public float GetWeight(int in_dim, int out_dim)
{
return((float)rng.NextGaussian(0, Math.Sqrt(2.0f / (in_dim + out_dim))));
}
public void Train()
{
string dir = "GAN_Data";
Directory.CreateDirectory($@"{dir}");
Directory.CreateDirectory($@"{dir}\Results");
Directory.CreateDirectory($@"{dir}\Sources");
Directory.CreateDirectory($@"{dir}\ResultsPRE");
Directory.CreateDirectory($@"{dir}\SourcesPRE");
#region GAN Variables
var sgd_disc = new Adam(0.0002f, 1e-6f);
var sgd_gen = new Adam(0.0002f, 1e-6f);
var sgd_dec = new Adam(0.0002f);
var fake_loss = new NamedLossFunction(NamedLossFunction.GANDiscFake, NamedLossFunction.GANDiscFake);
var real_loss = new NamedLossFunction(NamedLossFunction.GANDiscReal, NamedLossFunction.GANDiscReal);
var gen_loss = new NamedLossFunction(NamedLossFunction.GANGen, NamedLossFunction.GANGen);
var enc_loss = new Quadratic();
NRandom r_dataset = new NRandom(0);
NRandom r_latent = new NRandom(0);
Matrix data_vec = new Matrix(LatentSize, 1, MemoryFlags.ReadOnly, false);
Matrix d_real_loss = new Matrix(1, 1, MemoryFlags.ReadWrite, true);
Matrix d_fake_loss = new Matrix(1, 1, MemoryFlags.ReadWrite, true);
Matrix g_loss = new Matrix(1, 1, MemoryFlags.ReadWrite, true);
Matrix e_loss = new Matrix(OutputSize, 1, MemoryFlags.ReadWrite, true);
Matrix loss_reader = new Matrix(1, 1, MemoryFlags.ReadWrite, true);
float d_real_loss_f = 0;
float d_fake_loss_f = 0;
float g_loss_f = 0;
float d_real_class_f = 0, d_fake_class_f = 0, g_class_f = 0;
Matrix zero = new Matrix(1, 1, MemoryFlags.ReadWrite, true);
zero.Memory[0] = 0;
Matrix one = new Matrix(1, 1, MemoryFlags.ReadWrite, true);
one.Memory[0] = 1;
#endregion
#region Setup Database
AnimeDatasets dataset = new AnimeDatasets(StartSide, /*@"I:\Datasets\anime-faces\combined", @"I:\Datasets\anime-faces\combined_small");*/ @"I:\Datasets\VAE_Dataset\White", @"I:\Datasets\VAE_Dataset\White\conv");
dataset.InitializeDataset();
Matrix[] dataset_vec = new Matrix[dataset.TrainingFiles.Count];
float[][] dataset_f = new float[dataset.TrainingFiles.Count][];
for (int i = 0; i < dataset.TrainingFiles.Count; i++)
{
dataset_f[i] = new float[InputSize];
dataset_vec[i] = new Matrix(InputSize, 1, MemoryFlags.ReadOnly, false);
dataset.LoadImage(dataset.TrainingFiles[i], dataset_f[i]);
dataset_vec[i].Write(dataset_f[i]);
}
#endregion
//Pretrain generator
for (int i0 = 0; i0 < 5000; i0++)
{
int idx = 0;
idx = (r_dataset.Next() % dataset.TrainingFiles.Count);
var latent = encoder.ForwardPropagate(dataset_vec[idx]);
var res = generator.ForwardPropagate(latent);
e_loss.Clear();
enc_loss.LossDeriv(res[0], dataset_vec[idx], e_loss, 0.0f * sgd_dec.L2Val / sgd_dec.Net);
var enc_loss_v = generator_back.ComputeGradients(e_loss);
generator_back.ComputeLayerErrors(e_loss);
encoder_back.ComputeGradients(enc_loss_v);
encoder_back.ComputeLayerErrors(enc_loss_v);
sgd_dec.Update(0);
generator_back.UpdateLayers(sgd_gen);
encoder_back.UpdateLayers(sgd_dec);
if (i0 % BatchSize == 0)
{
dataset.SaveImage($@"{dir}\SourcesPRE\{i0 / BatchSize}.png", dataset_f[idx]);
dataset.SaveImage($@"{dir}\ResultsPRE\{i0 / BatchSize}.png", res[0].Read());
generator.Save($@"{dir}\pretrained_generator_fc.bin");
encoder.Save($@"{dir}\trained_encoder_fc.bin");
}
Console.Clear();
Console.WriteLine($"Iteration: {i0 / BatchSize} Sub-batch: {i0 % BatchSize}");
}
for (int i0 = 000; i0 < 5000 * BatchSize; i0++)
{
int idx = 0;
idx = (r_dataset.Next() % dataset.TrainingFiles.Count);
//Generate the fake data
for (int i1 = 0; i1 < LatentSize; i1++)
{
data_vec.Memory[i1] = (float)r_latent.NextGaussian(0, 1);//LatentSize;
}
var fake_result = generator.ForwardPropagate(data_vec);
if (i0 % BatchSize == 0)
{
dataset.SaveImage($@"{dir}\Sources\{i0 / BatchSize}.png", dataset_f[idx]);
