public Connection(UInt64 _conn_id, INet _net, Socket _socket, ISession _sess)
{
conn_id = _conn_id;
net = _net;
socket = _socket;
session = _sess;
}
public InfinityScroll(IBlue blue, IRed red, INet net, IShowData showData)
{
_blue = blue;
_red = red;
_net = net;
_showData = showData;
}
private static void AddNetInfos(IStep step, IODBLayer parentLayer, String NetName, IODBObject netItem)
{
if (NetName.Length == 0) //the default net
{
NetName = "$NONE$";
}
int lastNetNr = step.GetNets().Count;
INet outNet = step.GetNet(NetName);
int netNr = -1;
if (outNet == null)
{
netNr = step.AddNet(NetName, "", out outNet);
}
else
{
netNr = outNet.GetNetNumber();
}
netItem.PcbNetNumber = outNet.GetNetNumber(); //set the netinformation to the object
int newNr = outNet.AddLayerRef(parentLayer.GetLayerName(), lastNetNr); //each net saves a list of all layers are used.
if (newNr == lastNetNr)
{
lastNetNr++;
}
parentLayer.SetNetNumber(NetName, netItem.PcbNetNumber); //for each object the layer needs information that this net is used.
outNet.AddFID(PCBI.FidType.Copper, parentLayer.GetLayerName(), netItem.GetIndexOnLayer(), ""); //this is specific to ODB++, if you don't use it and save the data as ODB++ some information get lost.
}
protected static double GetPatternError(INet net, ILearningPattern pattern)
{
var actual = net.Propagate(pattern.Input);
var expected = pattern.Output;
var errors = new List<IFuzzyNumber>();
var i = 0;
foreach (var actualNumber in actual)
{
errors.Add(actualNumber.Sub(expected.ElementAt(i)));
i++;
}
var patternError = 0.0;
foreach (var errorNumber in errors)
{
var leftError = 0.0;
var rightError = 0.0;
errorNumber.ForeachLevel((alpha, level) =>
{
leftError += alpha * (level.X * level.X);
rightError += alpha * (level.Y * level.Y);
});
var currentOutputError = leftError + rightError;
patternError += currentOutputError;
}
return patternError / 2.0;
}
public static INetData Data(this INet net)
{
return(new NetData(
net.Id,
net.AllNeurons.Select(n => n.Data()).ToList(), net.AllConnections.Select(c => c.Data()).ToList(),
net.InputNeurons.Select(n => n.Id).ToList(), net.OutputNeurons.Select(n => n.Id).ToList()));
}
//here we have _gradient as sum of gradients, should make step in this direction
protected override bool LearnBatch(INet net, double currentLearningCycleError)
{
if (_prevStep != null)
{
var y = _gradient.Negate().Sum(_prevGradient); //yk
//_gradientDiffNorm = y.Norm;
//Console.WriteLine("Gradient diff norm: {0}", _gradientDiffNorm.GetMod().X);
//Console.WriteLine();
//if (IsNetLearned(currentLearningCycleError))
// return;
//its time to calculate b(k + 1)
_b = CalculateInvertedPseudoGaussian(_b, _prevStep, y);
}
var direction = CalculateMinimizeDirection(_b, _gradient); //pk - direction of next step
var step = MakeStep(direction, net, currentLearningCycleError); //step = alpha*pk
//var step = MakeStepGoldstein(direction, net, currentLearningCycleError, _gradient); //step = alpha*pk
if (step == null)
{
}
//Save step and grad
_prevStep = step;
_prevGradient = _gradient;
//clear gradient vector
_gradient = null;
return true;
}
public bool OnAccept(INet net)
{
if (net == null)
return false;
return true;
}
public static void SaveNet(this INet net, string filename)
{
using (var fs = new FileStream(filename, FileMode.Create))
{
net.SaveBinary(fs);
}
}
public void Execute(IPCBIWindow parent)
{
wdlg = new PCB_Investigator.PCBIWindows.PCBIWorkingDialog();
wdlg.SetAnimationStatus(false);
wdlg.SetStatusPercent(0);
wdlg.SetStatusText("Working");
wdlg.CanCancel(true);
IStep curStep = parent.GetCurrentStep();
if (curStep == null)
{
return;
}
Dictionary <string, double> smallestDiameterList = new Dictionary <string, double>();
