public static void GAN()
{
//Generator genFunc = new Generator();
//Discriminator disc = new Discriminator();
//NeuralNetwork generator = new NeuralNetwork(new int[] { 100, 128, 28 * 28 }, new IActivationFunction[] { null, new Tanh(), new Sigmoid() }, genFunc);
NeuralNetwork discriminator = new NeuralNetwork(new int[] { 28 * 28, 128, 1 }, new IActivationFunction[] { null, new Tanh(), new Sigmoid() }, new CrossEntropy(), new SGD());
Random rng = new Random(0);
real_in = new float[50000][];
expected_vec = new float[50000][];
for (int i = 0; i < real_in.Length; i++)
{
real_in[i] = new float[28 * 28];
expected_vec[i] = new float[11];
readMNIST(i);
}
float[] gen_out = new float[11];
gen_out[10] = 0;
for (int q = 0; q < 30; q++)
{
float[] Gz = null;
for (int i = 0; i < 50000; i++)
{
//if (expected_vec[i][0] != 1)
// continue;
//for (int j = 0; j < generator_input.Length; j++)
// generator_input[j] = (float)rng.NextDouble() * 1;
//Gz = generator.Activate(generator_input);
//var Dg = discriminator.Activate(Gz);
var Dz = discriminator.Activate(real_in[i]);
//disc.DiscriminatorValue = Dg[0] - 1;
//discriminator.Train(Gz, new float[] { 0 }, 3f / 1000);
//disc.DiscriminatorValue = Dz[0] + 0.001f;
discriminator.Train(real_in[i]); //, new float[] { expected_vec[i][0] }, 3f / 1000
//generator.Train(generator_input, real_in[i], 3f / 5000);
//Console.WriteLine($"Discriminator Output: {Dx.}");
}
//imsave(Gz);
Console.WriteLine($"Epoch {q} Finished.");
}
for (int i = 0; i < 50; i++)
{
var Dx = discriminator.Activate(real_in[i]);
float max = Dx.Max();
for (int k = 0; k < Dx.Length; k++)
{
if (expected_vec[i][k] == 1)
{
Console.WriteLine($"Expected:{k}");
}
if (Dx[k] == max)
{
Console.WriteLine($"Discriminator: {k}");
}
}
Console.WriteLine();
}
}
public static void GAN()
{
//Generator genFunc = new Generator();
//Discriminator disc = new Discriminator();
NeuralNetwork generator = new NeuralNetwork(new int[] { 100, 128, 256, 256, 512, 2048, 48 * 48 * 3 }, new IActivationFunction[] { null, new Tanh(), new Tanh(), new Tanh(), new Tanh(), new Tanh(), new ScaledSoftStep() }, new CrossEntropy(), new SGD());
NeuralNetwork discriminator = new NeuralNetwork(new int[] { 48 * 48 * 3, 2048, 512, 256, 256, 128, 1 }, new IActivationFunction[] { null, new Tanh(), new Tanh(), new Tanh(), new Tanh(), new Tanh(), new ScaledSoftStep() }, new CrossEntropy(), new SGD());
Random rng = new Random(0);
real_in = new float[1000][];
for (int i = 0; i < real_in.Length; i++)
{
real_in[i] = new float[48 * 48 * 3];
readMaki(i);
}
float max_gen_res = float.NegativeInfinity;
for (int q = 0; q < 40; q++)
{
float[] Gz = null;
for (int i = 0; i < real_in.Length; i++)
{
for (int j = 0; j < generator_input.Length; j++)
{
generator_input[j] = (float)rng.NextDouble();// NextGaussian(0, System.Math.Pow(100, 0.5f)) * 1;
}
Gz = generator.Activate(generator_input);
//Train discriminator
var(Dz_a, Dz_z) = discriminator.Train(real_in[i]);
var(Dz_nabla_w, Dz_nabla_b) = discriminator.Solver.Solve(Dz_a, Dz_z, new Vector(new float[] { 1 }));
(discriminator.Weights, discriminator.Biases) = discriminator.Optimizer.Optimize(discriminator.Weights, discriminator.Biases, Dz_nabla_w, Dz_nabla_b, 3f / 100);
var(DGz_a, DGz_z) = discriminator.Train(Gz);
var(DGz_nabla_w, DGz_nabla_b) = discriminator.Solver.Solve(DGz_a, DGz_z, new Vector(new float[] { 0 }));
(discriminator.Weights, discriminator.Biases) = discriminator.Optimizer.Optimize(discriminator.Weights, discriminator.Biases, DGz_nabla_w, DGz_nabla_b, 3f / 100);
//Train Generator
//Now compute the error from the output of D, but don't change anything until G
NeuralNetwork chained = new NeuralNetwork(generator, discriminator);
var(Gz_a_ex, Gz_z_ex) = chained.Train(generator_input);