dataset.SaveImage($@"{dir}\Results\{i0 / BatchSize}.png", fake_result[0].Read());
}
Console.Clear();
Console.WriteLine($"Iteration: {i0 / BatchSize} Sub-batch: {i0 % BatchSize}");
Console.WriteLine($"Discriminator Real Loss: {d_real_loss_f}\nDiscriminator Fake Loss: {d_fake_loss_f}\nGenerator Loss: {g_loss_f}\n");
Console.WriteLine($"Discriminator Real Prediction: {d_real_class_f}\nDiscriminator Fake Prediction: {d_fake_class_f}\nGenerator Prediction: {g_class_f}");
d_fake_loss.Clear();
d_real_loss.Clear();
g_loss.Clear();
zero.Memory[0] = (r_latent.Next() % 1000) / 10000f;
one.Memory[0] = 1 - (r_latent.Next() % 1000) / 10000f;
//Discriminator feed forward for real data
{
var d_real_class = discriminator.ForwardPropagate(dataset_vec[idx]);
real_loss.LossDeriv(d_real_class[0], one, d_real_loss, 0.01f * sgd_disc.L2Val / sgd_disc.Net);
var d_real_prop = discriminator_back.ComputeGradients(d_real_loss);
discriminator_back.ComputeLayerErrors(d_real_loss);
d_real_class_f = d_real_class[0].Memory[0];
real_loss.Loss(d_real_class[0], one, loss_reader, 0.01f * sgd_disc.L2Val / sgd_disc.Net);
d_real_loss_f = loss_reader.Memory[0];
loss_reader.Memory[0] = 0;
}
//Discriminator feed forward for fake data
{
var d_fake_class = discriminator.ForwardPropagate(fake_result);
fake_loss.LossDeriv(d_fake_class[0], zero, d_fake_loss, 0.01f * sgd_disc.L2Val / sgd_disc.Net);
var d_fake_prop = discriminator_back.ComputeGradients(d_fake_loss);
discriminator_back.ComputeLayerErrors(d_fake_loss);
d_fake_class_f = d_fake_class[0].Memory[0];
fake_loss.Loss(d_fake_class[0], zero, loss_reader, 0.01f * sgd_disc.L2Val / sgd_disc.Net);
d_fake_loss_f = loss_reader.Memory[0];
loss_reader.Memory[0] = 0;
}
//Update and reset discriminator
sgd_disc.Update(0);
discriminator_back.UpdateLayers(sgd_disc);
discriminator_back.ResetLayerErrors();
//Generate the fake data again
{
for (int i1 = 0; i1 < LatentSize; i1++)
{
data_vec.Memory[i1] = (float)r_latent.NextGaussian(0, 1);//LatentSize;
}
fake_result = generator.ForwardPropagate(data_vec);
var d_gen_class = discriminator.ForwardPropagate(fake_result);
//Compute discriminator crossentropy loss assuming fake is real and propagate
gen_loss.LossDeriv(d_gen_class[0], one, g_loss, 0.01f * sgd_gen.L2Val / sgd_gen.Net);
var d_err = discriminator_back.ComputeGradients(g_loss);
generator_back.ComputeGradients(d_err);
generator_back.ComputeLayerErrors(d_err);
g_class_f = d_gen_class[0].Memory[0];
gen_loss.Loss(d_gen_class[0], one, loss_reader, 0.01f * sgd_gen.L2Val / sgd_gen.Net);
g_loss_f = loss_reader.Memory[0];
loss_reader.Memory[0] = 0;
//Update generator
sgd_gen.Update(0);
generator_back.UpdateLayers(sgd_gen);
generator_back.ResetLayerErrors();
discriminator.ResetLayerErrors();
}
}
discriminator.Save($@"{dir}\network_final.bin");
Console.WriteLine("DONE.");
}
public void Train()
{
string dir = "GAN_Data";
Directory.CreateDirectory($@"{dir}");
Directory.CreateDirectory($@"{dir}\Results");
Directory.CreateDirectory($@"{dir}\Sources");
#region GAN Variables
var sgd_disc = new Adam(0.0002f, 1e-6f);
var sgd_gen = new Adam(0.0002f, 1e-6f);
var fake_loss = new NamedLossFunction(NamedLossFunction.GANDiscFake, NamedLossFunction.GANDiscFake);
var real_loss = new NamedLossFunction(NamedLossFunction.GANDiscReal, NamedLossFunction.GANDiscReal);
var gen_loss = new NamedLossFunction(NamedLossFunction.GANGen, NamedLossFunction.GANGen);
NRandom r_dataset = new NRandom(0);
NRandom r_latent = new NRandom(0);
Matrix data_vec = new Matrix(LatentSize, 1, MemoryFlags.ReadOnly, false);
Matrix d_real_loss = new Matrix(1, 1, MemoryFlags.ReadWrite, true);
Matrix d_fake_loss = new Matrix(1, 1, MemoryFlags.ReadWrite, true);
Matrix g_loss = new Matrix(1, 1, MemoryFlags.ReadWrite, true);
