value);
long size = 0;
using (Stream stream = new MemoryStream())
{
BinaryFormatter formatter = new BinaryFormatter();
formatter.Serialize(stream, (object)net);
size = stream.Length;
}
Conv conv1_44x44 = new Conv();
Hidden fully_160 = new Hidden(160, 16, true);
Pool pool = new Pool();
conv1_44x44.AddMap(new Map(new Kernel(5, 5), 48, 48));
conv1_44x44.InputValues(Input.Value);
pool.InputMaps(conv1_44x44.Maps);
var res = 8;
}
static bool ErrorTest()
{
Network network = new Network();
network.AddLayer(new FullyConnLayar(new Relu(), new Size(1, 1, 1)));
network.Compile(new Size(1, 1, 1));
double[,,] input = new double[1, 1, 1];
input[0, 0, 0] = 1;
double[,,] t = new double[1, 1, 1];
t[0, 0, 0] = 1;
double val = network.GetError(input, t);
return(val == 1);
}
public void XORNoBiasTest()
{
double[][] inputs = new double[][] { new double[] { -1, -1 }, new double[] { -1, 1 }, new double[] { 1, -1 }, new double[] { 1, 1 } };
double[] expectedOutput = new double[] { -1, 1, 1, -1 };
var network = new Network(
new IdentityLayer(2),
new DenseLayerNoBias(2, 1, new IdentityActivation(), new SquaredDistance()));
network.AddLayer(new DenseLayerNoBias(2, 2, new IdentityActivation(), new SquaredDistance()));
network.Initialize();
for (int i = 0; i < inputs.Length; i++)
{
network.Evaluate(inputs[i]);
Assert.AreNotEqual(expectedOutput[i], network.OutputLayer.Output[0]);
}
// Train network
int epoc = 0;
double error = 100;
while (++epoc < 1000 && error > 0.001)
{
error = 0;
for (int i = 0; i < inputs.Length; i++)
{
error += network.Train(inputs[i], new double[] { expectedOutput[i] }, 0.1);
}
}
//Assert.IsTrue(epoc < 1000);
//for (int i = 0; i < inputs.Length; i++)
//{
// layer.Evaluate(inputs[i]);
// Assert.IsTrue(Math.Abs(expectedOutput[i] - layer.Output[0]) < 0.01); // 1% error margin
//}
}
static bool PreTrain()
{
BitmapCatEnumerator enums = new BitmapCatEnumerator("Sorted", new System.Drawing.Size(24, 12));
BitmapCatEnumerator val = new BitmapCatEnumerator("Val", new System.Drawing.Size(24, 12));
Network network = new Network();
//network.LoadJSON(System.IO.File.ReadAllText("pretrained_2.neural"));
//network.CompileOnlyError();
network.AddLayer(new Conv2D(new Relu(), 3, 3, 10));
network.AddLayer(new MaxPool2D(new Relu(), 2, 2));
network.AddLayer(new Conv2D(new Relu(), 5, 5, 30));
network.AddLayer(new MaxPool2D(new Relu(), 2, 2));
network.AddLayer(new FullyConnLayar(new Relu(), new Size(1, 1, 256)));
network.AddLayer(new FullyConnLayar(new Sigmoid(), new Size(1, 1, 2)));
network.Compile(new Size(3, 12, 24), true);
network.Normalization();
network.Normalization();
MomentumParallel sgd = new MomentumParallel(network, 0.9, 1e-4);
sgd.need_max = false;
var pair = PretrainAutoEncoder.Action(network, sgd, enums, val, 2000, 32);
Console.WriteLine("{0}\n{1}", pair.Key, pair.Value);
Console.WriteLine("Start train");
sgd = new MomentumParallel(network, 0, 1e-5);
DateTime start = DateTime.Now;
sgd.TrainBatch(enums, 256, 1);
for (int i = 0; i < 100000; i++)
{
double[] errors = sgd.TrainBatchContinue(enums, 256, 1);
if ((DateTime.Now - start).TotalMinutes > 5)
{
System.IO.File.WriteAllText("train_" + i + ".neural", network.SaveJSON());
start = DateTime.Now;
Console.WriteLine("Saved at " + "train_" + i + ".neural");
}
}
return(true);
}
static void Main(string[] args)
{
Controler ffc = new Controler(1);
Network net = new Network(ffc.Node);
net.InitUnboundInputLayer(1).BasicConfiguration.ActivationType.Value = EActivationType.Linear;
net.AddLayer(1, EActivationType.Linear);
net.AutoLinkFeedforward();
Neuron neuron = net.LastLayer[0];
Synapse synapse = neuron.SourceSynapses[0];
ffc.ImportNetwork(net, false);
BasicConfig config = ffc.NeuralNetwork.BasicConfiguration;
config.BiasNeuronEnable.Value = true;
config.BiasNeuronOutput.Value = 1.0;
config.FlatspotEliminationEnable.Value = false;
config.WeightDecayEnable.Value = false;
config.SymmetryPreventionEnable.Value = false;
config.ManhattanTrainingEnable.Value = false;
