eta = str(datetime.timedelta(seconds=seconds)).split('.')[0] print( f'{i:3d} | loss: {loss:.4f} | t_total: {t_t:.3f} | t_data: {t_d:.3f} | t_forward: {t_f:.3f} | ' f't_loss: {t_l:.3f} | t_backward: {t_b:.3f} | t_update: {t_u:.3f} | lr: {lr:.5f} | ETA: {eta}' ) if i > 0 and i % 100 == 0: writer.add_scalar('loss', loss, global_step=i) i += 1 if i > cfg.iter: training = False if cfg.val_interval > 0 and i % cfg.val_interval == 0: save_name = f'{cfg.model}_{i}_{cfg.lr}.pth' torch.save(model.state_dict(), f'weights/{save_name}') print(f'Model saved as: {save_name}, begin validating.') timer.reset() cfg.mode = 'Val' cfg.to_val_aug() model.eval() miou = validate(model, cfg) model.train() writer.add_scalar('miou', miou, global_step=i) timer.start() writer.close()
def train(): startTime = time.time() args = parameters.parse_arguments() logging.basicConfig(filename=args.logfile, level=logging.INFO) logging.critical("\n\n" + args.log_header) logging.info(args) device = ("cuda" if torch.cuda.is_available() else "cpu") logging.info(f"TIME: {time.time() - startTime}s Using device {device}") logging.info(f"TIME: {time.time()-startTime}s Loading dataset") try: with open(os.path.join(args.datadir, "data.pkl"), "rb") as f: data = pickle.load(f) except: data = DataLoader(args.datadir, int(args.batchsize), shuffle=int(args.shuffle)) with open(os.path.join(args.datadir, "data.pkl"), "wb") as f: pickle.dump(data, f) data.batchSize = int(args.batchsize) logging.info(f"TIME: {time.time()-startTime}s Dataset Loaded") random.seed(args.seed) indices = list(range(len(data))) random.shuffle( indices ) # 0:floor((1-validationFrac)*len(data)) will be training data, rest will be validation data trainEndIndex = math.floor((1 - args.validation_frac) * (len(data))) model = UNet(in_channels=1, num_classes=2, start_filts=int(args.conv_filters), up_mode=args.mode, depth=int(args.depth), batchnorm=args.batchnorm) model.reset_params() model = model.to(device) optimizer = None if args.optimizer == 'adam': optimizer = optim.Adam(model.parameters(), lr=args.lrstart) logging.info(f"TIME: {time.time()-startTime}s Optimizer: adam") elif args.optimizer == 'sgd': optimizer = optim.SGD(model.parameters(), lr=args.lrstart, momentum=args.momentum) logging.info(f"TIME: {time.time()-startTime}s Optimizer: SGD") elif args.optimizer == 'rmsprop': optimizer = optim.RMSprop(model.parameters(), lr=args.lrstart) logging.info(f"TIME: {time.time()-startTime}s Optimizer: RMSProp") else: logging.error( f"TIME: {time.time()-startTime}s Incorrect optimizer given") scheduler = [] if args.lrscheduler == "steplr": scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=args.decay) logging.info(f"TIME: {time.time()-startTime}s LRScheduler: StepLR") elif args.lrscheduler == "exponentiallr": scheduler = optim.lr_scheduler.ExponentialLR(optimizer, gamma=args.decay) logging.info( f"TIME: {time.time()-startTime}s LRScheduler: exponentialLR") else: scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=int(args.epochs)) logging.info( f"TIME: {time.time()-startTime}s LRScheduler: lr shouldn't change with epochs" ) criteria = CombinedLoss(args.lambda_loss, args.loss_type) diceCoeff = DiceLoss() TL = [] VL = [] if not os.path.exists(os.path.join(os.getcwd(), "loss_files")): os.makedirs(os.path.join(os.getcwd(), "loss_files")) lossFile = open(os.path.join("loss_files", args.log_header + ".csv"), "w+") lossFile.write("Epoch,TrainLoss,ValidationLoss,Dice Coefficient\n") for epoch in tqdm(range(1, int(args.epochs) + 1), desc="Training model"): trainLoss = 0 valLoss = 0 trainingSample = 0 testSample = 0 netCoeff = 0 for i in range(len(data)): images, masks = data[i] images = torch.tensor(images.astype(np.float32)) masks = torch.tensor(masks.astype(np.float32)) images = images.to(device) masks = masks.to(device) images = torch.transpose(images, 1, 3) masks = torch.transpose(masks, 1, 3) if i in indices[:trainEndIndex]: trainingSample += images.shape[0] networkPred = model(images) if args.regularization == 'l1': reg = L1_regularization(model, args.reg_lamda1) loss = criteria(masks, networkPred) + reg elif args.regularization == 'l1l2': reg = L1L2_regularization(model, args.reg_lamda1, args.reg_lamda2) loss = criteria(masks, networkPred) + reg else: loss = criteria(masks, networkPred) loss.backward() trainLoss += loss.item() optimizer.step() model.zero_grad() else: with torch.no_grad(): testSample += images.shape[0] prediction = model(images) if (epoch % args.save_epochs == 0) or (epoch == 1) or (epoch == args.epochs): imgPath = os.path.join("validation_sample", args.log_header, f"epoch {epoch}") if not os.path.exists(imgPath): os.makedirs(imgPath) hrt = images[0, 0, :, :].to("cpu") plt.imshow(np.array(hrt), cmap='gray') plt.title("Heart Image") plt.savefig(os.path.join(imgPath, "heart.png")) plt.clf() # ax = figure.add_subplot(232, title="Mask 1 Predicted") msk1 = prediction[0, 0, :, :].to("cpu") plt.imshow(np.array(msk1), cmap='gray') plt.title("Predicted Mask 1") plt.savefig(os.path.join(imgPath, "pred-mask1.png")) plt.clf() # ax = figure.add_subplot(231, title="Mask 2 Predicted") msk2 = prediction[0, 1, :, :].to("cpu") plt.imshow(np.array(msk2), cmap='gray') plt.title("Predicted Mask 2") plt.savefig(os.path.join(imgPath, "pred-mask2.png")) plt.clf() msk = np.zeros((192, 192, 3)) msk[:, :, 0] = np.array(msk1) msk[:, :, 1] = np.array(msk2) plt.imshow(np.array(hrt), cmap='gray') plt.imshow(msk, cmap='jet', alpha=0.4) plt.title("predicted-RV") plt.savefig(os.path.join(imgPath, "pred-RV.png")) plt.clf() # ax = figure.add_subplot(231, title="Mask 2 Real") msk1 = masks[0, 0, :, :].to("cpu") plt.imshow(np.array(msk1), cmap='gray') plt.title("Actual Mask 1") plt.savefig(os.path.join(imgPath, "actual-mask1.png")) plt.clf() # ax = figure.add_subplot(231, title="Mask 2 Real") msk2 = masks[0, 1, :, :].to("cpu") plt.imshow(np.array(msk2), cmap='gray') plt.title("Actual Mask 2") plt.savefig(os.path.join(imgPath, "actual-mask2.png")) plt.clf() # plt.savefig(os.path.join("validation_sample", f"{args.log_header}-epoch {epoch}.png")) msk = np.zeros((192, 192, 3)) msk[:, :, 0] = np.array(msk1) msk[:, :, 1] = np.array(msk2) plt.imshow(np.array(hrt), cmap='gray') plt.imshow(msk, cmap='jet', alpha=0.4) plt.title("actual-RV") plt.savefig(os.path.join(imgPath, "actual-RV.png")) plt.clf() if args.regularization == 'l1': reg = L1_regularization(model, args.reg_lamda1) loss = criteria(masks, prediction) + reg elif args.regularization == 'l1l2': reg = L1L2_regularization(model, args.reg_lamda1, args.reg_lamda2) loss = criteria(masks, prediction) + reg else: loss = criteria(masks, prediction) valLoss += loss.item() coeff = diceCoeff(masks, prediction) netCoeff += torch.sum(1 - coeff).item() if (epoch % int(args.save_epochs) == 0) or (epoch == int(args.epochs)): if not os.path.exists(args.model_save_dir): os.makedirs(args.model_save_dir) # save model torch.save( { "epoch": epoch, "model_state_dict": model.state_dict(), "optimizer_state_dict": optimizer.state_dict(), }, os.path.join(args.model_save_dir, f"model-epoch({epoch}).hdf5")) logging.info( f"TIME: {time.time()-startTime}s Model state saved for epoch: {epoch}" ) logging.info( f"TIME: {time.time()-startTime}s TRAINING: Epoch: {epoch}, lr: {scheduler.get_last_lr()}, loss: {trainLoss/(2*trainingSample)}" ) logging.info( f"TIME: {time.time()-startTime}s VALIDATION: Epoch: {epoch}, lr: {scheduler.get_last_lr()}, loss: {valLoss/(2*testSample)}" ) TL.append(trainLoss / (2 * trainingSample)) VL.append(valLoss / (2 * testSample)) lossFile.write( f"{epoch},{trainLoss/(2*trainingSample)},{valLoss/(2*testSample)},{netCoeff/(2*testSample)}\n" ) scheduler.step( ) # https://www.deeplearningwizard.com/deep_learning/boosting_models_pytorch/lr_scheduling/ plt.plot(list(range(1, int(args.epochs) + 1)), TL, label="Training loss") plt.plot(list(range(1, int(args.epochs) + 1)), VL, label="Validation loss") plt.xlabel("Epoch") plt.ylabel("Loss") plt.legend(loc="best") if not os.path.exists(os.path.join(os.getcwd(), "plots")): os.makedirs(os.path.join(os.getcwd(), "plots")) plt.savefig(os.path.join("plots", args.log_header + ".png"))
CURRENT_EPOCH = 0 for epoch in range(CURRENT_EPOCH, N_EPOCHS): train_loss = train_iter(MODEL_NAME, 470, epoch, model, train_dataloader, criterion, optimizer, device=device, writer=writer, log_every=10) writer.add_scalar('EpochLoss/train', train_loss, epoch) with open('{}_train_{}.pth'.format(MODEL_NAME, epoch), 'wb') as fp: torch.save(model.state_dict(), fp) val_loss = validate(epoch, model, val_dataloader, criterion, device=device, writer=writer, log_every=20) writer.add_scalar('EpochLoss/val', val_loss, epoch) print('Epoch #{:2}:\ttrain loss: {:5.5}\tval loss: {:5.5}'.format( epoch, train_loss, val_loss)) if val_loss < best_val_loss: best_val_loss = val_loss
if args.check_diff: disable_dropout(model) if args.check_runtime: graph = Graph.create(model, input_shape=(3, height, width)) model.cuda() solvert = -1 if len(args.solution_file) > 0: solver_info, solution = load_solution(args.solution_file) else: input_ = torch.randn((bs, 3, height, width)).cuda() solver_info = SolverInfo(bs=bs, model_name=model_name, mode=mode) solver_info.extract( graph, input_, *list(model.state_dict(keep_vars=True).values())) solver_model = Model(solver_info, budget, args.solver, args.ablation) t0 = time() solution = solver_model.solve() solvert = time() - t0 del input_ torch.cuda.empty_cache() torch.cuda.reset_max_memory_allocated() input_ = torch.randn((bs, 3, height, width)).cuda() schedule = Schedule(graph, solver_info) schedule.init_schedule(solution, mode) torch.cuda.synchronize() start_event_monet = torch.cuda.Event(enable_timing=True)
predictions = torch.argmax(outputs, 1) running_test_iou += metrics.jaccard_similarity_score( labels.cpu().numpy().flatten(), predictions.cpu().numpy().flatten()) # Save results pred = predictions.cpu().numpy()[0] lab = labels.cpu().numpy()[0] print(np.amax(pred), np.amin(pred), np.amax(lab), np.amin(lab)) scipy.misc.imsave('checkpoints/pred_{}.png'.format(batch), pred) scipy.misc.imsave('checkpoints/lab_{}.png'.format(batch), lab) test_iou.append(running_test_iou / len(test_loader)) print('Test IoU: {:.3f}'.format(test_iou[-1])) # Checkpointing torch.save( { 'model': model.state_dict(), 'optimizer': optimizer.state_dict(), 'scheduler': scheduler.state_dict(), 'epoch': epoch, 'train_loss': train_loss, 'train_iou': train_iou, 'test_iou': test_iou, }, checkpoint_file)
