class Solver(object): """Solver for training and testing StarGAN.""" def __init__(self, data_loader, config): """Initialize configurations.""" # Data loader. self.data_loader = data_loader # Model configurations. self.c_dim = config.c_dim self.image_size = config.image_size self.g_conv_dim = config.g_conv_dim self.d_conv_dim = config.d_conv_dim self.g_repeat_num = config.g_repeat_num self.d_repeat_num = config.d_repeat_num self.lambda_app = config.lambda_app self.lambda_pose = config.lambda_pose self.lambda_rec = config.lambda_rec self.lambda_gp = config.lambda_gp # Training configurations. # self.dataset = config.dataset self.batch_size = config.batch_size self.num_iters = config.num_iters self.num_iters_decay = config.num_iters_decay self.g_lr = config.g_lr self.d_lr = config.d_lr self.n_critic = config.n_critic self.beta1 = config.beta1 self.beta2 = config.beta2 self.resume_iters = config.resume_iters # self.selected_attrs = config.selected_attrs # loss functions self.feature_loss = torch.nn.CosineEmbeddingLoss() self.pose_loss = torch.nn.PairwiseDistance() # Test configurations. self.test_iters = config.test_iters # Miscellaneous. self.use_tensorboard = config.use_tensorboard self.device = torch.device( 'cuda' if torch.cuda.is_available() else 'cpu') # Directories. self.log_dir = config.log_dir self.sample_dir = config.sample_dir self.model_save_dir = config.model_save_dir self.result_dir = config.result_dir # Step size. self.log_step = config.log_step self.sample_step = config.sample_step self.model_save_step = config.model_save_step self.lr_update_step = config.lr_update_step # Build the model and tensorboard. self.build_model() if self.use_tensorboard: self.build_tensorboard() def build_model(self): """Create a generator and a discriminator.""" self.G = Generator(self.g_conv_dim, self.c_dim, self.g_repeat_num) self.D = Discriminator(self.image_size, self.d_conv_dim, self.c_dim, self.d_repeat_num) self.g_optimizer = torch.optim.Adam(self.G.parameters(), self.g_lr, [self.beta1, self.beta2]) self.d_optimizer = torch.optim.Adam(self.D.parameters(), self.d_lr, [self.beta1, self.beta2]) self.print_network(self.G, 'G') self.print_network(self.D, 'D') self.G.to(self.device) self.D.to(self.device) """Build the feature extractor""" self.feature_model = f_model(model_path=DUMPED_MODEL, freeze_param=True).cuda() #.cuda() self.feature_model.eval() def print_network(self, model, name): """Print out the network information.""" num_params = 0 for p in model.parameters(): num_params += p.numel() print(model) print(name) print("The number of parameters: {}".format(num_params)) def restore_model(self, resume_iters): """Restore the trained generator and discriminator.""" print( 'Loading the trained models from step {}...'.format(resume_iters)) G_path = os.path.join(self.model_save_dir, '{}-G.ckpt'.format(resume_iters)) D_path = os.path.join(self.model_save_dir, '{}-D.ckpt'.format(resume_iters)) self.G.load_state_dict( torch.load(G_path, map_location=lambda storage, loc: storage)) self.D.load_state_dict( torch.load(D_path, map_location=lambda storage, loc: storage)) def build_tensorboard(self): """Build a tensorboard logger.""" from logger import Logger self.logger = Logger(self.log_dir) def update_lr(self, g_lr, d_lr): """Decay learning rates of the generator and discriminator.""" for param_group in self.g_optimizer.param_groups: param_group['lr'] = g_lr for param_group in self.d_optimizer.param_groups: param_group['lr'] = d_lr def reset_grad(self): """Reset the gradient buffers.""" self.g_optimizer.zero_grad() self.d_optimizer.zero_grad() def normalize(self, x): pass def denorm(self, x): """Convert the range from [-1, 1] to [0, 1].""" out = (x + 1) / 2 return out.clamp_(0, 1) def gradient_penalty(self, y, x): """Compute gradient penalty: (L2_norm(dy/dx) - 1)**2.""" weight = torch.ones(y.size()).to(self.device) dydx = torch.autograd.grad(outputs=y, inputs=x, grad_outputs=weight, retain_graph=True, create_graph=True, only_inputs=True)[0] dydx = dydx.view(dydx.size(0), -1) dydx_l2norm = torch.sqrt(torch.sum(dydx**2, dim=1)) return torch.mean((dydx_l2norm - 1)**2) pass def feat_extract(self, resized_data): # input: N * 3 * 224 * 224 resized_data = resized_data.to(self.device) #.cuda() # output: N * num_classes, N * inter_dim, N * C' * 7 * 7 out, inter_out, x = self.feature_model(resized_data) # return batch feature vector return out.clamp_(0, 1) def pose_extract(self, batch_img, bbx=None): # input [N, 224, 224, 3] in RGB # output [N, 2, 18] # permute = [2, 1, 0] # # crop_batch = np.zeros(shape=()) # for i in range(batch_img.shape[0]): # img = crop(batch_img[i], bbx[i]) # # RGB TO BGR # img = img[:, permute, : , :] # crop_batch[i, :, :] = img # pose_vectors = get_batch_body_vector(crop_batch.numpy()) batch_img = np.transpose(batch_img, (0, 2, 3, 1)) # print(batch_img.shape) pose_vectors = get_batch_body_vector(batch_img) pose_vectors = torch.from_numpy(pose_vectors) return pose_vectors def appreance_cos_similarity(self, feat_fake, feat_real): label = torch.ones(feat_real.shape[0]).cuda() cos_simi = self.feature_loss(feat_real, feat_fake, label) return cos_simi def compute_pose_loss(self, pose_fake, pose_real, mask): mask = mask.view(-1, 1, 18).double() pose_fake = torch.mul(pose_fake, mask) pose_real = torch.mul(pose_real.double(), mask) distance = self.pose_loss(pose_real, pose_fake) return distance.float().sum() def train(self): """Train StarGAN within a single dataset.""" # Fetch fixed inputs for debugging. data_iter = iter(self.data_loader) a_fixed, b_fixed, bbox_fixed, b_fixed_pose_feat, mask_fixed = next( data_iter) a_fixed = a_fixed.to(self.device) b_fixed = b_fixed.to(self.device) bbox_fixed = bbox_fixed.to(self.device) # c_fixed_list = self.create_labels(c_org, self.c_dim, self.dataset, self.selected_attrs) # Learning rate cache for decaying. g_lr = self.g_lr d_lr = self.d_lr # Start training from scratch or resume training. start_iters = 0 if self.resume_iters: start_iters = self.resume_iters self.restore_model(self.resume_iters) # Start training. print('Start training...') start_time = time.time() for step in range(start_iters, self.num_iters): # =================================================================================== # # 1. Preprocess input data # # =================================================================================== # # Fetch real images and labels. try: a_real, b_real, bbox, b_pose_feat, mask = next(data_iter) except: data_iter = iter(self.data_loader) a_real, b_real, bbox, b_pose_feat, mask = next(data_iter) a_real = a_real.to(self.device) # Input images. b_real = b_real.to(self.device) bbox = bbox.to(self.device) b_pose_feat = b_pose_feat.to(self.device) mask = mask.to(self.device) # extract appearance feature a_app_feat = self.feat_extract(a_real) a_app_feat = a_app_feat.to(self.device) # # extract pose feature # b_pose_feat = self.pose_extract(b_real) # =================================================================================== # # 2. Train the discriminator # # =================================================================================== # # Compute loss with real images. out_src = self.D(b_real) d_loss_real = -torch.mean(out_src) # d_loss_cls = self.classification_loss(out_cls, label_org, self.dataset) # Compute loss with fake images. # con_feat = torch.cat([a_app_feat, bbox/416.0], dim=1) con_feat = a_app_feat x_fake = self.G(b_real, con_feat) out_src = self.D(x_fake.detach()) d_loss_fake = torch.mean(out_src) # fake_app_feat = self.feat_extract(x_fake) # fake_pose_feat = self.pose_extract(x_fake, bbox) # d_loss_app = self.appreance_cos_similarity(fake_app_feat, a_app_feat) # d_loss_pose = - self.pose_loss(fake_pose_feat, b_pose_feat) # Compute loss for gradient penalty. alpha = torch.rand(b_real.size(0), 1, 1, 1).to(self.device) x_hat = (alpha * b_real.data + (1 - alpha) * x_fake.data).requires_grad_(True) out_src = self.D(x_hat) d_loss_gp = self.gradient_penalty(out_src, x_hat) # Backward and optimize. # d_loss = d_loss_real + d_loss_fake + self.lambda_app * d_loss_cls + self.lambda_gp * d_loss_gp # d_loss = d_loss_fake + d_loss_real + self.lambda_app * d_loss_app + self.lambda_pose * d_loss_pose # d_loss = d_loss_fake + d_loss_real + self.lambda_gp * d_loss_gp d_loss = d_loss_fake + d_loss_real + self.lambda_gp * d_loss_gp self.reset_grad() d_loss.backward() self.d_optimizer.step() # Logging. loss = {} loss['D/loss_real'] = d_loss_real.item() loss['D/loss_fake'] = d_loss_fake.item() # loss['D/loss_app'] = d_loss_app.item() # loss['D/loss_pose'] = d_loss_pose.item() loss['D/loss_gp'] = d_loss_gp.item() # =================================================================================== # # 3. Train the generator # # =================================================================================== # if (step + 1) % self.n_critic == 0: # Original-to-target domain. x_fake = self.G(b_real, con_feat) # print(x_fake[0,:,200:205,200:205]) out_src = self.D(x_fake) g_loss_fake = -torch.mean(out_src) crop_batch = torch.zeros((x_fake.shape[0], 3, 224, 224)) b = bbox.detach().cpu().numpy().astype(int) for i in range(x_fake.shape[0]): # img = crop(x_fake[i], bbox[i]) x1, x2, y1, y2 = b[i, 0], b[i, 0] + b[i, 2], b[ i, 1], b[i, 1] + b[i, 3] x1 = min(max(x1, 0), 416) x2 = min(max(x2, 0), 416) y1 = min(max(y1, 0), 416) y2 = min(max(y2, 0), 416) img = x_fake[i, :, x1:x2, y1:y2].cpu().data.numpy() img = img.transpose((1, 2, 0)) resized_img = np.zeros(shape=(224, 224, 3)) resized_img = cv2.resize(img, (224, 224), interpolation=cv2.INTER_AREA) crop_batch[i, :, :, :] = torch.from_numpy( resized_img.transpose((2, 0, 1))) fake_app_feat = self.feat_extract(crop_batch) fake_pose_feat = self.pose_extract(crop_batch.numpy()) # #**** debug ****# # fake_images = (x_fake.cpu().data).numpy() # permute = [2, 1, 0] # fake_images = fake_images[:, permute, :, :].transpose((0,2,3,1)) # resized_data = np.zeros(shape=(fake_images.shape[0], 224, 224, 3)) # for j in range(fake_images.shape[0]): # resized_data[j,:,:,:] = cv2.resize(fake_images[j,:,:,:], (224, 224), interpolation = cv2.INTER_AREA) # resized_data = np.transpose(resized_data, (0, 3, 1, 2)) # resized_tensor = torch.from_numpy(resized_data) # resized_tensor = resized_tensor.to(self.device, dtype=torch.float) # fake_app_feat = self.feat_extract(resized_tensor) # fake_pose_feat = self.pose_extract(resized_data, bbox) fake_app_feat = fake_app_feat.to(self.device) fake_pose_feat = fake_pose_feat.to(self.device) #**** debug ****# # g_loss_cls = self.classification_loss(out_cls, label_trg, self.dataset) g_loss_app = -self.appreance_cos_similarity( fake_app_feat, a_app_feat) # -similarity # print(fake_pose_feat.size(), b_pose_feat.size(), mask.size()) g_loss_pose = self.compute_pose_loss(fake_pose_feat, b_pose_feat, mask) # joints distance # Backward and optimize. # g_loss = g_loss_fake + self.lambda_rec * g_loss_rec + self.lambda_app * g_loss_cls # g_loss = g_loss_fake + self.lambda_app * g_loss_app + self.lambda_pose * g_loss_pose g_loss = g_loss_fake + self.lambda_app * g_loss_app + self.lambda_pose * g_loss_pose self.reset_grad() g_loss.backward() self.g_optimizer.step() # Logging. loss['G/loss_fake'] = g_loss_fake.item() # loss['G/loss_rec'] = g_loss_rec.item() loss['G/loss_app'] = g_loss_app.item() * self.lambda_app loss['G/loss_pose'] = g_loss_pose.item() * self.lambda_pose # =================================================================================== # # 4. Miscellaneous # # =================================================================================== # # Print out training information. if (step + 1) % self.log_step == 0: et = time.time() - start_time et = str(datetime.timedelta(seconds=et))[:-7] log = "Elapsed [{}], Iteration [{}/{}]".format( et, step + 1, self.num_iters) for tag, value in loss.items(): log += ", {}: {:.4f}".format(tag, value) print(log) if self.use_tensorboard: for tag, value in loss.items(): self.logger.scalar_summary(tag, value, i + 1) # Translate fixed images for debugging. if (step + 1) % self.sample_step == 0: # if (step + 1) % 1 == 0: with torch.no_grad(): # a fix: [N, 3, 224, 224] # a_real, b_real, bbox, b_pose_feat, mask a_resized = torch.zeros(size=(a_real.shape[0], 3, 416, 416)) b_drawed = torch.zeros(size=(a_real.shape[0], 3, 416, 416)) for i in range(a_real.shape[0]): img = a_real[i].cpu().data.numpy() img = img.transpose((1, 2, 0)) resized_img = np.zeros(shape=(416, 416, 3)) resized_img = cv2.resize(img, (416, 416), interpolation=cv2.INTER_AREA) a_resized[i, :, :, :] = torch.from_numpy( resized_img.transpose((2, 0, 1))) trans1 = transforms.ToPILImage() trans2 = transforms.ToTensor() b_img = trans1(b_real[i].cpu()) draw = ImageDraw.Draw(b_img) b = bbox[i].cpu().data.numpy().astype(int) x, y, w, h = b x2, y2 = x + w, y + h draw.rectangle([x, y, x2, y2], outline="green", width=20) b_drawed[i, :, :, :] = trans2(b_img) b_drawed = b_drawed.to(self.device) a_resized = a_resized.to(self.device) picture_list = [a_resized, b_drawed] a_visual_feat = self.feat_extract(a_real) # a feature: [N, 20]; bbox: [N,4] # con_visual_feat = torch.cat([a_visual_feat, bbox/416.0], dim=1) # [N, 24] con_visual_feat = a_visual_feat # print(b_real, con_visual_feat) x_fake = self.G(b_real, con_visual_feat) # [N, 3, 416, 416] # print(a_fixed.size(), b_fixed.size(), x_fake.size()) picture_list.append(x_fake) picture_concat = torch.cat(picture_list, dim=0) # print(picture_concat.size()) sample_path = os.path.join( self.sample_dir, '{}-images.jpg'.format(step + 1)) save_image(self.denorm(picture_concat.data.cpu()), sample_path, nrow=4, padding=0) print('Saved real and fake images into {}...'.format( sample_path)) # Save model checkpoints. if (step + 1) % self.model_save_step == 0: G_path = os.path.join(self.model_save_dir, '{}-G.ckpt'.format(step + 1)) D_path = os.path.join(self.model_save_dir, '{}-D.ckpt'.format(step + 1)) torch.save(self.G.state_dict(), G_path) torch.save(self.D.state_dict(), D_path) print('Saved model checkpoints into {}...'.format( self.model_save_dir)) # Decay learning rates. if (step + 1) % self.lr_update_step == 0 and (step + 1) > ( self.num_iters - self.num_iters_decay): g_lr -= (self.g_lr / float(self.num_iters_decay)) d_lr -= (self.d_lr / float(self.num_iters_decay)) self.update_lr(g_lr, d_lr) print('Decayed learning rates, g_lr: {}, d_lr: {}.'.format( g_lr, d_lr)) def test(self): """Translate images using StarGAN trained on a single dataset.""" # Load the trained generator. self.restore_model(self.test_iters) # Set data loader. data_loader = self.data_loader with torch.no_grad(): for i, (a_real, b_real) in enumerate(data_loader): # Prepare input images and target domain labels. a_real = a_real.to(self.device) b_real = b_real.to(self.device) # Translate images. a_fake_list = [a_real, b_real] a_fixed_feat = self.feat_extract(a_real) a_fake_list.append(self.G(b_real, a_fixed_feat)) # Save the translated images. x_concat = torch.cat(a_fake_list, dim=3) result_path = os.path.join(self.result_dir, '{}-images.jpg'.format(i + 1)) save_image(self.denorm(x_concat.data.cpu()), result_path, nrow=1, padding=0) print('Saved real and fake images into {}...'.format( result_path))
