static void Main(string[] args)
{
// 階層型ニューラルネットワークを構成
var nn = new HierarchicalNetwork(2, 2, 1);
nn[1, 0].Weights[0] = 10.0;
nn[1, 0].Weights[1] = 10.0;
nn[1, 0].Threshold = -5.0;
nn[1, 1].Weights[0] = 10.0;
nn[1, 1].Weights[1] = 10.0;
nn[1, 1].Threshold = -15.0;
nn[2, 0].Weights[0] = 10.0;
nn[2, 0].Weights[1] = -10.0;
nn[2, 0].Threshold = -5.0;
// 発火
nn.SetInputs(new ConstOutput(0.0), new ConstOutput(1.0));
nn.Fire();
// グラフを出力
var gen = new GraphGenerator();
using (var writer = new StreamWriter("graph.gv"))
{
gen.Generate(writer, nn);
}
}
static void Main(string[] args)
{
Func<double, double> f = ((double x) => 1 / (1 + Math.Exp(-x)));
Func<double, double> df = (x => f(x) * (1 - f(x)));
// ネットワークの情報読み込み
HierarchicalNetwork nn;
var examples = new List<Tuple<double[], double[]>>();
Console.Error.Write("path >");
var path = Console.ReadLine();
using (var reader = new StreamReader(path))
{
nn = new HierarchicalNetwork(reader.ReadLine().Split(' ').Select(v => int.Parse(v)).ToArray());
var exNum = int.Parse(reader.ReadLine());
for (var i = 0; i < exNum; ++i)
{
var io = reader.ReadLine().Split('|').ToList();
examples.Add(Tuple.Create(
io[0].Split(' ').Select(v => double.Parse(v)).ToArray(),
io[1].Split(' ').Select(v => double.Parse(v)).ToArray()
));
}
}
// 学習
var bp = new Backpropagation(nn, df, 0.8, 0.75);
var rnd = new Random();
using (var writer = new StreamWriter("error.csv"))
{
for (var i = 0; i < 1000; ++i)
{
var ex = examples.OrderBy(v => Guid.NewGuid()).ToArray(); // shuffle
var errSum = 0.0;
for (var j = 0; j < ex.Length; ++j)
{
bp.Train(ex[j].Item1, ex[j].Item2);
errSum += bp.Error;
}
writer.WriteLine("{0},{1}", i, errSum / ex.Length);
}
}
while (true)
{
Console.Error.Write("input >");
var inputs = Console.ReadLine().Split(' ').Select(v => new ConstOutput(double.Parse(v))).ToArray();
nn.SetInputs(inputs);
nn.Fire();
foreach (var o in nn.GetOutputs())
{
Console.WriteLine("{0:0.0########################################################}", o);
}
}
}
public Backpropagation(HierarchicalNetwork network, Func<double, double> diffTransferFunction, double learningRate, double stabilizationRate)
{
Network = network;
DiffTransferFunction = diffTransferFunction;
LearningRate = learningRate;
StabilizationRate = stabilizationRate;
Error = 0.0;
// allocation
var layers = Network.LayersCount;
delta = new double[layers][];
dw = new double[layers][][];
dh = new double[layers][];
delta[0] = new double[0];
dw[0] = new double[0][];
dh[0] = new double[0];
var prevCnt = Network.GetNeuronsCount(0);
for (var layer = 1; layer < layers; ++layer)
{
var cnt = Network.GetNeuronsCount(layer);
delta[layer] = new double[cnt];
dw[layer] = new double[cnt][];
dh[layer] = new double[cnt];
for (var i = 0; i < cnt; ++i)
{
delta[layer][i] = 0.0;
dw[layer][i] = new double[prevCnt];
dh[layer][i] = 0.0;
for (var j = 0; j < prevCnt; ++j)
{
dw[layer][i][j] = 0.0;
}
}
prevCnt = cnt;
}
}
public void Generate(StreamWriter writer, HierarchicalNetwork nn, string name = "hierarchical_nn")
{
writer.WriteLine("digraph {0} {{", name);
writer.WriteLine("\tgraph [rankdir=LR, ranksep=1.0];");
// subgraph in
writer.WriteLine("\tsubgraph in {");
writer.WriteLine("\t\tcolor=gray;");
writer.WriteLine("\t\tnode[style=invis, fixedsize=true, width=0, height=0];");
writer.WriteLine("\t\t{0};", string.Join(", ",
Enumerable.Range(0, nn.InputDimention)
.Select(i => string.Format("i{0}", i + 1))
));
writer.WriteLine("\t}");
// clusters
for (var i = 0; i < nn.LayersCount; ++i)
{
writer.WriteLine("\tsubgraph cluster{0} {{", i + 1);
writer.WriteLine("\t\tcolor=gray;");
if (i == 0)
{
writer.WriteLine("\t\tnode [label=\"\"]");
writer.WriteLine("\t\t{0};", string.Join(", ",
Enumerable.Range(0, nn.GetNeuronsCount(i))
.Select(j => string.Format("n{0}{1}", i + 1, j + 1))
));
}
else
{
for (var j = 0; j < nn.GetNeuronsCount(i); ++j)
{
writer.WriteLine("\t\tn{0}{1} [label=\"{2:0.00}\"];", i + 1, j + 1, -nn[i, j].Threshold);
}
}
writer.WriteLine("\t}");
}
// subgraph out
writer.WriteLine("\tsubgraph out {");
writer.WriteLine("\t\tnode[style=invis, fixedsize=true, width=0, height=0]");
writer.WriteLine("\t\t{0};", string.Join(", ",
Enumerable.Range(0, nn.OutputDimention)
.Select(i => string.Format("o{0}", i + 1))
));
writer.WriteLine("\t}");
// i? -> n1?
for (var i = 0; i < nn.InputDimention; ++i)
{
writer.WriteLine("\ti{0} -> n1{0} [arrowsize=0.6];", i + 1);
}
// n?? -> n??
for (var i = 0; i < nn.LayersCount - 1; ++i)
{
for (var j = 0; j < nn.GetNeuronsCount(i); ++j)
{
for (var k = 0; k < nn.GetNeuronsCount(i + 1); ++k)
{
if (k == 0)
{
writer.WriteLine("\tn{0}{1} -> n{2}{3} [taillabel=\"{4:0.00}\", headlabel=\"{5:0.00}\", arrowsize=0.6];",
i + 1, j + 1, i + 2, k + 1, nn[i, j].Output, nn[i + 1, k].Weights[j]);
}
else
{
writer.WriteLine("\tn{0}{1} -> n{2}{3} [headlabel=\"{4:0.00}\", arrowsize=0.6];",
i + 1, j + 1, i + 2, k + 1, nn[i + 1, k].Weights[j]);
}
}
}
}
// n?? -> o?
for (var i = 0; i < nn.OutputDimention; ++i)
{
writer.WriteLine("\tn{0}{1} -> o{1} [taillabel=\"{2:0.00}\", arrowsize=0.6];",
nn.LayersCount, i + 1, nn[nn.LayersCount - 1, i].Output);
}
writer.WriteLine("}");
}