StringBuilder sbResult = new StringBuilder();
wdlg.ShowWorkingDlgAsThread();
List <string> netNames = curStep.GetAllNetNames();
double value = 0;
double valueStep = ((100.0 / netNames.Count));
foreach (string netName in curStep.GetAllNetNames())
{
INet net = curStep.GetNet(netName);
wdlg.SetStatusText("Working on " + netName + "...");
value += valueStep;
wdlg.SetStatusPercent((int)(value));
List <IODBObject> allNetElements = net.GetAllNetObjects(parent);
if (allNetElements.Count == 0)
{
continue;
}
double smallestDiameter = allNetElements[0].GetDiameter();
foreach (IODBObject netElement in allNetElements)
{
double currentDiameter = netElement.GetDiameter();
if (currentDiameter < 0)
{
continue; //e.g. surfaces have no diameter
}
if (currentDiameter < smallestDiameter)
{
smallestDiameter = currentDiameter;
}
}
smallestDiameterList.Add(netName, smallestDiameter);
sbResult.AppendLine(netName + ": " + smallestDiameter.ToString() + " mils");
}
wdlg.Dispose();
PCB_Investigator.Localization.PCBILocalization.ShowMsgBox("All smallest Net Diameters:" + Environment.NewLine + sbResult.ToString(), "Result", MessageBoxButtons.OK, MessageBoxIcon.Information);
}
public YoutubeService(INet net)
{
if (net == null)
{
throw new ArgumentNullException("net");
}
this.net = net;
}
public void Load()
{
this.net = NetExtension.LoadNet($"{this.Prefix}net.dat");
FluentNet fluentNet = net as FluentNet;
this.inputWidth = fluentNet.InputLayers[0].InputWidth;
}
public static void SaveNetState(string filename, INet net)
{
using (var fs = new FileStream(filename, FileMode.Create, FileAccess.Write, FileShare.ReadWrite))
{
var bf = new BinaryFormatter();
bf.Serialize(fs, net);
}
}
public IConnection Create(INet net, Socket socket, ISession session)
{
++next_id;
Connection conn = new Connection(next_id, net, socket, session);
dict[next_id] = conn;
return(conn);
}
protected override void PrepareToLearning(INet net)
{
_b = Matrix.CreateI(net.WeightsCount, net.WeightsCount, () => new RealNumber(1), () => new RealNumber(0)); //b0
_weights = net.GetWeights(); //x0
_inputs = net.GetLastInputsForWeights();
_gradient = null;
_prevGradient = null;
_prevStep = null;
}
protected override void LearnPattern(INet net, ILearningPattern learningPattern, double currentPatternError)
{
//call only after net.propagation()
PropagateErrorOnLayers(net.Layers, learningPattern.Output);
CalculateWeightDelta(net);
//ChangeAndSetWeights(_deltas, net);
//_deltas = null;
}
public void Accept(INet sourceNet)
{
_network = Network.Load(sourceNet.Persistence()) as ActivationNetwork;
_teacher = new BackPropagationLearning(_network)
{
LearningRate = _learningRate,
Momentum = 0.9
};
}
public TrainingScheme(INet net, TrainerBase trainer, EntryContainer container, string label)
{
this.container = container;
this.label = label;
this.net = net;
this.trainer = trainer;
int n = container.ClassCount;
}
public bool Create()
{
m_pListenNet = libNet.Network.MyNetwork.Listen("0.0.0.0:1111",this);
if (m_pListenNet == null)
{
return false;
}
ThreadPool.QueueUserWorkItem(ClientProcesFunc, this);
return true;
}
public AdamTrainer(INet <T> net) : base(net)
{
if (typeof(T) == typeof(double))
{
this.Eps = (T)(ValueType)1e-8;
}
else if (typeof(T) == typeof(float))
{
this.Eps = (T)(ValueType)(float)1e-8;
}
}
private void MnistDemo()
{
Directory.CreateDirectory(mnistFolder);
string trainingLabelFilePath = Path.Combine(mnistFolder, trainingLabelFile);
string trainingImageFilePath = Path.Combine(mnistFolder, trainingImageFile);
string testingLabelFilePath = Path.Combine(mnistFolder, testingLabelFile);
string testingImageFilePath = Path.Combine(mnistFolder, testingImageFile);