config.LearningRate.Value = 0.005;
StochasticCoordinateGenerator scg = new StochasticCoordinateGenerator(0, 10, 100);
//RegularCoordinateGenerator rcg = new RegularCoordinateGenerator(-25, 25, 50);
DynamicSampleProvider dsp = new DynamicSampleProvider(my_func, scg); //rcg);
ffc.Provider = dsp; // new CachedSampleProvider(dsp);
Console.WriteLine("TARGET FUNCTION: 3*x+5");
Console.WriteLine("TARGET Synapse Weight = 3.0");
Console.WriteLine("TARGET Bias Weight = 5.0");
Console.WriteLine("TARGET Mean Squared Error <= 0.000000001");
Console.WriteLine();
Console.WriteLine("Synapse Weight: " + synapse.Weight + " - Bias Weight: " + neuron.BiasNeuronWeight);
Console.WriteLine("Initial MSE: " + ffc.EstimateMeanSquaredError());
Console.WriteLine();
ffc.TrainAllSamplesOnce();
Console.WriteLine("Synapse Weight: " + synapse.Weight + " - Bias Weight: " + neuron.BiasNeuronWeight);
Console.WriteLine("Trained MSE: " + ffc.EstimateMeanSquaredError());
Console.WriteLine();
ffc.TrainAllSamplesOnce();
Console.WriteLine("Synapse Weight: " + synapse.Weight + " - Bias Weight: " + neuron.BiasNeuronWeight);
Console.WriteLine("Trained MSE: " + ffc.EstimateMeanSquaredError());
Console.WriteLine();
ffc.TrainAllSamplesOnce();
Console.WriteLine("Synapse Weight: " + synapse.Weight + " - Bias Weight: " + neuron.BiasNeuronWeight);
Console.WriteLine("Trained MSE: " + ffc.EstimateMeanSquaredError());
Console.WriteLine();
Console.WriteLine("Auto Training, maximum 1000 Epochs");
Console.WriteLine();
if (ffc.TrainAllSamplesUntil(0.000000001, 1000))
{
Console.WriteLine("Synapse Weight: " + synapse.Weight + " - Bias Weight: " + neuron.BiasNeuronWeight);
Console.ForegroundColor = ConsoleColor.Green;
Console.WriteLine("SUCCEEDS auto training with MSE: " + ffc.EstimateMeanSquaredError());
Console.ResetColor();
}
else
{
Console.WriteLine("Synapse Weight: " + synapse.Weight + " - Bias Weight: " + neuron.BiasNeuronWeight);
Console.ForegroundColor = ConsoleColor.Red;
Console.WriteLine("FAILS auto training with MSE: " + ffc.EstimateMeanSquaredError());
Console.ResetColor();
}
Console.ReadKey();
}
public void RunDemo()
{
Console.WriteLine("### BASIC UNBOUND DEMO ###");
//Initialize the network manager.
//This constructor also creates the first
//network layer (Inputlayer).
Network network = new Network();
//You need to initialize (the size of) the
//input layer in an unbound scenario
network.InitUnboundInputLayer(3);
//Add the hidden layer with 4 neurons.
network.AddLayer(4);
//Add the output layer with 2 neurons.
network.AddLayer(2);
//Connect the neurons together using synapses.
//This is the easiest way to do it; I'll discuss
//other ways in more detail in another demo.
network.AutoLinkFeedforward();
//Push new input data
network.PushUnboundInput(new bool[] {false,true,false});
//... and output training data ...
network.PushUnboundTraining(new bool[] {false,true});
//Propagate the network using the bound input data.
//Internally, this is a two round process, to
//correctly handle feedbacks
network.CalculateFeedforward();
//Collect the network output and print it.
App.PrintArray(network.CollectOutput());
//Train the current pattern using Backpropagation (one step)!
network.TrainCurrentPattern(false,true);
//Print the output; the difference to (-1,1) should be
//smaller this time!
App.PrintArray(network.CollectOutput());
//Same one more time:
network.TrainCurrentPattern(false,true);
App.PrintArray(network.CollectOutput());
//Train another pattern:
Console.WriteLine("# new pattern:");
//this time we're using doubles directly, instead of booleans.
//5/1 are the default values for input/training values.
network.PushUnboundInput(new double[] {5d,-5d,-5d});
network.PushUnboundTraining(new double[] {1,1});
//calculate ...
network.CalculateFeedforward();
App.PrintArray(network.CollectOutput());
//... and train it one time
network.TrainCurrentPattern(false,true);
App.PrintArray(network.CollectOutput());
//what about the old pattern now?