dest='load', default=False, help='load file model') (options, args) = parser.parse_args() if (options.model == 1): net = UNet(3, 1) if options.load: net.load_state_dict(torch.load(options.load)) print('Model loaded from {}'.format(options.load)) if options.gpu: net.cuda() cudnn.benchmark = True try: train_net(net, options.epochs, options.batchsize, options.lr, gpu=options.gpu) except KeyboardInterrupt: torch.save(net.state_dict(), 'INTERRUPTED.pth') print('Saved interrupt') try: sys.exit(0) except SystemExit: os._exit(0)
def train(input_data_type, grade, seg_type, num_classes, batch_size, epochs, use_gpu, learning_rate, w_decay, pre_trained=False): logger.info('Start training using {} modal.'.format(input_data_type)) model = UNet(4, 4, residual=True, expansion=2) criterion = nn.CrossEntropyLoss() optimizer = optim.Adam(params=model.parameters(), lr=learning_rate, weight_decay=w_decay) if pre_trained: checkpoint = torch.load(pre_trained_path, map_location=device) model.load_state_dict(checkpoint['model_state_dict']) if use_gpu: ts = time.time() model.to(device) print("Finish cuda loading, time elapsed {}".format(time.time() - ts)) scheduler = lr_scheduler.StepLR( optimizer, step_size=step_size, gamma=gamma) # decay LR by a factor of 0.5 every 5 epochs data_set, data_loader = get_dataset_dataloader(input_data_type, seg_type, batch_size, grade=grade) since = time.time() best_model_wts = copy.deepcopy(model.state_dict()) best_iou = 0.0 epoch_loss = np.zeros((2, epochs)) epoch_acc = np.zeros((2, epochs)) epoch_class_acc = np.zeros((2, epochs)) epoch_mean_iou = np.zeros((2, epochs)) evaluator = Evaluator(num_classes) def term_int_handler(signal_num, frame): np.save(os.path.join(score_dir, 'epoch_accuracy'), epoch_acc) np.save(os.path.join(score_dir, 'epoch_mean_iou'), epoch_mean_iou) np.save(os.path.join(score_dir, 'epoch_loss'), epoch_loss) model.load_state_dict(best_model_wts) logger.info('Got terminated and saved model.state_dict') torch.save(model.state_dict(), os.path.join(score_dir, 'terminated_model.pt')) torch.save( { 'model_state_dict': model.state_dict(), 'optimizer_state_dict': optimizer.state_dict() }, os.path.join(score_dir, 'terminated_model.tar')) quit() signal.signal(signal.SIGINT, term_int_handler) signal.signal(signal.SIGTERM, term_int_handler) for epoch in range(epochs): logger.info('Epoch {}/{}'.format(epoch + 1, epochs)) logger.info('-' * 28) for phase_ind, phase in enumerate(['train', 'val']): if phase == 'train': model.train() logger.info(phase) else: model.eval() logger.info(phase) evaluator.reset() running_loss = 0.0 running_dice = 0.0 for batch_ind, batch in enumerate(data_loader[phase]): imgs, targets = batch imgs = imgs.to(device) targets = targets.to(device) # zero the learnable parameters gradients optimizer.zero_grad() with torch.set_grad_enabled(phase == 'train'): outputs = model(imgs) loss = criterion(outputs, targets) if phase == 'train': loss.backward() optimizer.step() preds = torch.argmax(F.softmax(outputs, dim=1), dim=1, keepdim=True) running_loss += loss * imgs.size(0) logger.debug('Batch {} running loss: {:.4f}'.format(batch_ind,\ running_loss)) # test the iou and pixelwise accuracy using evaluator preds = torch.squeeze(preds, dim=1) preds = preds.cpu().numpy() targets = targets.cpu().numpy() evaluator.add_batch(targets, preds) epoch_loss[phase_ind, epoch] = running_loss / len(data_set[phase]) epoch_acc[phase_ind, epoch] = evaluator.Pixel_Accuracy() epoch_class_acc[phase_ind, epoch] = evaluator.Pixel_Accuracy_Class() epoch_mean_iou[phase_ind, epoch] = evaluator.Mean_Intersection_over_Union() logger.info('{} loss: {:.4f}, acc: {:.4f}, class acc: {:.4f}, mean iou: {:.6f}'.format(phase,\ epoch_loss[phase_ind, epoch],\ epoch_acc[phase_ind, epoch],\ epoch_class_acc[phase_ind, epoch],\ epoch_mean_iou[phase_ind, epoch])) if phase == 'val' and epoch_mean_iou[phase_ind, epoch] > best_iou: best_iou = epoch_mean_iou[phase_ind, epoch] best_model_wts = copy.deepcopy(model.state_dict()) if phase == 'val' and (epoch + 1) % 10 == 0: logger.info('Saved model.state_dict in epoch {}'.format(epoch + 1)) torch.save( model.state_dict(), os.path.join(score_dir, 'epoch{}_model.pt'.format(epoch + 1))) print() time_elapsed = time.time() - since logger.info('Training completed in {}m {}s'.format(int(time_elapsed / 60),\ int(time_elapsed) % 60)) # load best model weights model.load_state_dict(best_model_wts) # save numpy results np.save(os.path.join(score_dir, 'epoch_accuracy'), epoch_acc) np.save(os.path.join(score_dir, 'epoch_mean_iou'), epoch_mean_iou) np.save(os.path.join(score_dir, 'epoch_loss'), epoch_loss) return model, optimizer