// Download Mnist files if needed
Console.WriteLine("Downloading Mnist training files...");
DownloadFile(urlMnist + trainingLabelFile, trainingLabelFilePath);
DownloadFile(urlMnist + trainingImageFile, trainingImageFilePath);
Console.WriteLine("Downloading Mnist testing files...");
DownloadFile(urlMnist + testingLabelFile, testingLabelFilePath);
DownloadFile(urlMnist + testingImageFile, testingImageFilePath);
// Load data
Console.WriteLine("Loading the datasets...");
this.training = MnistReader.Load(trainingLabelFilePath, trainingImageFilePath);
this.testing = MnistReader.Load(testingLabelFilePath, testingImageFilePath);
if (this.training.Count == 0 || this.testing.Count == 0)
{
Console.WriteLine("Missing Mnist training/testing files.");
Console.ReadKey();
return;
}
// Create network
this.net = FluentNet.Create(24, 24, 1)
.Conv(5, 5, 8).Stride(1).Pad(2)
.Relu()
.Pool(2, 2).Stride(2)
.Conv(5, 5, 16).Stride(1).Pad(2)
.Relu()
.Pool(3, 3).Stride(3)
.FullyConn(10)
.Softmax(10)
.Build();
this.trainer = new AdadeltaTrainer(this.net)
{
BatchSize = 20,
L2Decay = 0.001,
};
Console.WriteLine("Convolutional neural network learning...[Press any key to stop]");
do
{
var sample = this.SampleTrainingInstance();
this.Step(sample);
} while (!Console.KeyAvailable);
}
/// <exception cref="Exception">When called concurrently with <see cref="TickAsync"/>.</exception>
public INet CreatePlace(out Place place, string description = null)
{
if (_tickTask != null && !_tickTask.IsCompleted)
{
throw new Exception($"{nameof(CreateTransition)} called during tick.");
}
INet net = CreatePlace <Place>(out Place instance, description);
place = (Place)instance;
return(net);
}
public static INet LoadNet(string filename)
{
INet result = null;
if (File.Exists(filename))
{
using (var fs = new FileStream(filename, FileMode.Open))
{
result = SerializationExtensions.LoadBinary(fs);
}
}
return(result);
}
private void sendNetPacket(MsgType t, INet outmsg)
{
NetPacket msg = new NetPacket();
MemoryStream stream = new MemoryStream();
BinaryWriter buffer = new BinaryWriter(stream);
outmsg.Serialize(buffer);
if (buffer.BaseStream.Length + NetPacket.DEFAULT_FRAME_LEN > 512)
{
msg.message_type = (byte)MsgType.Multipart;
// calculate how many parts we have to split this into
int maxsize = 512 - (12 + NetPacket.DEFAULT_FRAME_LEN);
int parts = ((int)buffer.BaseStream.Length / maxsize) + 1;
this.multi_groupid++;
int bstart = 0;
for (int i = 0; i < parts; i++)
{
int bend = bstart + maxsize;
if (i + 1 == parts)
{
bend = bstart + (((int)buffer.BaseStream.Length) % maxsize);
}
Multipart wrapper = new Multipart();
wrapper.ID = (ushort)i;
wrapper.GroupID = this.multi_groupid;
wrapper.NumParts = (ushort)parts;
wrapper.Content = new byte[bend - bstart];
buffer.BaseStream.Read(wrapper.Content, bstart, bend - bstart);
MemoryStream pstream = new MemoryStream();
BinaryWriter pbuffer = new BinaryWriter(pstream);
wrapper.Serialize(pbuffer);
msg.content = pstream.ToArray();
msg.content_length = (ushort)pstream.Length;
this.sending_socket.Send(msg.MessageBytes());
bstart = bend;
}
}
else
{
msg.content = stream.ToArray();
msg.content_length = (ushort)msg.content.Length;
msg.message_type = (byte)t;
this.sending_socket.Send(msg.MessageBytes());
}
}
private void MnistDemo()
{
var datasets = new DataSets();
if (!datasets.Load(100))
{
return;
}
// Create network
this._net = FluentNet <double> .Create(24, 24, 1)
.Conv(5, 5, 8).Stride(1).Pad(2)
.Relu()
.Pool(2, 2).Stride(2)
.Conv(5, 5, 16).Stride(1).Pad(2)
.Relu()
.Pool(3, 3).Stride(3)
.FullyConn(10)
.Softmax(10)
.Build();
this._trainer = new SgdTrainer <double>(this._net)
{
LearningRate = 0.01,
BatchSize = 20,
L2Decay = 0.001,
Momentum = 0.9