Matrix loss_reader = new Matrix(1, 1, MemoryFlags.ReadWrite, true);
float d_real_loss_f = 0;
float d_fake_loss_f = 0;
float g_loss_f = 0;
float d_real_class_f = 0, d_fake_class_f = 0, g_class_f = 0;
Matrix zero = new Matrix(1, 1, MemoryFlags.ReadWrite, true);
zero.Memory[0] = 0;
Matrix one = new Matrix(1, 1, MemoryFlags.ReadWrite, true);
one.Memory[0] = 1;
#endregion
#region Setup Database
Reader dataset = new Reader();
dataset.InitializeTraining();
Matrix[] imgs = dataset.TrainingImages;
#endregion
for (int i0 = 000; i0 < 1000 * BatchSize; i0++)
{
int idx = 0;
idx = (r_dataset.Next() % imgs.Length);
//Generate the fake data
for (int i1 = 0; i1 < LatentSize; i1++)
{
data_vec.Memory[i1] = (float)r_latent.NextGaussian(0, 1);//LatentSize;
}
var fake_result = generator.ForwardPropagate(data_vec);
if (i0 % BatchSize == 0)
{
SaveImage($@"{dir}\Sources\{i0 / BatchSize}.png", imgs[idx].Read(), 28);
SaveImage($@"{dir}\Results\{i0 / BatchSize}.png", fake_result[0].Read(), 28);
}
Console.Clear();
Console.WriteLine($"Iteration: {i0 / BatchSize} Sub-batch: {i0 % BatchSize}");
Console.WriteLine($"Discriminator Real Loss: {d_real_loss_f}\nDiscriminator Fake Loss: {d_fake_loss_f}\nGenerator Loss: {g_loss_f}\n");
Console.WriteLine($"Discriminator Real Prediction: {d_real_class_f}\nDiscriminator Fake Prediction: {d_fake_class_f}\nGenerator Prediction: {g_class_f}");
d_fake_loss.Clear();
d_real_loss.Clear();
g_loss.Clear();
zero.Memory[0] = (r_latent.Next() % 100) / 1000f;
one.Memory[0] = 1 - (r_latent.Next() % 100) / 1000f;
//Discriminator feed forward for real data
{
var d_real_class = discriminator.ForwardPropagate(imgs[idx]);
real_loss.LossDeriv(d_real_class[0], one, d_real_loss, 0.01f * sgd_disc.L2Val / sgd_disc.Net);
var d_real_prop = discriminator_back.ComputeGradients(d_real_loss);
discriminator_back.ComputeLayerErrors(d_real_loss);
d_real_class_f = d_real_class[0].Memory[0];
real_loss.Loss(d_real_class[0], one, loss_reader, 0.01f * sgd_disc.L2Val);
d_real_loss_f = loss_reader.Memory[0];
loss_reader.Memory[0] = 0;
}
//Discriminator feed forward for fake data
{
var d_fake_class = discriminator.ForwardPropagate(fake_result);
fake_loss.LossDeriv(d_fake_class[0], zero, d_fake_loss, 0.01f * sgd_disc.L2Val / sgd_disc.Net);
var d_fake_prop = discriminator_back.ComputeGradients(d_fake_loss);
discriminator_back.ComputeLayerErrors(d_fake_loss);
d_fake_class_f = d_fake_class[0].Memory[0];
fake_loss.Loss(d_fake_class[0], zero, loss_reader, 0.01f * sgd_disc.L2Val / sgd_disc.Net);
d_fake_loss_f = loss_reader.Memory[0];
loss_reader.Memory[0] = 0;
}
//Update and reset discriminator
sgd_disc.Update(0);
discriminator_back.UpdateLayers(sgd_disc);
discriminator_back.ResetLayerErrors();
//Generate the fake data again
{
for (int i1 = 0; i1 < LatentSize; i1++)
{
data_vec.Memory[i1] = (float)r_latent.NextGaussian(0, 1);//LatentSize;
}
fake_result = generator.ForwardPropagate(data_vec);
var d_gen_class = discriminator.ForwardPropagate(fake_result);
//Compute discriminator crossentropy loss assuming fake is real and propagate
gen_loss.LossDeriv(d_gen_class[0], one, g_loss, 0.01f * sgd_gen.L2Val / sgd_gen.Net);
var d_err = discriminator_back.ComputeGradients(g_loss);
generator_back.ComputeGradients(d_err);
generator_back.ComputeLayerErrors(d_err);
g_class_f = d_gen_class[0].Memory[0];
gen_loss.Loss(d_gen_class[0], one, loss_reader, 0.01f * sgd_gen.L2Val / sgd_gen.Net);
g_loss_f = loss_reader.Memory[0];
loss_reader.Memory[0] = 0;
//Update generator
sgd_gen.Update(0);
generator_back.UpdateLayers(sgd_gen);
generator_back.ResetLayerErrors();
discriminator.ResetLayerErrors();
}
}
discriminator.Save($@"{dir}\network_final.bin");
Console.WriteLine("DONE.");
}