Console.WriteLine("# the old pattern again:");
network.PushUnboundInput(new double[] {-5d,5d,-5d});
network.PushUnboundTraining(new double[] {-1,1});
network.CalculateFeedforward();
App.PrintArray(network.CollectOutput());
Console.WriteLine("=== COMPLETE ===");
Console.WriteLine();
}
private Network initializeNetwork()
{
// layer 1
Node a1 = new Node("a1");
Node a2 = new Node("a2");
Node a3 = new Node("a3");
Layer layer1 = new Layer();
layer1.AddNode(a1);
layer1.AddNode(a2);
layer1.AddNode(a3);
// layer 2
Node b1 = new Node("b1");
Node b2 = new Node("b2");
Layer layer2 = new Layer();
layer2.AddNode(b1);
layer2.AddNode(b2);
// layer 3
Node c1 = new Node("c1");
Node c2 = new Node("c2");
Layer layer3 = new Layer();
layer3.AddNode(c1);
layer3.AddNode(c2);
// Connectors layer1 -> layer2
Connector con_a1_b1 = new Connector(a1, b1, 0.5f);
a1.AddForwardConnector(con_a1_b1);
b2.AddBackwardConnector(con_a1_b1);
Connector con_a2_b1 = new Connector(a2, b1, 0.5f);
Connector con_a2_b2 = new Connector(a2, b2, 0.5f);
a2.AddForwardConnector(con_a2_b1);
b1.AddBackwardConnector(con_a2_b1);
a2.AddForwardConnector(con_a2_b2);
b2.AddBackwardConnector(con_a2_b2);
Connector con_a3_b2 = new Connector(a3, b2, 0.5f);
a3.AddForwardConnector(con_a3_b2);
b2.AddBackwardConnector(con_a3_b2);
// Connectors layer2 -> layer3
Connector con_b1_c1 = new Connector(b1, c1, 0.5f);
Connector con_b1_c2 = new Connector(b1, c2, 0.5f);
b1.AddForwardConnector(con_b1_c1);
c1.AddBackwardConnector(con_b1_c1);
b1.AddForwardConnector(con_b1_c2);
c2.AddBackwardConnector(con_b1_c2);
Connector con_b2_c1 = new Connector(b2, c1, 0.5f);
Connector con_b2_c2 = new Connector(b2, c2, 0.5f);
b2.AddForwardConnector(con_b2_c1);
c1.AddBackwardConnector(con_b2_c1);
b2.AddForwardConnector(con_b2_c2);
c2.AddBackwardConnector(con_b2_c2);
// form the network
Network network = new Network();
network.AddLayer(layer1);
network.AddLayer(layer2);
return(network);
}
public void RunDemo()
{
Console.WriteLine("### NETWORK STRUCTURE DEMO ###");
//Initialize the network manager.
//This constructor also creates the first
//network layer (Inputlayer).
Network network = new Network();
//You need to initialize (the size of) the
//input layer in an unbound scenario
network.InitUnboundInputLayer(3);
//Add the hidden layer with 4 neurons.
network.AddLayer(4);
//Add the output layer with 2 neurons.
network.AddLayer(2);
//Instead of calling AutoLinkFeedforward()
//on this place, in this demo we'll connect
//the network together by our own!
Layer input = network.FirstLayer;
Layer hidden = input.TargetLayer;
Layer output = network.LastLayer;
//First we want to connect all neurons
//of the hidden layer to all neurons
//of the input layer (that's exactly
//what the AutoLinkFeedforward would
//do - but between all layers).
input.CrossLinkForward();
//Then we want to achieve a lateral
//feedback in the hidden layer
//(AutoLinkFeedforward does NOT do this):
hidden.CrossLinkLayer();
//Next we want to connect the first
//and the second Neuron of the hidden
//Layer to the first output neuron,
//and the third and fourth to the 2nd
//output neuron. Some of the synapses
//shall start with special weights:
hidden[0].ConnectToNeuron(output[0]);
hidden[1].ConnectToNeuron(output[0], 0.5);
hidden[2].ConnectToNeuron(output[1]);
hidden[3].ConnectToNeuron(output[1], -1.5);
//That's it. Now we can work with it,
//just we did on the Basic Unbound Demo:
network.PushUnboundInput(new bool[] { false, true, false });
network.PushUnboundTraining(new bool[] { false, true });
network.CalculateFeedforward();
App.PrintArray(network.CollectOutput());
network.TrainCurrentPattern(false, true);
App.PrintArray(network.CollectOutput());
network.TrainCurrentPattern(false, true);
App.PrintArray(network.CollectOutput());
//This demo may help you e.g. building your own
//INetworkStructureFactory implementations
//for the grid pattern matching building block.
//(You may also want to check out the default
//implementation!)