};
Console.WriteLine("Convolutional neural network learning...[Press any key to stop]");
do
{
var trainSample = datasets.Train.NextBatch(this._trainer.BatchSize);
Train(trainSample.Item1, trainSample.Item2, trainSample.Item3);
var testSample = datasets.Test.NextBatch(this._trainer.BatchSize);
Test(testSample.Item1, testSample.Item3, this._testAccWindow);
Console.WriteLine("Loss: {0} Train accuracy: {1}% Test accuracy: {2}%", this._trainer.Loss,
Math.Round(this._trainAccWindow.Items.Average() * 100.0, 2),
Math.Round(this._testAccWindow.Items.Average() * 100.0, 2));
Console.WriteLine("Example seen: {0} Fwd: {1}ms Bckw: {2}ms", this._stepCount,
Math.Round(this._trainer.ForwardTimeMs, 2),
Math.Round(this._trainer.BackwardTimeMs, 2));
} while (!Console.KeyAvailable);
}
public Application(IApplicationListener listener, ApplicationConfiguration config)
{
initializeGlfw();
setApplicationLogger(new ApplicationLogger());
this.config = ApplicationConfiguration.copy(config);
if (this.config.title == null)
{
this.config.title = listener.GetType().Name;
}
Gdx.app = this;
if (!config.disableAudio)
{
try
{
//this.audio = Gdx.audio = new OpenALAudio(config.audioDeviceSimultaneousSources, config.audioDeviceBufferCount, config.audioDeviceBufferSize);
}
catch (Exception t)
{
log("Lwjgl3Application", "Couldn't initialize audio, disabling audio", t);
//this.audio = Gdx.audio = new MockAudio();
}
}
else
{
//this.audio = Gdx.audio = new MockAudio();
}
files = Gdx.files = new Files();
net = Gdx.net = new Net();
clipboard = new Clipboard();
var window = createWindow(config, listener, null);
windows.Add(window);
loop();
cleanupWindows();
cleanup();
}
public static void SaveBinary(this INet net, Stream stream)
{
IFormatter formatter = new BinaryFormatter();
formatter.Serialize(stream, net);
}
public double Train(INet Net, int Epochs = 1000)
{
Dictionary<INode, double> deltas = new Dictionary<INode,double>();
foreach(var nodeLayer in Net.NodeLayers)
{
foreach(var node in nodeLayer.Nodes)
{
deltas.Add(node, 0.0);
}
}
for (int i = 0; i < Epochs; i++)
{
foreach (var dataSet in DataSets)
{
double error = 0;
var results = Net.Calculate(dataSet.Inputs, dataSet.Outputs, ref error);
// set delta of output nodes
for(int r = 0; r < results.Length; r++)
{
double delta = (dataSet.Outputs[r] - results[r]) * results[r] * (1 - results[r]);
deltas[Net.NodeLayers.Last().Nodes[r]] = delta;
}
for(int l = Net.NodeLayers.Length - 2; l >= 0; l--)
{
for(int l2 = 0; l2 < Net.NodeLayers[l].Nodes.Length; l2++)
{
var node = Net.NodeLayers[l].Nodes[l2];
double delta = 0;
foreach(var linkedNode in Net.NodeLayers[l+1].Nodes)
{
// add delta * weight of that node
delta += node.Result * (1 - node.Result) * linkedNode.Weights[l2] * deltas[linkedNode];
}
// save delta for other nodes
deltas[node] = delta;
}
}
foreach (var delta in deltas)
{
AdjustNode(delta.Key, dataSet.Inputs, delta.Value);
}
}
}
double SSE = 0;
foreach (var dataSet in DataSets)
{
double error = 0;
var result = Net.Calculate(dataSet.Inputs, dataSet.Outputs, ref error);
SSE += error;
}
return SSE;
}
public PropagationViewModel(Action<ViewModelBase> nextViewModel, string filename)
: base(nextViewModel)
{
_net = BinaryFileSerializer.LoadNetState(filename);
}
private static void SetPropagatedErrorToZero(INet net)
{
foreach (var layer in net.Layers)
{
layer.ForeachNeuron((i, neuron) =>
{
neuron.PropagatedError = new RealNumber(0.0);
});
}
}
public static void Run()
{
const string NetName = "net.dat";
var random = new Random();
var entryContainer = new EntryContainer();
#region Load Net
var convInputWith = 11; // Will extract 11x11 area
if (convInputWith % 2 == 0)
{
throw new ArgumentException("convInputWith must be odd");