def train_style_transfer(args): if not (args.train_data and args.valid_data): print("must chose train_data and valid_data") sys.exit() # make dataset trans = transforms.ToTensor() train_dataset = FaceDataset(args.train_data, transform=trans) label_dict = train_dataset.get_label_dict() valid_dataset = FaceDataset(args.valid_data, transform=trans) valid_dataset.give_label_dict(label_dict) train_loader = data_utils.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=1) valid_loader = data_utils.DataLoader(valid_dataset, batch_size=args.batch_size, shuffle=True, num_workers=1) train_size = len(train_dataset) valid_size = len(valid_dataset) loaders = {"train": train_loader, "valid": valid_loader} dataset_sizes = {"train": train_size, "valid": valid_size} if args.gpu: device = torch.device("cuda" if torch.cuda.is_available() else "cpu") else: device = torch.device("cpu") # make network if args.model_type == "VAE": net = Autoencoder(train_dataset.label_num()).to(device) optimizer = optim.Adam(net.parameters(), lr=args.lr, weight_decay=args.weight_decay) best_model_wts = net.state_dict() best_loss = 1e10 if args.generator_model and os.path.exists(args.generator_model): net.load_state_dict(torch.load(args.generator_model)) elif args.model_type == "VAEGAN": generator = Autoencoder(train_dataset.label_num()).to(device) discriminator = Discriminator().to(device) classifier = Classifier(train_dataset.label_num()).to(device) generator_optimizer = optim.Adam(generator.parameters(), lr=args.lr, weight_decay=args.weight_decay) discriminator_optimizer = optim.Adam(discriminator.parameters(), lr=args.lr * 0.1, weight_decay=args.weight_decay) best_generator_wts = generator.state_dict() best_discriminator_wts = discriminator.state_dict() best_generator_loss = 1e10 best_discriminator_loss = 1e10 if args.generator_model and os.path.exists(args.generator_model): generator.load_state_dict(torch.load(args.generator_model)) if args.discriminator_model and os.path.exists( args.discriminator_model): discriminator.load_state_dict(torch.load(args.discriminator_model)) if args.classifier_model: classifier.load_state_dict(torch.load(args.classifier_model)) # make loss function and optimizer criterion = nn.BCELoss(reduction="sum") classifier_criterion = nn.CrossEntropyLoss(reduction="sum") # initialize loss loss_history = {"train": [], "valid": []} # start training start_time = time.time() for epoch in range(args.epochs): print("epoch {}".format(epoch + 1)) for phase in ["train", "valid"]: if phase == "train": if args.model_type == "VAE": net.train(True) elif args.model_type == "VAEGAN": generator.train(True) discriminator.train(True) else: if args.model_type == "VAE": net.train(False) elif args.model_type == "VAEGAN": generator.train(False) discriminator.train(False) # initialize running loss generator_running_loss = 0.0 discriminator_running_loss = 0.0 for i, data in enumerate(loaders[phase]): inputs, label = data # wrap the in valiables if phase == "train": inputs = Variable(inputs).to(device) label = Variable(label).to(device) torch.set_grad_enabled(True) else: inputs = Variable(inputs).to(device) label = Variable(label).to(device) torch.set_grad_enabled(False) # zero gradients if args.model_type == "VAE": optimizer.zero_grad() mu, var, outputs = net(inputs, label) loss = loss_func(inputs, outputs, mu, var) if phase == "train": loss.backward() optimizer.step() generator_running_loss += loss.item() elif args.model_type == "VAEGAN": real_label = Variable( torch.ones((inputs.size()[0], 1), dtype=torch.float) - 0.2 * (torch.rand(inputs.size()[0], 1))).to(device) fake_label = Variable( torch.zeros((inputs.size()[0], 1), dtype=torch.float) + 0.2 * (torch.rand(inputs.size()[0], 1))).to(device) discriminator_optimizer.zero_grad() real_pred = discriminator(inputs) real_loss = criterion(real_pred, real_label) random_index = np.random.randint(0, train_dataset.label_num(), inputs.size()[0]) generate_label = Variable( torch.zeros_like(label)).to(device) for i, index in enumerate(random_index): generate_label[i][index] = 1 mu, var, outputs = generator(inputs, label) fake_pred = discriminator(outputs.detach()) fake_loss = criterion(fake_pred, fake_label) discriminator_loss = real_loss + fake_loss if phase == "train": discriminator_loss.backward() discriminator_optimizer.step() generator_optimizer.zero_grad() #class_loss = classifier_criterion(classifier(outputs), torch.max(label, 1)[1]) dis_loss = criterion(discriminator(outputs), real_label) gen_loss = loss_func(inputs, outputs, mu, var) generator_loss = dis_loss + gen_loss if phase == "train": generator_loss.backward() generator_optimizer.step() discriminator_running_loss += discriminator_loss.item() generator_running_loss += generator_loss.item() if args.model_type == "VAE": epoch_loss = generator_running_loss / dataset_sizes[ phase] * args.batch_size loss_history[phase].append(epoch_loss) print("{} loss {:.4f}".format(phase, epoch_loss)) if phase == "valid" and epoch_loss < best_loss: best_model_wts = net.state_dict() best_loss = epoch_loss elif args.model_type == "VAEGAN": epoch_generator_loss = generator_running_loss / dataset_sizes[ phase] * args.batch_size epoch_discriminator_loss = discriminator_running_loss / dataset_sizes[ phase] * args.batch_size print("{} generator loss {:.4f}".format( phase, epoch_generator_loss)) print("{} discriminator loss {:.4f}".format( phase, epoch_discriminator_loss)) if phase == "valid" and epoch_generator_loss < best_generator_loss: best_generator_wts = generator.state_dict() best_generator_loss = epoch_generator_loss if phase == "valid" and epoch_discriminator_loss < best_discriminator_loss: best_discriminator_wts = discriminator.state_dict() best_generator_loss = epoch_discriminator_loss elapsed_time = time.time() - start_time print("training complete in {:.0f}s".format(elapsed_time)) if args.model_type == "VAE": net.load_state_dict(best_model_wts) return net, label_dict elif args.model_type == "VAEGAN": generator.load_state_dict(best_generator_wts) discriminator.load_state_dict(best_discriminator_wts) return (generator, discriminator), label_dict
label.fill_(real_label) # fake labels are real for generator cost output = D(fake) errG = criterion(output, label) D_G_z2 = output.mean().item() errG.backward() optimizerG.step() # state_msg = f'{epoch+1:d}' state_msg = 'Epoch: %d Loss_D: %.4f Loss_G: %.4f D(x): %.4f D(G(z)): %.4f / %.4f' \ % (epoch, errD.item(), errG.item(), D_x, D_G_z1, D_G_z2) pbar.set_description(state_msg) writer.add_scalar('loss_D', errD, n_iter) writer.add_scalar('loss_G', errG, n_iter) writer.add_scalar('D(x)', D_x, n_iter) writer.add_scalar('D(G(z1))', D_G_z1, n_iter) writer.add_scalar('D(G(z2))', D_G_z2, n_iter) if n_iter % 100 == 0: fake = G(fixed_noise) real = vutils.make_grid(image, nrow=8, normalize=True, scale_each=True) writer.add_image('real', real, n_iter) fake = vutils.make_grid(fake.detach(), nrow=8, normalize=True, scale_each=True) writer.add_image('fake', fake, n_iter) n_iter += 1 # do checkpointing torch.save(G.state_dict(), f'{save_dir}/G_epoch_{epoch}.pth') torch.save(D.state_dict(), f'{save_dir}/D_epoch_{epoch}.pth')