Console.WriteLine("=== COMPLETE ===");
Console.WriteLine();
}
private void загрузитьСтандартнуюМодельToolStripMenuItem_Click(object sender, EventArgs e)
{
Network network = new Network();
network.AddLayer(new Dropout(new SimpleFunc(), 0.1));
network.AddLayer(new Conv2D(new Relu(), 9, 9, 16));
network.AddLayer(new MaxPool2D(new SimpleFunc(), 2, 2));
network.AddLayer(new Conv2D(new Relu(), 5, 5, 32));
network.AddLayer(new MaxPool2D(new SimpleFunc(), 2, 2));
network.AddLayer(new Conv2D(new Relu(), 3, 3, 64));
network.AddLayer(new MaxPool2D(new SimpleFunc(), 2, 2));
network.AddLayer(new Conv2D(new Relu(), 3, 3, 32));
network.AddLayer(new MaxPool2D(new SimpleFunc(), 2, 2));
network.AddLayer(new FullyConnLayar(new Relu(), new NeuralNetwork.Size(1, 1, 128)));
var pair = NetworkData.train.GetRandom(ref network);
var out_arr = pair.Value;
network.AddLayer(new FullyConnLayar(new Sigmoid(), new NeuralNetwork.Size(out_arr.GetLength(0), out_arr.GetLength(1), out_arr.GetLength(2))));
network.Compile(new NeuralNetwork.Size(3, NetworkData.image_size.Height, NetworkData.image_size.Width), true);
network.Normalization();
network.Normalization();
NetworkData.network = network;
string message = "";
for (int i = 0; i < network.layers.Count; i++)
{
message += network.layers[i].GetType().Name + " {" + network.layers[i].output_size[0] + "," + network.layers[i].output_size[1] + "," + network.layers[i].output_size[2] + "}\n";
}
WriteNetwork();
MessageBox.Show(message);
}
private void button1_Click(object sender, EventArgs e)
{
if (openFileDialog1.ShowDialog() == DialogResult.OK)
{
Bitmap map = new Bitmap(openFileDialog1.FileName);
BoxerDLL.BoxCut2D cutter = new BoxCut2D(7, 10, 0.143f, 0.1f);
List <Bitmap> maps = cutter.Cut(map);
List <double[, , ]> rgbs = new List <double[, , ]>();
for (int i = 0; i < maps.Count; i++)
{
maps[i] = new Bitmap(maps[i], new System.Drawing.Size(24, 12));
rgbs.Add(ImageDataConverter.GetRGBArray(maps[i]));
}
RectangleF[] rects = cutter.Rects1D(map.Width, map.Height);
BitmapCatEnumerator enums = new BitmapCatEnumerator("Sorted", new System.Drawing.Size(24, 12));
BitmapCatEnumerator val = new BitmapCatEnumerator("Val", new System.Drawing.Size(24, 12));
Network network = new Network();
network.AddLayer(new Dropout(new Relu(), 0.05));
network.AddLayer(new Conv2D(new Relu(), 7, 7, 32));
network.AddLayer(new MaxPool2D(new Relu(), 2, 2));
network.AddLayer(new Dropout(new Relu(), 0.05));
network.AddLayer(new Conv2D(new Relu(), 3, 3, 64));
network.AddLayer(new MaxPool2D(new Relu(), 2, 2));
network.AddLayer(new Dropout(new Relu(), 0.05));
network.AddLayer(new FullyConnLayar(new Sigmoid(), new NeuralNetwork.Size(1, 1, 256)));
network.AddLayer(new FullyConnLayar(new Sigmoid(), new NeuralNetwork.Size(1, 1, 2)));
network.Compile(new NeuralNetwork.Size(3, 12, 24), true);
network.Normalization();
pictureBox1.Image = map;
MomentumParallel sgd = new MomentumParallel(network, 0.9, 1e-4);
int last = 0;
for (int i = 0; i < 1000; i++)
{
var errors = sgd.TrainBatchPercent(enums, 32, 1);
if (errors.Last().Value > 90 || i - last > 20)
{
last = i;
Color[] colors = new Color[rects.Length];
Color plus = Color.FromArgb(100, Color.Red);
Color minus = Color.FromArgb(100, Color.Green);
for (int j = 0; j < maps.Count; j++)
{
double[,,] output = network.GetOutput(rgbs[j]);
colors[j] = output[0, 0, 0] > output[0, 0, 1] ? plus : minus;
}
pictureBox1.Image = DrawColoredRects(map, rects, colors);
var validation = network.GetError(val);
label2.Text = "Validation\nError: " + validation.Key + "\nPercent: " + validation.Value;
}
label1.Text = i + "\nError: " + errors.Last().Key + "\nPercent: " + errors.Last().Value;
Update();
}
}
}
static void LinePredictor()
{
StreamReader rd = new StreamReader("data.txt");
string classes_line = rd.ReadLine();
string[] classes = classes_line.Split(' ');
List <double[, , ]> inputs = new List <double[, , ]>();
List <double [, , ]> outputs = new List <double[, , ]>();
while (!rd.EndOfStream)
{
inputs.Add(LineToArr3(rd.ReadLine()));
outputs.Add(LineToArr3(rd.ReadLine()));
}
int output_len = outputs[0].Length;
int nancount = 0;
for (int i = 0; i < inputs.Count; i++)
{
var l = inputs[i];
for (int j = 0; j < l.Length; j++)
{
if (double.IsNaN(l[0, 0, j]) || double.IsInfinity(l[0, 0, j]))
{
nancount++;
inputs.RemoveAt(i);
outputs.RemoveAt(i);
i--;
break;
}
}
}
ArrDataEnumerator enumerator = new ArrDataEnumerator(inputs, outputs);
DataCategoryVisualisation visualisation = new DataCategoryVisualisation(enumerator, 2, 6, 600, classes);
visualisation.ShowDialog();
Network network = new Network();
network.loss = new MaxLoss();
network.AddLayer(new FullyConnLayar(new LeackyRelu(0.01), new NeuralNetwork.Size(1, 1, 64)));
network.AddLayer(new FullyConnLayar(new Sigmoid(), new NeuralNetwork.Size(1, 1, output_len)));
network.Compile(new NeuralNetwork.Size(1, 1, inputs[0].GetLength(2)), true);
network.AddNoise();
Optimizer parallel = new MomentumParallel(network, 0.9, 0.0005);
enumerator.SplitTrainVal(0.8, out ArrDataEnumerator train, out ArrDataEnumerator val);
Learner learner = new Learner()
{
optimizer = parallel,
train_data = train,
val_data = val
};
learner.metrics.Add(new ArgMaxMetrics());
learner.metrics.Add(new ArgmaxCount(0, 0, 4));
learner.metrics.Add(new ArgmaxTrainCount(0, 0, 4));
learner.callbacks.Add(new MetricSaveCallback("val_argmax"));
learner.Learn(8, 50, 10000);
}
public void RunDemo()
{
Console.WriteLine("### NETWORK STRUCTURE DEMO ###");
//Initialize the network manager.
//This constructor also creates the first
//network layer (Inputlayer).
Network network = new Network();
//You need to initialize (the size of) the
//input layer in an unbound scenario
network.InitUnboundInputLayer(3);
//Add the hidden layer with 4 neurons.
network.AddLayer(4);
//Add the output layer with 2 neurons.
network.AddLayer(2);
//Instead of calling AutoLinkFeedforward()
//on this place, in this demo we'll connect
//the network together by our own!
Layer input = network.FirstLayer;
Layer hidden = input.TargetLayer;
Layer output = network.LastLayer;
//First we want to connect all neurons
//of the hidden layer to all neurons
//of the input layer (that's exactly
//what the AutoLinkFeedforward would
//do - but between all layers).
input.CrossLinkForward();
//Then we want to achieve a lateral
//feedback in the hidden layer
//(AutoLinkFeedforward does NOT do this):
hidden.CrossLinkLayer();
//Next we want to connect the first
//and the second Neuron of the hidden
//Layer to the first output neuron,
//and the third and fourth to the 2nd
//output neuron. Some of the synapses
//shall start with special weights:
hidden[0].ConnectToNeuron(output[0]);
hidden[1].ConnectToNeuron(output[0],0.5);
hidden[2].ConnectToNeuron(output[1]);
hidden[3].ConnectToNeuron(output[1],-1.5);
//That's it. Now we can work with it,
//just we did on the Basic Unbound Demo:
network.PushUnboundInput(new bool[] {false,true,false});
network.PushUnboundTraining(new bool[] {false,true});
network.CalculateFeedforward();
App.PrintArray(network.CollectOutput());
network.TrainCurrentPattern(false,true);
App.PrintArray(network.CollectOutput());
network.TrainCurrentPattern(false,true);
App.PrintArray(network.CollectOutput());
//This demo may help you e.g. building your own
//INetworkStructureFactory implementations
//for the grid pattern matching building block.
//(You may also want to check out the default
//implementation!)
Console.WriteLine("=== COMPLETE ===");
Console.WriteLine();
}
public void RunDemo()
{
Console.WriteLine("### BASIC BOUND DEMO ###");
//Prepare you're input and training data
//to bind to the network
double[] input = new double[] {-5d,5d,-5d};
double[] training = new double[] {-1,1};
//Initialize the network manager.
//This constructor also creates the first
//network layer (Inputlayer).