}
// Load IA or initialize new network if not found - Direction choice
INet singleNet = NetExtension.LoadOrCreateNet(NetName, () =>
{
var net = FluentNet.Create(convInputWith, convInputWith, 3)
.Conv(3, 3, 16).Stride(2)
.Tanh()
.Conv(2, 2, 16)
.Tanh()
.FullyConn(100)
.Relu()
.FullyConn(5)
.Softmax(5).Build();
return(net);
});
#endregion
#region Load data
var hltFiles = Directory.EnumerateFiles(@"..\..\..\games\2609\", "*.hlt").ToList(); // erdman games downloaded with HltDownloader
int total = hltFiles.Count;
Console.WriteLine($"Loading {total} games...");
foreach (var file in hltFiles)
{
Console.WriteLine(total--);
HltReader reader = new HltReader(file);
var playerId = -1;
var playerToCopy = reader.PlayerNames.FirstOrDefault(o => o.StartsWith("erdman"));
if (playerToCopy == null)
{
Console.WriteLine("Player not found");
continue;
}
playerId = reader.PlayerNames.IndexOf(playerToCopy) + 1;
var width = reader.Width;
var height = reader.Height;
int lastmoveCount = 1;
for (var frame = 0; frame < reader.FrameCount - 1; frame++)
{
var currentFrame = reader.GetFrame(frame);
var map = currentFrame.map;
var moves = currentFrame.moves;
var helper = new Helper(map, (ushort)playerId);
bool foundInFrame = false;
int moveCount = 0;
// moves
for (ushort x = 0; x < width; x++)
{
for (ushort y = 0; y < height; y++)
{
if (map[x, y].Owner == playerId)
{
foundInFrame = true;
moveCount++;
if (random.NextDouble() < 1.0 / lastmoveCount)
{
var convVolume = map.GetVolume(convInputWith, playerId, x, y); // Input
var direction = moves[y][x]; // Output
var entry1 = new Entry(new[] { convVolume }, direction, x, y, frame, file.GetHashCode());
entryContainer.Add(entry1);
// Data augmentation
var entry2 = new Entry(new[] { convVolume.Flip(VolumeUtilities.FlipMode.LeftRight) }, (int)Helper.FlipLeftRight((Direction)direction), x, y, frame, file.GetHashCode());
entryContainer.Add(entry2);
var entry3 = new Entry(new[] { convVolume.Flip(VolumeUtilities.FlipMode.UpDown) }, (int)Helper.FlipUpDown((Direction)direction), x, y, frame, file.GetHashCode());
entryContainer.Add(entry3);
var entry4 = new Entry(new[] { convVolume.Flip(VolumeUtilities.FlipMode.Both) }, (int)Helper.FlipBothWay((Direction)direction), x, y, frame, file.GetHashCode());
entryContainer.Add(entry4);
}
}
}
}
lastmoveCount = moveCount;
if (!foundInFrame)
{
// player has died
break;
}
}
}
var length = entryContainer.Shuffle();
Console.WriteLine(entryContainer.Summary);
#endregion
#region Training
var trainer = new AdamTrainer(singleNet)
{
BatchSize = 1024, LearningRate = 0.1, Beta1 = 0.9, Beta2 = 0.99, Eps = 1e-8
};
var trainingScheme = new TrainingScheme(singleNet, trainer, entryContainer, "single");
bool save = true;
double lastValidationAcc = 0.0;
do
{
for (int i = 0; i < 1000; i++)
{
if (i > 5)
{
trainer.L2Decay = 0.001;
}
Console.WriteLine($"Epoch #{i + 1}");
if (i % 15 == 0)
{
trainer.LearningRate = Math.Max(trainer.LearningRate / 10.0, 0.00001);
}
trainingScheme.RunEpoch();
#region Save Nets
if (save)
{
// Save if validation accuracy has improved
if (trainingScheme.ValidationAccuracy > lastValidationAcc)
{
lastValidationAcc = trainingScheme.ValidationAccuracy;
singleNet.SaveNet(NetName);
}
}
#endregion
if (Console.KeyAvailable)
{
break;
}
}
} while (!Console.KeyAvailable);
#endregion
}
public SgdTrainer(INet <double> net) : base(net)
{
}
private IVector MakeStep(IVector direction, INet net, double currentError)
{
const int maximumNumberOfTry = 50;
var numberOfTry = 0;
double error;
var step = direction.Mul(_alpha); ;
var oldWeights = _weights;
//can change alpha by minimizing it in f(xk + alpha*direction)
do
{
if (numberOfTry > maximumNumberOfTry || _alpha < 0.000000000001)