var(Gz_nabla_w_ex, Gz_nabla_b_ex) = chained.Solver.Solve(Gz_a_ex, Gz_z_ex, new Vector(new float[] { 1 }));
Matrix[] Gz_nabla_w = new ArraySegment <Matrix>(Gz_nabla_w_ex, 0, generator.LayerCount).Array;
Vector[] Gz_nabla_b = new ArraySegment <Vector>(Gz_nabla_b_ex, 0, generator.LayerCount).Array;
(generator.Weights, generator.Biases) = generator.Optimizer.Optimize(generator.Weights, generator.Biases, Gz_nabla_w, Gz_nabla_b, 3f / 50);
if ((i % 25 == 0))
{
imsave(Gz, i);
max_gen_res = DGz_a[2][0];
}
Console.WriteLine($"[{i}]Discriminator Real: {Dz_a.Last()[0]} Fake: {DGz_a.Last()[0]}");
}
imsave(Gz, q * 1000);
Console.WriteLine($"Epoch {q} Finished.");
}
SaveWeights(generator.Weights, "gen_weights.bin");
SaveBiases(generator.Biases, "gen_biases.bin");
SaveWeights(discriminator.Weights, "dis_weights.bin");
SaveBiases(discriminator.Biases, "dis_biases.bin");
}
public static void GAN()
{
//Generator genFunc = new Generator();
//Discriminator disc = new Discriminator();
NeuralNetwork generator = new NeuralNetwork(new int[] { 100, 128, 28 * 28 }, new IActivationFunction[] { null, new Tanh(), new Sigmoid() }, new CrossEntropy(), new SGD());
NeuralNetwork discriminator = new NeuralNetwork(new int[] { 28 * 28, 128, 1 }, new IActivationFunction[] { null, new Tanh(), new Sigmoid() }, new CrossEntropy(), new SGD());
Random rng = new Random(0);
real_in = new float[50000][];
expected_vec = new float[50000][];
for (int i = 0; i < real_in.Length; i++)
{
real_in[i] = new float[28 * 28];
expected_vec[i] = new float[11];
readMNIST(i);
}
float[] gen_out = new float[11];
gen_out[10] = 0;
float max_gen_res = 0;
for (int q = 0; q < 30; q++)
{
float[] Gz = null;
for (int i = 0; i < 50000; i++)
{
if (expected_vec[i][6] != 1)
{
continue;
}
for (int j = 0; j < generator_input.Length; j++)
{
generator_input[j] = (float)rng.NextGaussian(0, System.Math.Pow(28 * 28, 0.5f)) * 1;
}
Gz = generator.Activate(generator_input);
//Train discriminator
var(Dz_a, Dz_z) = discriminator.Train(real_in[i]);
var(Dz_nabla_w, Dz_nabla_b) = discriminator.Solver.Solve(Dz_a, Dz_z, new Vector(new float[] { 1 }));
(discriminator.Weights, discriminator.Biases) = discriminator.Optimizer.Optimize(discriminator.Weights, discriminator.Biases, Dz_nabla_w, Dz_nabla_b, 3f / 1000);
var(DGz_a, DGz_z) = discriminator.Train(Gz);
var(DGz_nabla_w, DGz_nabla_b) = discriminator.Solver.Solve(DGz_a, DGz_z, new Vector(new float[] { 0 }));
(discriminator.Weights, discriminator.Biases) = discriminator.Optimizer.Optimize(discriminator.Weights, discriminator.Biases, DGz_nabla_w, DGz_nabla_b, 3f / 1000);
//Train Generator
//Now compute the error from the output of D, but don't change anything until G
NeuralNetwork chained = new NeuralNetwork(generator, discriminator);
var(Gz_a_ex, Gz_z_ex) = chained.Train(generator_input);
var(Gz_nabla_w_ex, Gz_nabla_b_ex) = chained.Solver.Solve(Gz_a_ex, Gz_z_ex, new Vector(new float[] { 1 }));
Matrix[] Gz_nabla_w = new ArraySegment <Matrix>(Gz_nabla_w_ex, 0, generator.LayerCount).Array;
Vector[] Gz_nabla_b = new ArraySegment <Vector>(Gz_nabla_b_ex, 0, generator.LayerCount).Array;
(generator.Weights, generator.Biases) = generator.Optimizer.Optimize(generator.Weights, generator.Biases, Gz_nabla_w, Gz_nabla_b, 3f / 1000);
if (max_gen_res < DGz_a[2][0])
{
imsave(Gz);
max_gen_res = DGz_a[2][0];
}
Console.WriteLine($"[{i}]Discriminator Real: {Dz_a[2][0]} Fake: {DGz_a[2][0]}");
}
//imsave(Gz);
Console.WriteLine($"Epoch {q} Finished.");
}
for (int i = 0; i < 50; i++)
{
var Dx = discriminator.Activate(real_in[i]);
float max = Dx.Max();
for (int k = 0; k < Dx.Length; k++)
{
if (expected_vec[i][k] == 1)
{
Console.WriteLine($"Expected:{k}");
}
if (Dx[k] == max)
{
Console.WriteLine($"Discriminator: {k}");
}
}
Console.WriteLine();
}
}