def main(args): # check if cuda available # device = 'cuda:0' if torch.cuda.is_available() else 'cpu' # define dataset and dataloader # dir_name = "{}_{}_{}_{}_{}".format(watch.year, watch.month, watch.day, watch.hour, watch.minute) args = parse_args() log_dir = os.path.join(os.getcwd(), "log_runs") dataset = AirfoilDataset() airfoil_x = dataset.get_x() airfoil_dim = airfoil_x.shape[0] airfoil_dataloader = DataLoader(dataset, batch_size=8, shuffle=True) # hyperparameters latent_dim = 16 lr_dis = args.lr_dis # discriminator learning rate lr_gen = args.lr_gen # generator learning rate num_epochs = args.num_epochs disc_hidden_dim = args.dis_hidden gen_hidden_dim = args.gen_hidden # build the model dis = Discriminator(input_dim=airfoil_dim, hidden_dim=disc_hidden_dim, lr=lr_dis).to(device) gen = Generator(latent_dim=latent_dim, hidden_dim=gen_hidden_dim, airfoil_dim=airfoil_dim, lr=lr_gen).to(device) print("Distrminator model:\n", dis) print("Generator model:\n", gen) path = "./gan_model.ckpt" # train the GAN model save_model_interval = 10 disc_loss_arr = [] disc_fake_prob_std_arr = [] disc_fake_prob_arr = [] gen_loss_arr = [] gen_update_interval = 2 for epoch in range(num_epochs): epoch_disc_loss = 0.0 epoch_gen_loss = 0.0 disc_fake_prob = 0.0 disc_fake_prob_std = 0.0 num_batches = 0 for n_batch, (local_batch, __) in enumerate(airfoil_dataloader): num_batches += 1 y_real = local_batch.to(device) batch_size = y_real.shape[0] y_real = torch.tensor(y_real).float().to(device) # train generator z = torch.randn((batch_size, latent_dim)).to(device) # train discriminator with torch.no_grad(): gen_sample = gen(z) disc_real = dis(y_real) disc_fake = dis(gen_sample) loss_dis = torch.log(disc_real) + torch.log(1.0 - disc_fake) loss_dis = -0.5 * loss_dis.mean() dis.optimizer.zero_grad() loss_dis.backward() dis.optimizer.step() epoch_disc_loss += loss_dis.item() # train generator gen_sample = gen(z) fake_prob = dis(gen_sample) loss_gen = -torch.log(fake_prob) loss_gen = loss_gen.mean() if (num_batches % gen_update_interval == 0): gen.optimizer.zero_grad() loss_gen.backward() gen.optimizer.step() disc_fake_prob_std += torch.std(fake_prob).item() disc_fake_prob += fake_prob.clone().detach().mean().item() epoch_gen_loss += loss_gen.item() # print loss while training if (n_batch + 1) % 30 == 0: print( "Epoch: [{}/{}], Batch: {}, Discriminator loss: {}, Generator loss: {}" .format(epoch, num_epochs, n_batch, loss_dis.item(), loss_gen.item())) print("entry made") if ((epoch + 1) % save_model_interval == 0): saveGAN(dis.state_dict(), gen.state_dict(), path) disc_fake_prob_arr.append(disc_fake_prob / num_batches) disc_fake_prob_std_arr.append((disc_fake_prob_std / num_batches)) disc_loss_arr.append(epoch_disc_loss / num_batches) gen_loss_arr.append(epoch_gen_loss / num_batches) # test trained GAN model num_samples = 100 # create random noise noise = torch.randn((num_samples, latent_dim)).to(device) # noise = torch.randn((num_samples, noise_dim)).to(device) # generate airfoils gen_airfoils = gen(noise) # if 'cuda' in device: if device.type == 'cuda': gen_airfoils = gen_airfoils.detach().cpu().numpy() else: gen_airfoils = gen_airfoils.detach().numpy() # gen_airfoils = gen_airfoils.detach().cpu().numpy() # plot generated airfoils plot_airfoils(airfoil_x, gen_airfoils, "gan_generated_airfoils.png") plot_prop(disc_loss_arr, "disc_loss") plot_prop(gen_loss_arr, "gen_loss") plot_prop( disc_fake_prob_arr, "disc_fake_prob_lrd_{}_lrg_{}_epochs_{}".format( lr_dis, lr_gen, num_epochs), disc_fake_prob_std_arr)