budget = float(args.budget) bs = int(args.bs) model_name = args.model.split(".")[-1][:-2] mode = args.mode time_limit = args.time_limit print("Memory budget ", budget, " GB") print("Batch size ", bs) print("Model", model_name) print("Mode", mode) if args.model == 'unet': height, width = 416, 608 model = UNet(n_channels=3, n_classes=1, height=height, width=width) else: height, width = 224, 224 model = eval(args.model, {'torch': torch, 'torchvision': torchvision}) if 'mobilenet_v2' in args.model: model = torch.nn.Sequential( model.features, torch.nn.AdaptiveAvgPool2d((1, 1)), torch.nn.Flatten(start_dim=1), model.classifier[0], model.classifier[1]) graph = Graph.create(model, input_shape=(3, height, width)) model.cuda() input_ = torch.randn([bs,3,height,width]).cuda() solver_info = SolverInfo(bs, mode=mode, model_name=model_name) solver_info.extract(graph, input_, *list(model.state_dict(keep_vars=True).values())) solution = solve_ilp_gurobi(solver_info, budget, time_limit=time_limit, ablation=args.ablation)
# scheduler.step() dice_a, dice_b, loss_a, loss_b = validate_net(net=net, loader=test_loader, name='Test ', epoch='final') # save results here save_path = os.path.join(result_path, project_name) if os.path.isdir(save_path) is False: os.makedirs(save_path) np.savez(os.path.join(save_path, 'params'), num_train=num_train, num_val=num_val, num_test=num_test, epochs=epochs, learning_rate=learning_rate, batch_size=batch_size, val_dice_a=val_dice_a, val_dice_b=val_dice_b, val_loss_a=val_loss_a, val_loss_b=val_loss_b, test_dice_a=dice_a, test_dice_b=dice_b, test_loss_a=loss_a, test_loss_b=loss_b) torch.save(net.state_dict(), os.path.join(save_path, 'net'))
# Num of iters for training num_iters = 2400 # Num of iters when to save image save_frequency = 50 z0 = z(shape=(img.height, img.width), channels=3) out_avg = None z0_saved = z0.detach().clone() noise = z0.detach().clone() z0_noise_std = 1. / 30. last_lost = None last_model_state = net.state_dict() for i in range(num_iters): optimizer.zero_grad() if z0_noise_std > 0: z0 = z0_saved + (noise.normal_() * z0_noise_std) output = net(z0) # Optimizer loss = mse(output, x) loss.backward() optimizer.step() cpu_loss = loss.data.cpu()
writer.add_scalar('psnr/train_psnr', psnr, global_step=step) writer.add_scalar('total_loss/g_loss_total', G_l_t, global_step=step) writer.add_scalar('total_loss/d_loss', D_l, global_step=step) writer.add_scalar('G_loss_total/g_loss', g_l, global_step=step) writer.add_scalar('G_loss_total/fl_loss', fl_l, global_step=step) writer.add_scalar('G_loss_total/inte_loss', inte_l, global_step=step) writer.add_scalar('G_loss_total/grad_loss', grad_l, global_step=step) writer.add_scalar('psnr/train_psnr', psnr, global_step=step) if step % int(train_cfg.iters / 100) == 0: writer.add_image('image/G_frame', save_G_frame, global_step=step) writer.add_image('image/target', save_target, global_step=step) if step % train_cfg.save_interval == 0: model_dict = {'net_g': generator.state_dict(), 'optimizer_g': optimizer_G.state_dict(), 'net_d': discriminator.state_dict(), 'optimizer_d': optimizer_D.state_dict()} torch.save(model_dict, f'weights/{train_cfg.dataset}_{step}.pth') print(f'\nAlready saved: \'{train_cfg.dataset}_{step}.pth\'.') if step % train_cfg.val_interval == 0: auc = val(train_cfg, model=generator) writer.add_scalar('results/auc', auc, global_step=step) generator.train() step += 1 if step > train_cfg.iters: training = False model_dict = {'net_g': generator.state_dict(), 'optimizer_g': optimizer_G.state_dict(), 'net_d': discriminator.state_dict(), 'optimizer_d': optimizer_D.state_dict()} torch.save(model_dict, f'weights/latest_{train_cfg.dataset}_{step}.pth')
def train_mc(self, dataloader, physics, epochs, lr, ckp_interval, schedule, residual=True, pretrained=None, task='', loss_type='l2', cat=True, report_psnr=False, lr_cos=False): save_path = './ckp/{}_mc_{}'.format(get_timestamp(), 'res' if residual else '', task) os.makedirs(save_path, exist_ok=True) generator = UNet(in_channels=self.in_channels, out_channels=self.out_channels, compact=4, residual=residual, circular_padding=True, cat=cat).to(self.device) if pretrained: checkpoint = torch.load(pretrained) generator.load_state_dict(checkpoint['state_dict']) if loss_type == 'l2': criterion_mc = torch.nn.MSELoss().to(self.device) if loss_type == 'l1': criterion_mc = torch.nn.L1Loss().to(self.device) optimizer = Adam(generator.parameters(), lr=lr['G'], weight_decay=lr['WD']) if report_psnr: log = LOG(save_path, filename='training_loss', field_name=['epoch', 'loss_fc', 'psnr', 'mse']) else: log = LOG(save_path, filename='training_loss', field_name=['epoch', 'loss_fc']) for epoch in range(epochs): adjust_learning_rate(optimizer, epoch, lr['G'], lr_cos, epochs, schedule) loss = closure_mc(generator, dataloader, physics, optimizer, criterion_mc, self.dtype, self.device, report_psnr) log.record(epoch + 1, *loss) if report_psnr: print('{}\tEpoch[{}/{}]\tmc={:.4e}\tpsnr={:.4f}\tmse={:.4e}'. format(get_timestamp(), epoch, epochs, *loss)) else: print('{}\tEpoch[{}/{}]\tmc={:.4e}'.format( get_timestamp(), epoch, epochs, *loss)) if epoch % ckp_interval == 0 or epoch + 1 == epochs: state = { 'epoch': epoch, 'state_dict': generator.state_dict(), 'optimizer': optimizer.state_dict() } torch.save( state, os.path.join(save_path, 'ckp_{}.pth.tar'.format(epoch))) log.close()