Network network = new Network();
//Bind your input array (to the already
//existing input layer)
network.BindInputLayer(input);
//Add the hidden layer with 4 neurons.
network.AddLayer(4);
//Add the output layer with 2 neurons.
network.AddLayer(2);
//bind your training array to the output layer.
//Always do this AFTER creating the layers.
network.BindTraining(training);
//Connect the neurons together using synapses.
//This is the easiest way to do it; I'll discuss
//other ways in more detail in another demo.
network.AutoLinkFeedforward();
//Propagate the network using the bound input data.
//Internally, this is a two round process, to
//correctly handle feedbacks
network.CalculateFeedforward();
//Collect the network output and print it.
App.PrintArray(network.CollectOutput());
//Train the current pattern using Backpropagation (one step)!
network.TrainCurrentPattern(false,true);
//Print the output; the difference to (-1,1) should be
//smaller this time!
App.PrintArray(network.CollectOutput());
//Same one more time:
network.TrainCurrentPattern(false,true);
App.PrintArray(network.CollectOutput());
//Train another pattern:
Console.WriteLine("# new pattern:");
input[0] = 5d;
input[1] = -5d;
training[0] = 1;
//calculate ...
network.CalculateFeedforward();
App.PrintArray(network.CollectOutput());
//... and train it one time
network.TrainCurrentPattern(false,true);
App.PrintArray(network.CollectOutput());
//what about the old pattern now?
Console.WriteLine("# the old pattern again:");
input[0] = -5d;
input[1] = 5d;
training[0] = -1;
network.CalculateFeedforward();
App.PrintArray(network.CollectOutput());
Console.WriteLine("=== COMPLETE ===");
Console.WriteLine();
}
private void RebuildCurrentNetworkStructure()
{
currentNetwork = new Network();
currentNeuronMap = new Hashtable();
currentSynapseMap = new Hashtable();
//Build Layers and Neurons
NeuralDataSet.LayersRow[] layerRows = SelectLayersFromNetwork(currentNetworkRow);
if(layerRows.Length == 0)
return;
NeuralDataSet.NeuronsRow[] neuronRows = SelectNeuronsFromLayer(layerRows[0]);
currentNetwork.InitUnboundInputLayer(neuronRows.Length);
Layer[] layers = new Layer[layerRows.Length];
layers[0] = currentNetwork.FirstLayer;
AppendNeuronsToNeuronMap(currentNeuronMap,neuronRows,layers[0]);
for(int i=1;i<layerRows.Length;i++)
{
neuronRows = SelectNeuronsFromLayer(layerRows[i]);
layers[i] = currentNetwork.AddLayer(neuronRows.Length);
AppendNeuronsToNeuronMap(currentNeuronMap,neuronRows,layers[i]);
}
//Build Synapses
NeuralDataSet.SynapsesRow[] synapseRows = SelectSynapsesFromNetwork(currentNetworkRow);
for(int i=0;i<synapseRows.Length;i++)
currentSynapseMap.Add(synapseRows[i].syID,
((Neuron)currentNeuronMap[synapseRows[i].syFK_neuronSource]).ConnectToNeuron((Neuron)currentNeuronMap[synapseRows[i].syFK_neuronTarget]));
}
public void BuildNetwork()
{
network = new Network(node);
if(!groupedHiddenLayer)
{
network.AddLayer(32); //Hidden layer with 32 neurons
network.AddLayer(16); //Output layer with 16 neuron
network.BindInputLayer(input); //Bind Input Data
network.PushUnboundTraining(outFF);
network.AutoLinkFeedforward(); //Create synapses between the layers for typical feedforward networks.
}
else
{
network.AddLayer(64); //Hidden layer with 64 neurons
network.AddLayer(16); //Output layer with 16 neuron
network.BindInputLayer(input); //Bind Input Data
network.PushUnboundTraining(outFF);
network.FirstLayer.CrossLinkForward();
Layer hidden = network.FirstLayer.TargetLayer;
Layer output = network.LastLayer;
for(int i=0;i<output.Count;i++)
{
hidden[i*4].ConnectToNeuron(output[i]);
hidden[i*4+1].ConnectToNeuron(output[i]);
hidden[i*4+2].ConnectToNeuron(output[i]);
hidden[i*4+3].ConnectToNeuron(output[i]);
}
}
}
public void RunDemo()
{
Console.WriteLine("### BASIC UNBOUND DEMO ###");
//Initialize the network manager.
//This constructor also creates the first
//network layer (Inputlayer).
Network network = new Network();
//You need to initialize (the size of) the
//input layer in an unbound scenario
network.InitUnboundInputLayer(3);
//Add the hidden layer with 4 neurons.
network.AddLayer(4);
//Add the output layer with 2 neurons.
network.AddLayer(2);
//Connect the neurons together using synapses.