break;
_weights = oldWeights.Sum(step); //x(k+1) = xk + sk
net.SetWeights(_weights); //content of _weights now shared between net and _weights vector
error = GetBatchError(net);
_alpha /= 2.1;
step = direction.Mul(_alpha);
numberOfTry++;
} while (error > currentError);
if (numberOfTry > maximumNumberOfTry || _alpha < 0.000000000001)
{
Console.WriteLine("Simple step was performed. Too little alpha: {0:0.#############}.", _alpha);
//step = direction.Mul(_alpha);
_weights = oldWeights.Sum(direction.Mul(0.1));
net.SetWeights(_weights);
//AddLittleCorrectionToWeights(net.Layers);
//_gradient = null;
_alpha = 1.0;
return null;
}
_alpha = 1.0;
return step;
}
private void ChangeAndSetWeights(IVector delta, INet net)
{
_weights = _weights.Sum(delta);
net.SetWeights(_weights);
}
protected virtual double CalculatePatternError(INet net, ILearningPattern pattern)
{
return GetPatternError(net, pattern);
}
protected double GetBatchError(INet net)
{
return _patterns.Sum(learningPattern => CalculatePatternError(net, learningPattern));
}
protected abstract bool LearnBatch(INet net, double currentLearningCycleError);
public void LearnNet(INet net)
{
_worker.RunWorkerAsync(net);
}
protected abstract void PrepareToLearning(INet net);
internal Place(INet net, string description)
{
_net = net;
_description = description;
}
private IVector MakeStepGoldstein(IVector direction, INet net, double currentError, IVector gradient)
{
const int maximumNumberOfTry = 50;
var numberOfTry = 0;
double error;
var step = direction.Mul(_alpha); ;
var c = 0.1;
var p = _random.NextDouble()*0.3 + 0.3;
var oldWeights = _weights;
//can change alpha by minimizing it in f(xk + alpha*direction)
var threshold = 0.0;
do
{
if (numberOfTry > maximumNumberOfTry || _alpha < 0.00001)
break;
_weights = oldWeights.Sum(step); //x(k+1) = xk + sk
net.SetWeights(_weights); //content of _weights now shared between net and _weights vector
error = GetBatchError(net);
threshold = gradient.Mul(step).Mul(_alpha).Mul(c).GetMod().X;
_alpha *= p;
step = direction.Mul(_alpha);
numberOfTry++;
} while (error > currentError + threshold);
if (numberOfTry > maximumNumberOfTry || _alpha < 0.00001)
{
_weights = oldWeights;
net.SetWeights(_weights);
//will make gradient descent on next step
_gradient = null;
_b = Matrix.CreateI(net.WeightsCount, net.WeightsCount, () => new RealNumber(1), () => new RealNumber(0)); //b0
_alpha = 1.0;
Console.WriteLine("Reset matrix");
return null;
}
_alpha = 1.0;
return step;
}
protected override bool LearnBatch(INet net, double currentLearningCycleError)
{
ChangeAndSetWeights(_deltas, net);
_deltas = null;
return true;
}
private void CalculateWeightDelta(INet net)
{
var gradient = CreateWeightsGradient(net.Layers);
var currentDelta = gradient.Mul(_alpha);
_deltas = _deltas == null ? currentDelta : _deltas.Sum(currentDelta);
}
//here we summarize gradient of each pattern
protected override void LearnPattern(INet net, ILearningPattern learningPattern, double currentPatternError)
{
PropagateErrorOnLayers(net.Layers, learningPattern.Output); //nablaF(xk)
var currentGradient = CreateWeightsGradient(net.Layers);
_gradient = _gradient == null ? currentGradient : _gradient.Sum(currentGradient);
}
protected override void PrepareToLearning(INet net)
{
_weights = net.GetWeights();
_outputs = net.GetLastInputsForWeights();
}
public override void OnRecvice(INet net, byte[] data, short len)
{
}
public NeuroEvolveBot(INet net, int seed = 3, string name = "NeuroEvolve")
{
rand = new Random(seed);
network = net;
Name = name;
}
protected int K; // iteration counter
protected TrainerBase(INet net)
{
this.Net = net;
this.BatchSize = 1;
}
public SgdTrainer(INet <T> net, T learningRate) : base(net)
{
this.LearningRate = learningRate;
this.Optimizer = new GradientDescentOptimizer <T>(learningRate);
}
protected TrainerBase(INet <T> net)
{
this.Net = net;
}
public AdamTrainer(INet <float> net) : base(net)
{
}
public SgdTrainer(INet <T> net) : base(net)
{
}
public static void Run()
{
var random = new Random(RandomUtilities.Seed);
int normalInputWidth = 19;
int earlyInputWidth = 19;
int strongInputWidth = 19;
string NetName = $"net.dat";
string NetName_early = $"net_early.dat";
string NetName_strong = $"net_strong.dat";
var entryContainer = new EntryContainer();
var entryContainer_early = new EntryContainer();
var entryContainer_strong = new EntryContainer();
#region Load Net
INet singleNet = NetExtension.LoadOrCreateNet(NetName, () =>
{
var net = FluentNet.Create(normalInputWidth, normalInputWidth, 3)
.Conv(5, 5, 16).Stride(5).Pad(2)
.Tanh()
.Conv(3, 3, 16).Stride(1).Pad(1)
.Tanh()
.FullyConn(100)
.Relu()
.FullyConn(5)
.Softmax(5).Build();
return(net);
});
INet singleNet_early = NetExtension.LoadOrCreateNet(NetName_early, () =>
{
var net = FluentNet.Create(earlyInputWidth, earlyInputWidth, 3)
.Conv(5, 5, 16).Stride(5).Pad(2)
.Tanh()
.Conv(3, 3, 16).Stride(1).Pad(1)
.Tanh()
.FullyConn(100)
.Relu()
.FullyConn(5)
.Softmax(5).Build();
return(net);
});
INet singleNet_strong = NetExtension.LoadOrCreateNet(NetName_strong, () =>
{
var net = FluentNet.Create(strongInputWidth, strongInputWidth, 3)
.Conv(5, 5, 16).Stride(5).Pad(2)
.Tanh()
.Conv(3, 3, 16).Stride(1).Pad(1)
.Tanh()
.FullyConn(100)
.Relu()
.FullyConn(5)
.Softmax(5).Build();
return(net);
});
#endregion
#region Load data
var hltFiles = Directory.EnumerateFiles(@"..\..\..\games\2609\", "*.hlt").ToList(); // erdman games downloaded with HltDownloader
int total = hltFiles.Count;
Console.WriteLine($"Loading {total} games...");
foreach (var file in hltFiles)
{
Console.WriteLine(total--);
HltReader reader = new HltReader(file);
var playerId = -1;
var playerToCopy = reader.PlayerNames.FirstOrDefault(o => o.StartsWith("erdman"));
if (playerToCopy != null)
{
playerId = reader.PlayerNames.IndexOf(playerToCopy) + 1;
}
if (playerId != -1)
{
var width = reader.Width;
var height = reader.Height;
int lastmoveCount = 1;
for (var frame = 0; frame < reader.FrameCount - 1; frame++)
{
bool earlyGame = lastmoveCount < 25;
var currentFrame = reader.GetFrame(frame);
var map = currentFrame.map;
var moves = currentFrame.moves;
var helper = new Helper(map, (ushort)playerId);
bool foundInFrame = false;
int moveCount = 0;
// moves
for (ushort x = 0; x < width; x++)
{
for (ushort y = 0; y < height; y++)
{
if (map[x, y].Owner == playerId)
{
bool strong = map[x, y].Strength > 200;
foundInFrame = true;
moveCount++;
if ((earlyGame && random.NextDouble() < 1.5 / lastmoveCount) || (strong && random.NextDouble() < 1.5 / lastmoveCount) || random.NextDouble() < 1.0 / lastmoveCount)
{
var w = normalInputWidth;
var container = entryContainer;
if (earlyGame)
{
w = earlyInputWidth;
container = entryContainer_early;
}
else if (strong)
{
w = strongInputWidth;
container = entryContainer_strong;
}
var convVolume = map.GetVolume(w, playerId, x, y);
var direction = moves[y][x];
var entry1 = new Entry(new[] { convVolume }, direction, x, y, frame, file.GetHashCode());
container.Add(entry1);
var entry2 = new Entry(new[] { convVolume.Flip(VolumeUtilities.FlipMode.LeftRight) }, (int)Helper.FlipLeftRight((Direction)direction), x, y, frame, file.GetHashCode());
container.Add(entry2);
var entry3 = new Entry(new[] { convVolume.Flip(VolumeUtilities.FlipMode.UpDown) }, (int)Helper.FlipUpDown((Direction)direction), x, y, frame, file.GetHashCode());
container.Add(entry3);
