private void Go()
{
var lijstje = DataSetGenerator.Generate();
int size = 500;
var img = new Image <Rgba32>(size, size);
MLImager.AddPointsToImage(img, lijstje);
using (var fs = new FileStream("output.png", FileMode.Create, FileAccess.Write, FileShare.Read))
{
img.SaveAsPng(fs);
}
lijstje.Shuffle();
var testData = lijstje.Take(lijstje.Count / 2).ToList();
var trainData = lijstje.Skip(lijstje.Count / 2).ToList();
foreach (var inp in INPUTS.INPUTSDICT)
{
state[inp.Key] = false;
}
state["x"] = true;
state["y"] = true;
//state["xTimesY"] = true;
//var networkShape = new List<int>() { 2, 1, 1 };
var networkShape = new List <int>()
{
2, 8, 8, 8, 8, 8, 8, 1
};
var inputIds = new List <string>()
{
"x", "y"
};
var network = NetworkBuilder.BuildNetwork(networkShape, Activations.TANH, Activations.TANH, null, inputIds, false);
var w = Stopwatch.StartNew();
for (int i = 0; i < 100000; i++)
{
foreach (var trainPoint in trainData)
{
var input = MLHelper.ConstructInput(state, trainPoint.X, trainPoint.Y);
NetworkBuilder.ForwardProp(network, input);
NetworkBuilder.BackProp(network, trainPoint.Label, Errors.SQUARE);
if ((i + 1) % batchSize == 0)
{
NetworkBuilder.UpdateWeights(network, learningrate, regularizationRate);
}
}
lossTrain = MLHelper.GetLoss(network, state, trainData);
lossTest = MLHelper.GetLoss(network, state, testData);
Console.WriteLine($"{i}: LossTrain: {lossTrain} LossTest: {lossTest}");
if (w.Elapsed.TotalSeconds > 1)
{
var img2 = MLImager.GenerateImage(network, state, size);
MLImager.AddPointsToImage(img2, lijstje);
try
{
using (var fs = new FileStream("output2.png", FileMode.Create, FileAccess.Write, FileShare.Read))
{
img2.SaveAsPng(fs);
}
}
catch (Exception ex)
{
}
w.Restart();
}
//for (int y = 1; y < network.Count; y++)
//{
// var layer = network[y];
// Console.WriteLine(y);
// foreach (var node in layer)
// {
// var sb = new StringBuilder();
// foreach (var link in node.InputLinks)
// {
// sb.Append($"Lnk: {link.Weight} ");
// }
// Console.WriteLine($" {sb.ToString()}");
// }
// Console.WriteLine();
//}
//double cccc = 0;
//foreach (var testPoint in testData)
//{
// var input = ConstructInput(testPoint.X, testPoint.Y);
// var result = NetworkBuilder.ForwardProp(network, input);
// var res = Math.Abs(result - testPoint.Label);
// cccc += res;
//}
//Console.WriteLine($"Res: {cccc}");
}
}
public static Network BuildNetwork()
{
var builder = new NetworkBuilder("LeNet")
.AddInputLayer(new InputLayerVertex("data"))
.AddLayerBlock(
b => b.AddLayer(new ConvolutionalLayerVertex("conv1", new ConvolutionalLayerParameter(
numberOfOutput: 64,
kernelSize: 7,
padding: 3,
stride: 2,
biasTerm: true)))
.AddActivation(new ActivationLayerVertex("conv1_relu"))
.AddBatchNorm(new BatchNormalizationLayerVertex("bn_conv1"))
.AddScale(new ScaleLayerVertex("scale_conv1")))
.AddLayer(new PoolingLayerVertex("pool1", new PoolingLayerParameter(
PoolingLayerKind.Max,
kernelSize: 3,
stride: 2)))
.AddResidualBlock(
left: b => b.AddLayerBlock(
lb => lb.AddLayer(new ConvolutionalLayerVertex("res2a_branch1", new ConvolutionalLayerParameter(
numberOfOutput: 256,
kernelSize: 1,
padding: 0,
stride: 1,
biasTerm: false)))
.AddBatchNorm(new BatchNormalizationLayerVertex("bn2a_branch1"))
.AddScale(new ScaleLayerVertex("scale2a_branch1"))),
right: b => b.AddLayerBlock(
lb => lb.AddLayer(new ConvolutionalLayerVertex("res2a_branch2a", new ConvolutionalLayerParameter(
numberOfOutput: 64,
kernelSize: 1,
padding: 0,
stride: 1,
biasTerm: false)))
.AddBatchNorm(new BatchNormalizationLayerVertex("bn2a_branch2a"))
.AddScale(new ScaleLayerVertex("scale2a_branch2a"))
.AddActivation(new ActivationLayerVertex("res2a_branch2a_relu")))
.AddLayerBlock(
lb => lb.AddLayer(new ConvolutionalLayerVertex("res2a_branch2b", new ConvolutionalLayerParameter(
numberOfOutput: 64,
kernelSize: 3,
padding: 1,
stride: 1,
biasTerm: false)))
.AddBatchNorm(new BatchNormalizationLayerVertex("bn2a_branch2b"))
.AddScale(new ScaleLayerVertex("scale2a_branch2b"))
.AddActivation(new ActivationLayerVertex("res2a_branch2b_relu")))
.AddLayerBlock(
lb => lb.AddLayer(new ConvolutionalLayerVertex("res2a_branch2c", new ConvolutionalLayerParameter(
numberOfOutput: 256,
kernelSize: 1,
padding: 0,
stride: 1,
biasTerm: false)))
.AddBatchNorm(new BatchNormalizationLayerVertex("bn2a_branch2c"))
.AddScale(new ScaleLayerVertex("scale2a_branch2c"))))
.AddEltwise(new EltwiseLayerVertex("res2a"))
.AddActivation(new ActivationLayerVertex("res2a_relu"));
return(builder.BuildNetwork());
}