public void Train() { string dir = "ND_OPT_ConvAutoencoder_Data"; Directory.CreateDirectory($@"{dir}"); Directory.CreateDirectory($@"{dir}\Results"); Directory.CreateDirectory($@"{dir}\Sources"); AnimeDatasets a_dataset = new AnimeDatasets(StartSide, @"I:\Datasets\VAE_Dataset\White", @"I:\Datasets\VAE_Dataset\White\conv");//@"I:\Datasets\anime-faces\combined", @"I:\Datasets\anime-faces\combined_small"); a_dataset.InitializeDataset(); AnimeDatasets b_dataset = new AnimeDatasets(EndSide, @"I:\Datasets\VAE_Dataset\White", @"I:\Datasets\VAE_Dataset\White\conv");//@"I:\Datasets\anime-faces\combined", @"I:\Datasets\anime-faces\combined_small"); b_dataset.InitializeDataset(); Adam sgd = new Adam(0.001f); Quadratic quadratic = new Quadratic(); NRandom r = new NRandom(0); NRandom r2 = new NRandom(0); Matrix loss_deriv = new Matrix(OutputSize, 1, MemoryFlags.ReadWrite, true); #region Setup Database Matrix data_vec = new Matrix(LatentSize, 1, MemoryFlags.ReadOnly, false); Matrix[] a_dataset_vec = new Matrix[a_dataset.TrainingFiles.Count]; float[][] a_dataset_f = new float[a_dataset.TrainingFiles.Count][]; Matrix[] b_dataset_vec = new Matrix[a_dataset.TrainingFiles.Count]; float[][] b_dataset_f = new float[a_dataset.TrainingFiles.Count][]; for (int i = 0; i < a_dataset.TrainingFiles.Count; i++) { a_dataset_f[i] = new float[InputSize]; a_dataset_vec[i] = new Matrix(InputSize, 1, MemoryFlags.ReadOnly, false); a_dataset.LoadImage(a_dataset.TrainingFiles[i], a_dataset_f[i]); a_dataset_vec[i].Write(a_dataset_f[i]); b_dataset_f[i] = new float[OutputSize]; b_dataset_vec[i] = new Matrix(OutputSize, 1, MemoryFlags.ReadOnly, false); b_dataset.LoadImage(b_dataset.TrainingFiles[i], b_dataset_f[i]); b_dataset_vec[i].Write(b_dataset_f[i]); } #endregion for (int i0 = 000; i0 < 20000 * BatchSize; i0++) { int idx = (r.Next() % (a_dataset.TrainingFiles.Count / 2)); var out_img = superres_enc_front.ForwardPropagate(a_dataset_vec[idx]); quadratic.LossDeriv(out_img[0], b_dataset_vec[idx], loss_deriv, 0); superres_dec_back.ResetLayerErrors(); superres_dec_back.ComputeGradients(loss_deriv); superres_dec_back.ComputeLayerErrors(loss_deriv); superres_dec_back.UpdateLayers(sgd); loss_deriv.Clear(); if (i0 % BatchSize == 0) { a_dataset.SaveImage($@"{dir}\Sources\{i0 / BatchSize}.png", a_dataset_f[idx]); b_dataset.SaveImage($@"{dir}\Results\{i0 / BatchSize}.png", out_img[0].Read()); } Console.Clear(); Console.Write($"Iteration: {i0 / BatchSize}, Sub-Batch: {i0 % BatchSize}"); } superres_enc_front.Save($@"{dir}\network_final.bin"); Console.WriteLine("DONE."); }
public void Check() { front = InputLayer.Create(3, 1); back = ActivationLayer.Create <Sigmoid>(); conv = ConvLayer.Create(2, 1); fc = FCLayer.Create(1, 1); front.Append( conv.Append( ConvLayer.Create(3, 3, 2).Append( ActivationLayer.Create <LeakyReLU>().Append( fc.Append( FCLayer.Create(1, 1).Append( back )))))); front.SetupInternalState(); front.InitializeWeights(new UniformWeightInitializer(0, 0)); //ConstantWeightInitializer()); var lossFunc = new Quadratic(); var optimizer = new SGD(0.7f); Matrix x0 = new Matrix(9, 1, MemoryFlags.ReadWrite, false); x0.Write(new float[] { 0, 0, 0, 0, 0, 0, 0, 0, 0 }); Matrix x1 = new Matrix(9, 1, MemoryFlags.ReadWrite, false); x1.Write(new float[] { 0.1f, 0.2f, 0.3f, 0.4f, 0.5f, 0.6f, 