var entry4 = new Entry(new[] { convVolume.Flip(VolumeUtilities.FlipMode.Both) }, (int)Helper.FlipBothWay((Direction)direction), x, y, frame, file.GetHashCode());
container.Add(entry4);
}
}
}
}
lastmoveCount = moveCount;
if (!foundInFrame)
{
// player has died
break;
}
}
}
else
{
Console.WriteLine("not found");
}
}
var length = entryContainer.Shuffle();
Console.WriteLine("normal: " + entryContainer.Summary);
length = entryContainer_early.Shuffle();
Console.WriteLine("early: " + entryContainer_early.Summary);
length = entryContainer_strong.Shuffle();
Console.WriteLine("strong " + entryContainer_strong.Summary);
#endregion
#region Training
var trainer = new AdamTrainer(singleNet)
{
BatchSize = 1024, LearningRate = 0.01, Beta1 = 0.9, Beta2 = 0.99, Eps = 1e-8
};
var trainingScheme = new TrainingScheme(singleNet, trainer, entryContainer, "single");
var trainer_early = new AdamTrainer(singleNet_early)
{
BatchSize = 1024, LearningRate = 0.01, Beta1 = 0.9, Beta2 = 0.99, Eps = 1e-8
};
var trainingScheme_early = new TrainingScheme(singleNet_early, trainer_early, entryContainer_early, "single_early");
var trainer_strong = new AdamTrainer(singleNet_strong)
{
BatchSize = 1024, LearningRate = 0.01, Beta1 = 0.9, Beta2 = 0.99, Eps = 1e-8
};
var trainingScheme_strong = new TrainingScheme(singleNet_strong, trainer_strong, entryContainer_strong, "single_strong");
bool save = true;
double lastValidationAcc = 0.0;
double lastValidationAcc_early = 0.0;
double lastValidationAcc_strong = 0.0;
double lastTrainAcc = 0.0;
double lastTrainAcc_early = 0.0;
double lastTrainAcc_strong = 0.0;
do
{
var normal = Task.Factory.StartNew(() =>
{
for (int i = 0; i < 50; i++)
{
if (i > 5)
{
trainer.L2Decay = 0.05;
}
Console.WriteLine($"[normal] Epoch #{i + 1}");
if (i % 50 == 0)
{
trainer.LearningRate = Math.Max(trainer.LearningRate / 5.0, 0.00001);
}
trainingScheme.RunEpoch();
#region Save Nets
if (save)
{
if (trainingScheme.ValidationAccuracy > lastValidationAcc)
{
lastValidationAcc = trainingScheme.ValidationAccuracy;
lastTrainAcc = trainingScheme.TrainAccuracy;
singleNet.SaveNet(NetName);
}
}
#endregion
if (Console.KeyAvailable)
{
break;
}
}
});
var early = Task.Factory.StartNew(() =>
{
for (int i = 0; i < 50; i++)
{
if (i > 5)
{
trainer_early.L2Decay = 0.05;
}
Console.WriteLine($"[early] Epoch #{i + 1}");
if (i % 50 == 0)
{
trainer_early.LearningRate = Math.Max(trainer_early.LearningRate / 5.0, 0.00001);
}
trainingScheme_early.RunEpoch();
#region Save Nets
if (save)
{
if (trainingScheme_early.ValidationAccuracy > lastValidationAcc_early)
{
lastValidationAcc_early = trainingScheme_early.ValidationAccuracy;
lastTrainAcc_early = trainingScheme_early.TrainAccuracy;
singleNet_early.SaveNet(NetName_early);
}
}
#endregion
if (Console.KeyAvailable)
{
break;
}
}
});
var strong = Task.Factory.StartNew(() =>
{
for (int i = 0; i < 50; i++)
{
if (i > 5)
{
trainer_strong.L2Decay = 0.05;
}
Console.WriteLine($"[strong] Epoch #{i + 1}");
if (i % 50 == 0)
{
trainer_strong.LearningRate = Math.Max(trainer_strong.LearningRate / 5.0, 0.00001);
}
trainingScheme_strong.RunEpoch();
#region Save Nets
if (save)
{
if (trainingScheme_strong.ValidationAccuracy > lastValidationAcc_strong)
{
lastValidationAcc_strong = trainingScheme_strong.ValidationAccuracy;
lastTrainAcc_strong = trainingScheme_strong.TrainAccuracy;
singleNet_strong.SaveNet(NetName_strong);
}
}
#endregion
if (Console.KeyAvailable)
{
break;
}
}
});
Task.WaitAll(new[] { normal, strong, early });
}while (!Console.KeyAvailable);
#endregion
}
public SgdTrainer(INet <float> net) : base(net)
{
}
protected abstract void LearnPattern(INet net, ILearningPattern learningPattern, double currentPatternError);