def train_ei_adv(self, dataloader, physics, transform, epochs, lr, alpha, ckp_interval, schedule, residual=True, pretrained=None, task='', loss_type='l2', cat=True, report_psnr=False, lr_cos=False): save_path = './ckp/{}_ei_adv_{}'.format(get_timestamp(), task) os.makedirs(save_path, exist_ok=True) generator = UNet(in_channels=self.in_channels, out_channels=self.out_channels, compact=4, residual=residual, circular_padding=True, cat=cat) if pretrained: checkpoint = torch.load(pretrained) generator.load_state_dict(checkpoint['state_dict']) discriminator = Discriminator( (self.in_channels, self.img_width, self.img_height)) generator = generator.to(self.device) discriminator = discriminator.to(self.device) if loss_type == 'l2': criterion_mc = torch.nn.MSELoss().to(self.device) criterion_ei = torch.nn.MSELoss().to(self.device) if loss_type == 'l1': criterion_mc = torch.nn.L1Loss().to(self.device) criterion_ei = torch.nn.L1Loss().to(self.device) criterion_gan = torch.nn.MSELoss().to(self.device) optimizer_G = Adam(generator.parameters(), lr=lr['G'], weight_decay=lr['WD']) optimizer_D = Adam(discriminator.parameters(), lr=lr['D'], weight_decay=0) if report_psnr: log = LOG(save_path, filename='training_loss', field_name=[ 'epoch', 'loss_mc', 'loss_ei', 'loss_g', 'loss_G', 'loss_D', 'psnr', 'mse' ]) else: log = LOG(save_path, filename='training_loss', field_name=[ 'epoch', 'loss_mc', 'loss_ei', 'loss_g', 'loss_G', 'loss_D' ]) for epoch in range(epochs): adjust_learning_rate(optimizer_G, epoch, lr['G'], lr_cos, epochs, schedule) adjust_learning_rate(optimizer_D, epoch, lr['D'], lr_cos, epochs, schedule) loss = closure_ei_adv(generator, discriminator, dataloader, physics, transform, optimizer_G, optimizer_D, criterion_mc, criterion_ei, criterion_gan, alpha, self.dtype, self.device, report_psnr) log.record(epoch + 1, *loss) if report_psnr: print( '{}\tEpoch[{}/{}]\tfc={:.4e}\tti={:.4e}\tg={:.4e}\tG={:.4e}\tD={:.4e}\tpsnr={:.4f}\tmse={:.4e}' .format(get_timestamp(), epoch, epochs, *loss)) else: print( '{}\tEpoch[{}/{}]\tfc={:.4e}\tti={:.4e}\tg={:.4e}\tG={:.4e}\tD={:.4e}' .format(get_timestamp(), epoch, epochs, *loss)) if epoch % ckp_interval == 0 or epoch + 1 == epochs: state = { 'epoch': epoch, 'state_dict_G': generator.state_dict(), 'state_dict_D': discriminator.state_dict(), 'optimizer_G': optimizer_G.state_dict(), 'optimizer_D': optimizer_D.state_dict() } torch.save( state, os.path.join(save_path, 'ckp_{}.pth.tar'.format(epoch))) log.close()
global_step=step) writer.add_scalar('psnr/train_psnr', psnr, global_step=step) if step % int(train_cfg.iters / 100) == 0: writer.add_image('image/G_frame', save_G_frame, global_step=step) writer.add_image('image/target', save_target, global_step=step) if step % train_cfg.save_interval == 0: model_dict = { 'net_g': generator.state_dict(), 'optimizer_g': optimizer_G.state_dict(), 'net_d': discriminator.state_dict(), 'optimizer_d': optimizer_D.state_dict() } torch.save(model_dict, args.model_dir + "model-%02d.pth" % step) print("Already saved:", args.model_dir + "model-%02d.pth" % step) if step % train_cfg.val_interval == 0 or step == 1: auc = val(train_cfg, model=generator) print(step, "auc score", auc) writer.add_scalar('results/auc', auc, global_step=step) generator.train()
def train(args): ''' -------------------------Hyperparameters-------------------------- ''' EPOCHS = args.epochs START = 0 # could enter a checkpoint start epoch ITER = args.iterations # per epoch LR = args.lr MOM = args.momentum # LOGInterval = args.log_interval BATCHSIZE = args.batch_size TEST_BATCHSIZE = args.test_batch_size NUMBER_OF_WORKERS = args.workers DATA_FOLDER = args.data TESTSET_FOLDER = args.testset ROOT = args.run WEIGHT_DIR = os.path.join(ROOT, "weights") CUSTOM_LOG_DIR = os.path.join(ROOT, "additionalLOGS") CHECKPOINT = os.path.join(WEIGHT_DIR, str(args.model) + str(args.name) + ".pt") useTensorboard = args.tb # check existance of data if not os.path.isdir(DATA_FOLDER): print("data folder not existant or in wrong layout.\n\t", DATA_FOLDER) exit(0) # check existance of testset if TESTSET_FOLDER is not None and not os.path.isdir(TESTSET_FOLDER): print("testset folder not existant or in wrong layout.