0.7f, 0.8f, 0.9f }); Matrix x2 = new Matrix(9, 1, MemoryFlags.ReadWrite, false); x2.Write(new float[] { 0, 0, 0, 1, 1, 1, 1, 1, 1 }); Matrix x3 = new Matrix(9, 1, MemoryFlags.ReadWrite, false); x3.Write(new float[] { 1, 1, 1, 1, 1, 1, 1, 1, 1 }); var loss_vec = new Matrix(1, 1, MemoryFlags.ReadWrite, true); var x = new Matrix[] { x0, x1, x2, x3 }; var y = new Matrix[] { new Matrix(1, 1, MemoryFlags.ReadWrite, true), new Matrix(1, 1, MemoryFlags.ReadWrite, true), new Matrix(1, 1, MemoryFlags.ReadWrite, true), new Matrix(1, 1, MemoryFlags.ReadWrite, true), }; y[0].Write(new float[] { 0.53f }); y[1].Write(new float[] { 0.77f }); y[2].Write(new float[] { 0.88f }); y[3].Write(new float[] { 1.1f }); float delta = 1e-1f; float orig_loss_deriv = 0; float norm_conv = 0.0f, norm_conv_net = 0.0f; float norm_fc = 0.0f, norm_fc_net = 0.0f; for (int epoch = 0; epoch < 1; epoch++) { for (int idx = 1; idx < x.Length - 2; idx++) { { var output = front.ForwardPropagate(x[idx]); //Compute loss deriv loss_vec.Clear(); lossFunc.LossDeriv(output[0], y[idx], loss_vec, 0); orig_loss_deriv = loss_vec.Memory[0]; back.ComputeGradients(loss_vec); back.ComputeLayerErrors(loss_vec); } { //Save weights and apply deltas var conv_l = conv.CurrentLayer as ConvLayer; for (int f_i = 0; f_i < conv_l.FilterCnt; f_i++) { for (int i_i = 0; i_i < conv_l.InputDepth; i_i++) { for (int f_y = 0; f_y < conv_l.FilterSz; f_y++) { for (int f_x = 0; f_x < conv_l.FilterSz; f_x++) { var w_delta = conv_l.WeightErrors[f_i][i_i].Memory[f_y * conv_l.FilterSz + f_x]; conv_l.Weights[f_i][i_i].Memory[f_y * conv_l.FilterSz + f_x] += delta; var output = front.ForwardPropagate(x[idx]); loss_vec.Clear(); lossFunc.Loss(output[0], y[idx], loss_vec, 0); var y1 = loss_vec.Memory[0]; conv_l.Weights[f_i][i_i].Memory[f_y * conv_l.FilterSz + f_x] -= 2 * delta; output = front.ForwardPropagate(x[idx]); loss_vec.Clear(); lossFunc.Loss(output[0], y[idx], loss_vec, 0); var y0 = loss_vec.Memory[0]; conv_l.Weights[f_i][i_i].Memory[f_y * conv_l.FilterSz + f_x] += delta; var deriv = (y1 - y0) / (2 * delta); var norm = ((w_delta - deriv) * (w_delta - deriv)) / ((w_delta + deriv) * (w_delta + deriv)); norm_conv += norm; norm_conv_net++; } } } } var fc_l = fc.CurrentLayer as FCLayer; for (int i = 0; i < fc_l.Weights.Rows * fc_l.Weights.Columns; i++) { var w_delta = fc_l.WeightDelta.Memory[i]; fc_l.Weights.Memory[i] += delta; var output = front.ForwardPropagate(x[idx]); loss_vec.Clear(); lossFunc.Loss(output[0], y[idx], loss_vec, 0); var y1 = loss_vec.Memory[0]; fc_l.Weights.Memory[i] -= 2 * delta; output = front.ForwardPropagate(x[idx]); loss_vec.Clear(); lossFunc.Loss(output[0], y[idx], loss_vec, 0); var y0 = loss_vec.Memory[0]; fc_l.Weights.Memory[i] += delta; var deriv = (y1 - y0) / (2 * delta); var norm = ((w_delta - deriv) * (w_delta - deriv)) / ((w_delta + deriv) * (w_delta + deriv)); norm_fc += norm; norm_fc_net++; } } { back.ResetLayerErrors(); back.ComputeGradients(loss_vec); back.ComputeLayerErrors(loss_vec); back.UpdateLayers(optimizer); } } } Console.WriteLine($"Conv Norm {norm_conv / norm_conv_net}"); Console.WriteLine($"FC Norm {norm_fc / norm_fc_net}"); Console.ReadLine(); }
public void Train() { string dir = "GAN_Data"; Directory.CreateDirectory($@"{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."); }