//This is the easiest way to do it; I'll discuss
//other ways in more detail in another demo.
network.AutoLinkFeedforward();
//Push new input data
network.PushUnboundInput(new bool[] { false, true, false });
//... and output training data ...
network.PushUnboundTraining(new bool[] { false, true });
//Propagate the network using the bound input data.
//Internally, this is a two round process, to
//correctly handle feedbacks
network.CalculateFeedforward();
//Collect the network output and print it.
App.PrintArray(network.CollectOutput());
//Train the current pattern using Backpropagation (one step)!
network.TrainCurrentPattern(false, true);
//Print the output; the difference to (-1,1) should be
//smaller this time!
App.PrintArray(network.CollectOutput());
//Same one more time:
network.TrainCurrentPattern(false, true);
App.PrintArray(network.CollectOutput());
//Train another pattern:
Console.WriteLine("# new pattern:");
//this time we're using doubles directly, instead of booleans.
//5/1 are the default values for input/training values.
network.PushUnboundInput(new double[] { 5d, -5d, -5d });
network.PushUnboundTraining(new double[] { 1, 1 });
//calculate ...
network.CalculateFeedforward();
App.PrintArray(network.CollectOutput());
//... and train it one time
network.TrainCurrentPattern(false, true);
App.PrintArray(network.CollectOutput());
//what about the old pattern now?
Console.WriteLine("# the old pattern again:");
network.PushUnboundInput(new double[] { -5d, 5d, -5d });
network.PushUnboundTraining(new double[] { -1, 1 });
network.CalculateFeedforward();
App.PrintArray(network.CollectOutput());
Console.WriteLine("=== COMPLETE ===");
Console.WriteLine();
}
public void RunDemo()
{
Console.WriteLine("### BASIC BOUND DEMO ###");
//Prepare you're input and training data
//to bind to the network
double[] input = new double[] { -5d, 5d, -5d };
double[] training = new double[] { -1, 1 };
//Initialize the network manager.
//This constructor also creates the first
//network layer (Inputlayer).
Network network = new Network();
//Bind your input array (to the already
//existing input layer)
network.BindInputLayer(input);
//Add the hidden layer with 4 neurons.
network.AddLayer(4);
//Add the output layer with 2 neurons.
network.AddLayer(2);
//bind your training array to the output layer.
//Always do this AFTER creating the layers.
network.BindTraining(training);
//Connect the neurons together using synapses.
//This is the easiest way to do it; I'll discuss
//other ways in more detail in another demo.
network.AutoLinkFeedforward();
//Propagate the network using the bound input data.
//Internally, this is a two round process, to
//correctly handle feedbacks
network.CalculateFeedforward();
//Collect the network output and print it.
App.PrintArray(network.CollectOutput());
//Train the current pattern using Backpropagation (one step)!
network.TrainCurrentPattern(false, true);
//Print the output; the difference to (-1,1) should be
//smaller this time!
App.PrintArray(network.CollectOutput());
//Same one more time:
network.TrainCurrentPattern(false, true);
App.PrintArray(network.CollectOutput());
//Train another pattern:
Console.WriteLine("# new pattern:");
input[0] = 5d;
input[1] = -5d;
training[0] = 1;
//calculate ...
network.CalculateFeedforward();
App.PrintArray(network.CollectOutput());
//... and train it one time
network.TrainCurrentPattern(false, true);
App.PrintArray(network.CollectOutput());
//what about the old pattern now?