\n\t", DATA_FOLDER) exit(0) ''' ---------------------------preparations--------------------------- ''' # CUDA for PyTorch use_cuda = torch.cuda.is_available() device = torch.device("cuda:0" if use_cuda else "cpu") print("using device: ", str(device)) # loading the validation samples to make online evaluations path_to_valX = args.valX path_to_valY = args.valY valX = None valY = None if path_to_valX is not None and path_to_valY is not None \ and os.path.exists(path_to_valX) and os.path.exists(path_to_valY) \ and os.path.isfile(path_to_valX) and os.path.isfile(path_to_valY): with torch.no_grad(): valX, valY = torch.load(path_to_valX, map_location='cpu'), \ torch.load(path_to_valY, map_location='cpu') ''' ---------------------------loading dataset and normalizing--------------------------- ''' # Dataloader Parameters train_params = { 'batch_size': BATCHSIZE, 'shuffle': True, 'num_workers': NUMBER_OF_WORKERS } test_params = { 'batch_size': TEST_BATCHSIZE, 'shuffle': False, 'num_workers': NUMBER_OF_WORKERS } # create a folder for the weights and custom logs if not os.path.isdir(WEIGHT_DIR): os.makedirs(WEIGHT_DIR) if not os.path.isdir(CUSTOM_LOG_DIR): os.makedirs(CUSTOM_LOG_DIR) labelsNorm = None # NORMLABEL # normalizing on a trainingset wide mean and std mean = None std = None if args.norm: print('computing mean and std over trainingset') # computes mean and std over all ground truths in dataset to tackle the problem of numerical insignificance mean, std = computeMeanStdOverDataset('CONRADataset', DATA_FOLDER, train_params, device) print('\niodine (mean/std): {}\t{}'.format(mean[0], std[0])) print('water (mean/std): {}\t{}\n'.format(mean[1], std[1])) labelsNorm = transforms.Normalize(mean=[0, 0], std=std) m2, s2 = computeMeanStdOverDataset('CONRADataset', DATA_FOLDER, train_params, device, transform=labelsNorm) print("new mean and std are:") print('\nnew iodine (mean/std): {}\t{}'.format(m2[0], s2[0])) print('new water (mean/std): {}\t{}\n'.format(m2[1], s2[1])) traindata = CONRADataset(DATA_FOLDER, True, device=device, precompute=True, transform=labelsNorm) testdata = None if TESTSET_FOLDER is not None: testdata = CONRADataset(TESTSET_FOLDER, False, device=device, precompute=True, transform=labelsNorm) else: testdata = CONRADataset(DATA_FOLDER, False, device=device, precompute=True, transform=labelsNorm) trainingset = DataLoader(traindata, **train_params) testset = DataLoader(testdata, **test_params) ''' ----------------loading model and checkpoints--------------------- ''' if args.model == "unet": m = UNet(2, 2).to(device) print( "using the U-Net architecture with {} trainable params; Good Luck!" .format(count_trainables(m))) else: m = simpleConvNet(2, 2).to(device) o = optim.SGD(m.parameters(), lr=LR, momentum=MOM) loss_fn = nn.MSELoss() test_loss = None train_loss = None if len(os.listdir(WEIGHT_DIR)) != 0: checkpoints = os.listdir(WEIGHT_DIR) checkDir = {} latestCheckpoint = 0 for i, checkpoint in enumerate(checkpoints): stepOfCheckpoint = int( checkpoint.split(str(args.model) + str(args.name))[-1].split('.pt')[0]) checkDir[stepOfCheckpoint] = checkpoint latestCheckpoint = max(latestCheckpoint, stepOfCheckpoint) print("[{}] {}".format(stepOfCheckpoint, checkpoint)) # if on development machine, prompt for input, else just take the most recent one if 'faui' in os.uname()[1]: toUse = int(input("select checkpoint to use: ")) else: toUse = latestCheckpoint checkpoint = torch.load(os.path.join(WEIGHT_DIR, checkDir[toUse])) m.load_state_dict(checkpoint['model_state_dict']) m.to(device) # pushing weights to gpu o.load_state_dict(checkpoint['optimizer_state_dict']) train_loss = checkpoint['train_loss'] test_loss = checkpoint['test_loss'] START = checkpoint['epoch'] print("using checkpoint {}:\n\tloss(train/test): {}/{}".format( toUse, train_loss, test_loss)) else: print("starting from scratch") ''' -----------------------------training----------------------------- ''' global_step = 0 # calculating initial loss if test_loss is None or train_loss is None: print("calculating initial loss") m.eval() print("testset...") test_loss = calculate_loss(set=testset, loss_fn=loss_fn, length_set=len(testdata), dev=device, model=m) print("trainset...") train_loss = calculate_loss(set=trainingset, loss_fn=loss_fn, length_set=len(traindata), dev=device, model=m) ## SSIM and R value R = [] SSIM = [] performanceFLE = os.path.join(CUSTOM_LOG_DIR, "performance.csv") with open(performanceFLE, 'w+') as f: f.write( "step, SSIMiodine, SSIMwater, Riodine, Rwater, train_loss, test_loss\n" ) print("computing ssim and r coefficents to: {}".format(performanceFLE)) # printing runtime information print( "starting training at {} for {} epochs {} iterations each\n\t{} total". format(START, EPOCHS, ITER, EPOCHS * ITER)) print("\tbatchsize: {}\n\tloss: {}\n\twill save results to \"{}\"".format( BATCHSIZE, train_loss, CHECKPOINT)) print( "\tmodel: {}\n\tlearningrate: {}\n\tmomentum: {}\n\tnorming output space: {}" .format(args.model, LR, MOM, args.norm)) #start actual training loops for e in range(START, START + EPOCHS): # iterations will not be interupted with validation and metrics for i in range(ITER): global_step = (e * ITER) + i # training m.train() iteration_loss = 0 for x, y in tqdm(trainingset): x, y = x.to(device=device, dtype=torch.float), y.to(device=device, dtype=torch.float) pred = m(x) loss = loss_fn(pred, y) iteration_loss += loss.item() o.zero_grad() loss.backward() o.step() print("\niteration {}: --accumulated loss {}".format( global_step, iteration_loss)) # validation, saving and logging print("\nvalidating") m.eval() # disable dropout batchnorm etc print("testset...") test_loss = calculate_loss(set=testset, loss_fn=loss_fn, length_set=len(testdata), dev=device, model=m) print("trainset...") train_loss = calculate_loss(set=trainingset, loss_fn=loss_fn, length_set=len(traindata), dev=device, model=m) print("calculating SSIM and R coefficients") currSSIM, currR = performance(set=testset, dev=device, model=m, bs=TEST_BATCHSIZE) print("SSIM (iod/water): {}/{}\nR (iod/water): {}/{}".format( currSSIM[0], currSSIM[1], currR[0], currR[1])) with open(performanceFLE, 'a') as f: newCSVline = "{}, {}, {}, {}, {}, {}, {}\n".format( global_step, currSSIM[0], currSSIM[1], currR[0], currR[1], train_loss, test_loss) f.write(newCSVline) print("wrote new line to csv:\n\t{}".format(newCSVline)) ''' if valX and valY were set in preparations, use them to perform analytics. if not, use the first sample from the testset to perform analytics ''' with torch.no_grad(): truth, pred = None, None IMAGE_LOG_DIR = os.path.join(CUSTOM_LOG_DIR, str(global_step)) if not os.path.isdir(IMAGE_LOG_DIR): os.makedirs(IMAGE_LOG_DIR) if valX is not None and valY is not None: batched = np.zeros((BATCHSIZE, *valX.numpy().shape)) batched[0] = valX.numpy() batched = torch.from_numpy(batched).to(device=device, dtype=torch.float) pred = m(batched) pred = pred.cpu().numpy()[0] truth = valY.numpy() # still on cpu assert pred.shape == truth.shape else: for x, y in testset: # x, y in shape[2,2,480,620] [b,c,h,w] x, y = x.to(device=device, dtype=torch.float), y.to(device=device, dtype=torch.float) pred = m(x) pred = pred.cpu().numpy()[ 0] # taking only the first sample of batch truth = y.cpu().numpy()[ 0] # first projection for evaluation advanvedMetrics(truth, pred, mean, std, global_step, args.norm, IMAGE_LOG_DIR) print("logging") CHECKPOINT = os.path.join( WEIGHT_DIR, str(args.model) + str(args.name) + str(global_step) + ".pt") torch.save( { 'epoch': e + 1, # end of this epoch; so resume at next. 'model_state_dict': m.state_dict(), 'optimizer_state_dict': o.state_dict(), 'train_loss': train_loss, 'test_loss': test_loss }, CHECKPOINT) print('\tsaved weigths to: ', CHECKPOINT) if logger is not None and train_loss is not None: logger.add_scalar('test_loss', test_loss, global_step=global_step) logger.add_scalar('train_loss', train_loss, global_step=global_step) logger.add_image("iodine-prediction", pred[0].reshape(1, 480, 620), global_step=global_step) logger.add_image("water-prediction", pred[1].reshape(1, 480, 620), global_step=global_step) # logger.add_image("water-prediction", wat) print( "\ttensorboard updated with test/train loss and a sample image" ) elif train_loss is not None: print("\tloss of global-step {}: {}".format( global_step, train_loss)) elif not useTensorboard: print("\t(tb-logging disabled) test/train loss: {}/{} ".format( test_loss, train_loss)) else: print("\tno loss accumulated yet") # saving final results print("saving upon exit") torch.save( { 'epoch': EPOCHS, 'model_state_dict': m.state_dict(), 'optimizer_state_dict': o.state_dict(), 'train_loss': train_loss, 'test_loss': test_loss }, CHECKPOINT) print('\tsaved progress to: ', CHECKPOINT) if logger is not None and train_loss is not None: logger.add_scalar('test_loss', test_loss, global_step=global_step) logger.add_scalar('train_loss', train_loss, global_step=global_step)
def main(): parser = argparse.ArgumentParser(description="Train the model") parser.add_argument('-trainf', "--train_filepath", type=str, default=None, required=True, help="training dataset filepath.") parser.add_argument('-validf', "--val_filepath", type=str, default=None, help="validation dataset filepath.") parser.add_argument("--shuffle", action="store_true", default=False, help="Shuffle the dataset") parser.add_argument("--debug", action="store_true", default=False) parser.add_argument('--epochs', type=int, default=30, metavar='N', help='number of epochs to train (default: 30)') parser.add_argument("--batch-size", type=int, default=32, help="Batch size") parser.add_argument('--img-shape', type=str, default="(1,512,512)", help='Image shape (default "(1,512,512)"') parser.add_argument("--num-cpu", type=int, default=10, help="Number of CPUs to use in parallel for dataloader.") parser.add_argument('--cuda', type=int, default=0, help='CUDA visible device (use CPU if -1, default: 0)') parser.add_argument('--cuda-non-deterministic', action='store_true', default=False, help="sets flags for non-determinism when using CUDA (potentially fast)") parser.add_argument('-lr', type=float, default=0.0005, help='Learning rate') parser.add_argument('--seed', type=int, default=0, help='Seed (numpy and cuda if GPU is used.).') parser.add_argument('--log-dir', type=str, default=None, help='Save the results/model weights/logs under the directory.') args = parser.parse_args() # TODO: support image reshape img_shape = tuple(map(int, args.img_shape.strip()[1:-1].split(","))) if args.log_dir: os.makedirs(args.log_dir, exist_ok=True) best_model_path = os.path.join(args.log_dir, "model_weights.pth") else: best_model_path = None if args.seed is not None: np.random.seed(args.seed) torch.manual_seed(args.seed) if args.cuda >= 0: if args.cuda_non_deterministic: printBlue("Warning: using CUDA non-deterministc. Could be faster but results might not be reproducible.") else: printBlue("Using CUDA deterministc. Use --cuda-non-deterministic might accelerate the training a bit.") # Make CuDNN Determinist torch.backends.cudnn.deterministic = not args.cuda_non_deterministic # torch.cuda.manual_seed(args.seed) torch.cuda.manual_seed_all(args.seed) # TODO [OPT] enable multi-GPUs ? device = torch.device("cuda:{}".format(args.cuda) if torch.cuda.is_available() and (args.cuda >= 0) else "cpu") # ================= Build dataloader ================= # DataLoader # transform_normalize = transforms.Normalize(mean=[0.5, 0.5, 0.5], # std=[0.5, 0.5, 0.5]) transform_normalize = transforms.Normalize(mean=[0.5], std=[0.5]) # Warning: DO NOT use geometry transform (do it in the dataloader instead) data_transform = transforms.Compose([ # transforms.ToPILImage(mode='F'), # mode='F' for one-channel image # transforms.Resize((256, 256)) # NO # transforms.RandomResizedCrop(256), # NO # transforms.RandomHorizontalFlip(p=0.5), # NO # WARNING, ISSUE: transforms.ColorJitter doesn't work with ToPILImage(mode='F'). # Need custom data augmentation functions: TODO # transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2), # TODO TODO transforms.ToTensor(), # already done in the dataloader transform_normalize # ]) data_loader_params = {'batch_size': args.batch_size, 'shuffle': args.shuffle, 'num_workers': args.num_cpu, # 'sampler': balanced_sampler, 'drop_last': True, 'pin_memory': False } train_set = LiTSDataset(args.train_filepath, dtype=np.float32, transform=data_transform, ) valid_set = LiTSDataset(args.val_filepath, dtype=np.float32, transform=data_transform, ) dataloader_train = torch.utils.data.DataLoader(train_set, **data_loader_params) dataloader_valid = torch.utils.data.DataLoader(valid_set, **data_loader_params) # =================== Build model =================== model = UNet(in_ch=1, out_ch=3, # there are 3 classes: 0: background, 1: liver, 2: tumor depth=4, start_ch=32, inc_rate=2, padding=True, batch_norm=True, spec_norm=False, dropout=0.5, up_mode='upconv', include_top=True) model.to(device) optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, betas=(0.9, 0.95)) # TODO best_valid_loss = float('inf') for epoch in range(args.epochs): for valid_mode, dataloader in enumerate([dataloader_train, dataloader_valid]): n_batch_per_epoch = len(dataloader) if args.debug: n_batch_per_epoch = 1 # infinite dataloader allows several update per iteration (for special models e.g. GAN) dataloader = infinite_dataloader(dataloader) if valid_mode: printYellow("Switch to validation mode.") model.eval() prev_grad_mode = torch.is_grad_enabled() torch.set_grad_enabled(False) else: model.train() st = time.time() cum_loss = 0 for iter_ind in range(n_batch_per_epoch): supplement_logs = "" # reset cumulated losses at the begining of each batch # loss_manager.reset_losses() # TODO TODO optimizer.zero_grad() img, msk = next(dataloader) img, msk = img.to(device), msk.to(device) # TODO this is ugly: convert dtype and convert the shape from (N, 1, 512, 512) to (N, 512, 512) msk = msk.to(torch.long).squeeze(1) msk_pred = model(img) # label_weights is determined according the liver_ratio & tumor_ratio loss = CrossEntropyLoss(msk_pred, msk, label_weights=[1., 10., 100.], device=device) if valid_mode: pass else: loss.backward() optimizer.step() loss = loss.item() # release cum_loss += loss if valid_mode: print("\r--------(valid) {:.2%} Loss: {:.3f} (time: {:.1f}s) |supp: {}".format( (iter_ind+1)/n_batch_per_epoch, cum_loss/(iter_ind+1), time.time()-st, supplement_logs), end="") else: print("\rEpoch: {:3}/{} {:.2%} Loss: {:.3f} (time: {:.1f}s) |supp: {}".format( (epoch+1), args.epochs, (iter_ind+1)/n_batch_per_epoch, cum_loss/(iter_ind+1), time.time()-st, supplement_logs), end="") print() if valid_mode: torch.set_grad_enabled(prev_grad_mode) valid_mean_loss = cum_loss/(iter_ind+1) # validation (mean) loss of the current epoch if best_model_path and (valid_mean_loss < best_valid_loss): printGreen("Valid loss decreases from {:.5f} to {:.5f}, saving best model.".format( best_valid_loss, valid_mean_loss)) best_valid_loss = valid_mean_loss # Only need to save the weights torch.save(model.state_dict(), best_model_path) return best_valid_loss
class Noise2Noise(object): """Implementation of Noise2Noise from Lehtinen et al. (2018).""" def __init__(self, params, trainable): """Initializes model.""" self.p = params self.trainable = trainable self._compile() #初始化模型 def _compile(self): """ Compiles model (architecture, loss function, optimizers, etc.). 初始化 网络、损失函数、优化器等 """ print('Noise2Noise: Learning Image Restoration without Clean Data (Lethinen et al., 2018)') # Model (3x3=9 channels for Monte Carlo since it uses 3 HDR buffers) 已删除蒙特卡洛相关代码 if self.p.noise_type == 'mc': self.is_mc = True self.model = UNet(in_channels=9) else: self.is_mc = False self.model = UNet(in_channels=3) # Set optimizer and loss, if in training mode # 如果 为训练,则初始化优化器和损失 if self.trainable: self.optim = