private void StartAI()
{
BuilderInstance <float> .Volume = new VolumeBuilder(); // Needed for GPU, must be done on AI thread
// Must be done after setting BuilderInstance<float>.Volume
this._yhat = SerializationExtensions.Load <float>("FaceDetection", false)[0];
while (true)
{
if (this._queueIn.TryDequeue(out var bitmap))
{
bitmap = BitmapTool.ResizeImage(bitmap, 300, 300);
var boxes = new List <BoundingBox>();
var allExtracts =
BitmapTool.SlideWindow(bitmap, 100, 100 / 9);
// .Concat(BitmapTool.SlideWindow(bitmap, 40, 40 / 3));
//.Concat(BitmapTool.SlideWindow(bitmap, 50, 50/3));
foreach (var extracts in allExtracts.Batch(10))
{
boxes.AddRange(BitmapTool.Evaluate(extracts.ToList(), this._yhat, IMAGE_WIDTH, IMAGE_HEIGHT, 0.8f));
}
this._queueOut.Enqueue(boxes);
}
}
}
public void GraphMl()
{
var a = new Const <double>(1.0, "one");
var b = new Const <double>(2.0, "two");
var add = new Add <double>(a, b);
var activation = new Activation <double>(add, ActivationType.Relu);
activation.Save("test");
var result = SerializationExtensions.Load <double>("test", false);
}
public void GraphMl()
{
var cns = new ConvNetSharp <double>();
var a = cns.Const(1.0, "one");
var b = cns.Const(2.0, "two");
var add = a + b;
var activation = cns.Relu(add);
activation.Save("test");
var result = SerializationExtensions.Load <double>("test", false);
}
public Form1()
{
InitializeComponent();
this.openFileDialog1.InitialDirectory = Directory.GetCurrentDirectory();
if (!string.IsNullOrEmpty(Settings.Default.LastFilename))
{
SetFilename(Settings.Default.LastFilename);
}
BuilderInstance <float> .Volume = new VolumeBuilder();
; // Needed for GPU, must be done on AI thread
// Must be done after setting BuilderInstance<float>.Volume
this._yhat = SerializationExtensions.Load <float>("FaceDetection", false)[0];
}
/// <summary>
/// This sample shows how to serialize and deserialize a ConvNetSharp.Flow network
/// 1) Graph creation
/// 2) Dummy Training (only use a single data point)
/// 3) Serialization
/// 4) Deserialization
/// </summary>
private static void Main()
{
var cns = new ConvNetSharp <double>();
// 1) Graph creation
var input = cns.PlaceHolder("x"); // input
var dense1 = cns.Dense(input, 20) + cns.Variable(BuilderInstance <double> .Volume.From(new double[20].Populate(0.1), new Shape(20)), "bias1", true);
var relu = cns.Relu(dense1);
var dense2 = cns.Dense(relu, 10) + cns.Variable(new Shape(10), "bias2", true);
var softmax = cns.Softmax(dense2); // output
var output = cns.PlaceHolder("y"); // ground truth
var cost = new SoftmaxCrossEntropy <double>(cns, softmax, output);
var x = BuilderInstance <double> .Volume.From(new[] { 0.3, -0.5 }, new Shape(2));
var y = BuilderInstance <double> .Volume.From(new[] { 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0 }, new Shape(10));
var dico = new Dictionary <string, Volume <double> > {
{ "x", x }, { "y", y }
};
var count = 0;
var optimizer = new GradientDescentOptimizer <double>(cns, 0.01);
using (var session = new Session <double>())
{
session.Differentiate(cost); // computes dCost/dW at every node of the graph
// 2) Dummy Training (only use a single data point)
double currentCost;
do
{
currentCost = Math.Abs(session.Run(cost, dico, false).ToArray().Sum());
Console.WriteLine($"cost: {currentCost}");
session.Run(optimizer, dico);
count++;
} while (currentCost > 1e-2);
Console.WriteLine($"{count}");
// Forward pass with original network
var result = session.Run(softmax, new Dictionary <string, Volume <double> > {
{ "x", x }
});
Console.WriteLine("probability that x is class 0: " + result.Get(0));
}
// 3) Serialization
softmax.Save("MyNetwork");
// 4) Deserialization
var deserialized = SerializationExtensions.Load <double>("MyNetwork", false)[0]; // first element is the model (second element is the cost if it was saved along)
using (var session = new Session <double>())
{
// Forward pass with deserialized network
var result = session.Run(deserialized, new Dictionary <string, Volume <double> > {
{ "x", x }
});
Console.WriteLine("probability that x is class 0: " + result.Get(0)); // This should give exactly the same result as previous network evaluation
}
Console.ReadLine();
}
public void GenerateText()
{
this._output = SerializationExtensions.Load <float>("MyNetwork", false)[0];
var initState = BuilderInstance <float> .Volume.SameAs(new Shape(this._stateSize, 1, 1, 1));
Volume <float> temperature = 1.0f; // increase this for more creativity (and more spelling mistakes)
Volume <float> dropoutProba = 0.0f;
var input = new List <Volume <float> >();
initState.Clear();
// Seed
var inputchar = new char[this._numSteps];
inputchar[0] = 'n';
inputchar[1] = 'o';
for (var i = 0; i < this._numSteps; i++)
{
Console.Write(inputchar[i]);
input.Add(BuilderInstance <float> .Volume.SameAs(new Shape(this._data.Vocabulary.Count, 1, 1, 1)));
}
do
{
using (var session = new Session <float>())
{
for (var i = 0; i < this._numSteps; i++)
{
input[i].Clear();
input[i].Set(this._data.Vocabulary.IndexOf(inputchar[i]), 0, 0, 0, 1);
}
var dico = new Dictionary <string, Volume <float> >
{
{ "initState", initState },
{ "temperature", temperature },
{ "dropoutProba", dropoutProba }
};
for (var i = 0; i < this._numSteps; i++)
{
dico["x" + (i + 1)] = input[i];
}
var result = session.Run(this._output, dico);
var c3 = ToChar(3, result);
Console.Write(c3);
for (var i = 1; i < this._numSteps; i++)
{
inputchar[i - 1] = inputchar[i];
}
inputchar[this._numSteps - 1] = c3;
initState = session.GetVariableByName(this._output, "initState").Result.Clone(); // re inject
}
} while (!Console.KeyAvailable);
}
/// <summary>
/// Solves y = x * W + b (CPU single version)
/// for y = 1 and x = -2
///
/// This also demonstrates how to save and load a graph
/// </summary>
public static void Example1()
{
var cns = new ConvNetSharp <float>();
// Graph creation
Op <float> cost;
Op <float> fun;
if (File.Exists("test.graphml"))
{
Console.WriteLine("Loading graph from disk.");
var ops = SerializationExtensions.Load <float>("test", true);
fun = ops[0];
cost = ops[1];
}
else
{
var x = cns.PlaceHolder("x");
var y = cns.PlaceHolder("y");
var W = cns.Variable(1.0f, "W", true);
var b = cns.Variable(2.0f, "b", true);
fun = x * W + b;
cost = (fun - y) * (fun - y);
}
var optimizer = new AdamOptimizer <float>(cns, 0.01f, 0.9f, 0.999f, 1e-08f);
using (var session = new Session <float>())
{
session.Differentiate(cost); // computes dCost/dW at every node of the graph
float currentCost;
do
{
var dico = new Dictionary <string, Volume <float> > {
{ "x", -2.0f }, { "y", 1.0f }
};
currentCost = session.Run(cost, dico);
Console.WriteLine($"cost: {currentCost}");
var result = session.Run(fun, dico);
session.Run(optimizer, dico);
} while (currentCost > 1e-5);
float finalW = session.GetVariableByName(fun, "W").Result;
float finalb = session.GetVariableByName(fun, "b").Result;
Console.WriteLine($"fun = x * {finalW} + {finalb}");
fun.Save("test", cost);
// Display graph
var vm = new ViewModel <float>(cost);
var app = new Application();
app.Run(new GraphControl {
DataContext = vm
});
}
Console.ReadKey();
}
private static void FacePresence()
{
var batchSize = 1000;
int width = 32;
int height = 32;
BuilderInstance <float> .Volume = new VolumeBuilder(); // For GPU
var imageLoader = new ImageLoader();
var randomImageLoader = new ImageLoader(true, 2);
// Load Dataset - Faces
var faces1 = LfwCropLoader.LoadDataset(@"..\..\..\Dataset\lfwcrop_grey", width, height);
var faces2 = imageLoader.LoadDataset(@"..\..\..\Dataset\custom\faces", width, height); // dump you own face images here
// Load Dataset - Non-faces
var nonFaces1 = randomImageLoader.LoadDataset(@"..\..\..\Dataset\scene_categories", width, height);
var nonFaces2 = randomImageLoader.LoadDataset(@"..\..\..\Dataset\TextureDatabase", width, height);
var nonFaces3 = randomImageLoader.LoadDataset(@"..\..\..\Dataset\cars_brad_bg", width, height);