Console.WriteLine("# the old pattern again:");
input[0] = -5d;
input[1] = 5d;
training[0] = -1;
network.CalculateFeedforward();
App.PrintArray(network.CollectOutput());
Console.WriteLine("=== COMPLETE ===");
Console.WriteLine();
}
static void Main(string[] args)
{
Controler ffc = new Controler(1);
Network net = new Network(ffc.Node);
net.InitUnboundInputLayer(1).BasicConfiguration.ActivationType.Value = EActivationType.Linear;
net.AddLayer(1, EActivationType.Linear);
net.AutoLinkFeedforward();
Neuron neuron = net.LastLayer[0];
Synapse synapse = neuron.SourceSynapses[0];
ffc.ImportNetwork(net, false);
BasicConfig config = ffc.NeuralNetwork.BasicConfiguration;
config.BiasNeuronEnable.Value = true;
config.BiasNeuronOutput.Value = 1.0;
config.FlatspotEliminationEnable.Value = false;
config.WeightDecayEnable.Value = false;
config.SymmetryPreventionEnable.Value = false;
config.ManhattanTrainingEnable.Value = false;
config.LearningRate.Value = 0.005;
StochasticCoordinateGenerator scg = new StochasticCoordinateGenerator(0,10,100);
//RegularCoordinateGenerator rcg = new RegularCoordinateGenerator(-25, 25, 50);
DynamicSampleProvider dsp = new DynamicSampleProvider(my_func, scg); //rcg);
ffc.Provider = dsp; // new CachedSampleProvider(dsp);
Console.WriteLine("TARGET FUNCTION: 3*x+5");
Console.WriteLine("TARGET Synapse Weight = 3.0");
Console.WriteLine("TARGET Bias Weight = 5.0");
Console.WriteLine("TARGET Mean Squared Error <= 0.000000001");
Console.WriteLine();
Console.WriteLine("Synapse Weight: " + synapse.Weight + " - Bias Weight: " + neuron.BiasNeuronWeight);
Console.WriteLine("Initial MSE: " + ffc.EstimateMeanSquaredError());
Console.WriteLine();
ffc.TrainAllSamplesOnce();
Console.WriteLine("Synapse Weight: " + synapse.Weight + " - Bias Weight: " + neuron.BiasNeuronWeight);
Console.WriteLine("Trained MSE: " + ffc.EstimateMeanSquaredError());
Console.WriteLine();
ffc.TrainAllSamplesOnce();
Console.WriteLine("Synapse Weight: " + synapse.Weight + " - Bias Weight: " + neuron.BiasNeuronWeight);
Console.WriteLine("Trained MSE: " + ffc.EstimateMeanSquaredError());
Console.WriteLine();
ffc.TrainAllSamplesOnce();
Console.WriteLine("Synapse Weight: " + synapse.Weight + " - Bias Weight: " + neuron.BiasNeuronWeight);
Console.WriteLine("Trained MSE: " + ffc.EstimateMeanSquaredError());
Console.WriteLine();
Console.WriteLine("Auto Training, maximum 1000 Epochs");
Console.WriteLine();
if(ffc.TrainAllSamplesUntil(0.000000001, 1000))
{
Console.WriteLine("Synapse Weight: " + synapse.Weight + " - Bias Weight: " + neuron.BiasNeuronWeight);
Console.ForegroundColor = ConsoleColor.Green;
Console.WriteLine("SUCCEEDS auto training with MSE: " + ffc.EstimateMeanSquaredError());
Console.ResetColor();
}
else
{
Console.WriteLine("Synapse Weight: " + synapse.Weight + " - Bias Weight: " + neuron.BiasNeuronWeight);
Console.ForegroundColor = ConsoleColor.Red;
Console.WriteLine("FAILS auto training with MSE: " + ffc.EstimateMeanSquaredError());
Console.ResetColor();
}
Console.ReadKey();
}
private void загрузитьБольшуюМодельToolStripMenuItem_Click(object sender, EventArgs e)
{
Network network = new Network();
network.AddLayer(new Conv2D(new Relu(), 5, 5, 16));
network.AddLayer(new MaxPool2D(new Relu(), 2, 2));
network.AddLayer(new Conv2D(new Relu(), 5, 5, 20));
network.AddLayer(new MaxPool2D(new Relu(), 2, 2));
network.AddLayer(new Conv2D(new Relu(), 3, 3, 32));
network.AddLayer(new MaxPool2D(new Relu(), 2, 2));
network.AddLayer(new Conv2D(new Relu(), 3, 3, 54));
network.AddLayer(new MaxPool2D(new Relu(), 2, 2));
network.AddLayer(new Conv2D(new Relu(), 3, 3, 80));
network.AddLayer(new MaxPool2D(new Relu(), 2, 2));
network.AddLayer(new FullyConnLayar(new Sigmoid(), new NeuralNetwork.Size(1, 1, 100)));
network.AddLayer(new FullyConnLayar(new Sigmoid(), new NeuralNetwork.Size(1, 1, 2)));
network.Compile(new NeuralNetwork.Size(3, NetworkData.image_size.Height, NetworkData.image_size.Width), true);
network.Normalization();
network.Normalization();
NetworkData.network = network;
string message = "";
for (int i = 0; i < network.layers.Count; i++)
{
message += network.layers[i].GetType().Name + " {" + network.layers[i].output_size[0] + "," + network.layers[i].output_size[1] + "," + network.layers[i].output_size[2] + "}\n";
}
WriteNetwork();
MessageBox.Show(message);
}
/// <summary>
/// Standard constructor of NetTrainer, should be used in most cases.
/// </summary>
/// <param name = "dalManager">Database gateway</param>
public NetTrainer(DaoGateway dalManager)
{
Network net = new Network();
net.AddLayer(new BasicLayer(new ActivationTANH(), true, DEFAULT_FINGERPRINT_SIZE));
net.AddLayer(new BasicLayer(new ActivationTANH(), true, DEFAULT_HIDDEN_NEURONS_COUNT));
net.AddLayer(new BasicLayer(new ActivationTANH(), false, OUT_PUT_NEURONS));
net.Structure.FinalizeStructure();
net.Reset();
Init(net, dalManager);
}