var nonFaces4 = randomImageLoader.LoadDataset(@"..\..\..\Dataset\houses", width, height);
var nonFaces5 = imageLoader.LoadDataset(@"..\..\..\Dataset\custom\non_faces", width, height); // dump you own non-face images here
var facesDataset = new FaceDetectionDataset(width, height);
facesDataset.TrainSet.AddRange(faces1);
facesDataset.TrainSet.AddRange(faces2);
facesDataset.TrainSet.AddRange(nonFaces1);
facesDataset.TrainSet.AddRange(nonFaces2);
facesDataset.TrainSet.AddRange(nonFaces3);
facesDataset.TrainSet.AddRange(nonFaces4);
facesDataset.TrainSet.AddRange(nonFaces5);
Console.WriteLine(" Done.");
ConvNetSharp <float> cns;
// Model
Op <float> softmax = null;
if (File.Exists("FaceDetection.json"))
{
Console.WriteLine("Loading model from disk...");
softmax = SerializationExtensions.Load <float>("FaceDetection", false)[0]; // first element is the model (second element is the cost if it was saved along)
cns = softmax.Graph; // Deserialization creates its own graph that we have to use. TODO: make it simplier in ConvNetSharp
}
else
{
cns = new ConvNetSharp <float>();
}
var x = cns.PlaceHolder("x");
var dropProb = cns.PlaceHolder("dropProb");
if (softmax == null)
{
// Inspired by https://github.com/PCJohn/FaceDetect
var layer1 = cns.Relu(cns.Conv(x, 5, 5, 4, 2) + cns.Variable(new Shape(1, 1, 4, 1), "bias1", true));
var layer2 = cns.Relu(cns.Conv(layer1, 3, 3, 16, 2) + cns.Variable(new Shape(1, 1, 16, 1), "bias2", true));
var layer3 = cns.Relu(cns.Conv(layer2, 3, 3, 32) + cns.Variable(new Shape(1, 1, 32, 1), "bias3", true));
var flatten = cns.Flatten(layer3);
var dense1 = cns.Dropout(cns.Relu(cns.Dense(flatten, 600)) + cns.Variable(new Shape(1, 1, 600, 1), "bias4", true), dropProb);
var dense2 = cns.Dense(dense1, 2) + cns.Variable(new Shape(1, 1, 2, 1), "bias5", true);
softmax = cns.Softmax(dense2);
}
var y = cns.PlaceHolder("y");
// Cost
var cost = new SoftmaxCrossEntropy <float>(cns, softmax, y);
// Optimizer
var optimizer = new AdamOptimizer <float>(cns, 1e-4f, 0.9f, 0.999f, 1e-16f);
//if (File.Exists("loss.csv"))
//{
// File.Delete("loss.csv");
//}
Volume <float> trainingProb = 0.5f;
Volume <float> testingProb = 0.0f;
// Training
using (var session = new Session <float>())
{
session.Differentiate(cost); // computes dCost/dW at every node of the graph
var iteration = 0;
double currentCost;
do
{
var batch = facesDataset.GetBatch(batchSize);
var input = batch.Item1;
var output = batch.Item2;
var dico = new Dictionary <string, Volume <float> > {
{ "x", input }, { "y", output }, { "dropProb", trainingProb }
};
var stopwatch = Stopwatch.StartNew();
// session.Run(softmax, dico);
Debug.WriteLine(stopwatch.ElapsedMilliseconds);
currentCost = session.Run(cost, dico);
Console.WriteLine($"cost: {currentCost}");
File.AppendAllLines("loss.csv", new[] { currentCost.ToString(CultureInfo.InvariantCulture) });
session.Run(optimizer, dico);
if (iteration++ % 100 == 0)
{
// Test on a on random picture
var test = facesDataset.GetBatch(100);
dico = new Dictionary <string, Volume <float> > {
{ "x", test.Item1 }, { "dropProb", testingProb }
};
var result = session.Run(softmax, dico);
int correct = 0;
for (int i = 0; i < 100; i++)
{
var class0Prob = result.Get(0, 0, 0, i);
var class1Prob = result.Get(0, 0, 1, i);
if ((test.Item3[i].IsFace && class1Prob > class0Prob) || (!test.Item3[i].IsFace && class0Prob > class1Prob))
{
correct++;
}
}
Console.WriteLine($"Test: {correct}%");
File.AppendAllLines("accuracy.csv", new[] { correct.ToString() });
var filename = test.Item3[0].Filename;
softmax.Save("FaceDetection");
}
} while (currentCost > 1e-5 && !Console.KeyAvailable);
softmax.Save("FaceDetection");
}
}
private static void FaceLocalization()
{
BuilderInstance <float> .Volume = new VolumeBuilder(); // For GPU
// Load Dataset
var datasetPath = @"C:\Pro\Github\FaceLocalisation\Face\Dataset\Helen"; // contains folders from helen dataset (annotation, helen_1 ,..)
var batchSize = 5; // my GTX 760 cannot take more...
var dataSet = HelenLoader.LoadDataset(datasetPath, 256, 256, 50);
Console.WriteLine(" Done.");
ConvNetSharp <float> cns;
// Model
Op <float> yhat = null;
if (File.Exists("FaceDetection.json"))
{
Console.WriteLine("Loading model from disk...");
yhat = SerializationExtensions.Load <float>("FaceDetection", false)[0]; // first element is the model (second element is the cost if it was saved along)
cns = yhat.Graph; // Deserialization creates its own graph that we have to use. TODO: make it simplier in ConvNetSharp
}
else
{
cns = new ConvNetSharp <float>();
}
var x = cns.PlaceHolder("x");
if (yhat == null)
{
// Inspired by http://cs231n.stanford.edu/reports/2017/pdfs/222.pdf
var alpha = 0.2f;
var layer1 = cns.Dropout(cns.Pool(cns.LeakyRelu(cns.Conv(x, 3, 3, 32, 1, 1) + cns.Variable(new Shape(1, 1, 32, 1), "bias1", true), alpha), 2, 2, 0, 0, 2, 2), 0.1f);
var layer2 = cns.Pool(cns.LeakyRelu(cns.Conv(layer1, 3, 3, 64, 1, 1) + cns.Variable(new Shape(1, 1, 64, 1), "bias2", true), alpha), 2, 2, 0, 0, 2, 2);
var layer3 = cns.Dropout(cns.Pool(cns.LeakyRelu(cns.Conv(layer2, 3, 3, 128, 1, 1) + cns.Variable(new Shape(1, 1, 128, 1), "bias3", true), alpha), 2, 2, 0, 0, 2, 2),
0.1f);
var layer4 = cns.Pool(cns.LeakyRelu(cns.Conv(layer3, 3, 3, 64, 1, 1) + cns.Variable(new Shape(1, 1, 64, 1), "bias4", true), alpha), 2, 2, 0, 0, 2, 2);
var layer5 = cns.Pool(cns.LeakyRelu(cns.Conv(layer4, 3, 3, 16, 1, 1) + cns.Variable(new Shape(1, 1, 16, 1), "bias5", true), alpha), 2, 2, 0, 0, 2, 2);
var flatten = cns.Flatten(layer5);
var dense1 = cns.Conv(flatten, 1, 1, 128);
yhat = cns.Conv(dense1, 1, 1, 4);
//x.Evaluated += (sender, args) => { }; // I use this to place a break point and check Volume dimensions / debug
}
var y = cns.PlaceHolder("y");
// Cost
var cost = cns.Sum((yhat - y) * (yhat - y), Shape.From(1));
// Optimizer
var optimizer = new AdamOptimizer <float>(cns, 0.01f, 0.9f, 0.999f, 1e-08f);
if (File.Exists("loss.csv"))
{
File.Delete("loss.csv");
}
// Training
using (var session = new Session <float>())
{
session.Differentiate(cost); // computes dCost/dW at every node of the graph
var iteration = 0;
double currentCost;
do
{
var batch = dataSet.GetBatch(batchSize);
var input = batch.Item1;
var output = batch.Item2;
var dico = new Dictionary <string, Volume <float> > {
{ "x", input }, { "y", output }
};
currentCost = session.Run(cost, dico);
Console.WriteLine($"cost: {currentCost}");
File.AppendAllLines("loss.csv", new[] { currentCost.ToString(CultureInfo.InvariantCulture) });
session.Run(optimizer, dico);
if (iteration++ % 100 == 0)
{
// Test on a on random picture
var test = dataSet.GetBatch(1);
dico = new Dictionary <string, Volume <float> > {
{ "x", test.Item1 }
};
var result = session.Run(yhat, dico);
var image = (Bitmap)Image.FromFile(test.Item3[0].Filename);
BitmapTool.DrawBoundingBox(image, new BoundingBox {
x1 = result.Get(0), y1 = result.Get(1), x2 = result.Get(2), y2 = result.Get(3)
}, Color.Blue);
BitmapTool.DrawBoundingBox(image, test.Item3[0].BoundingBox, Color.Green); // correct answer
image.Save($"iteration_{iteration}.jpg");
yhat.Save("FaceDetection");
}
} while (currentCost > 1e-5 && !Console.KeyAvailable);
yhat.Save("FaceDetection");
}
}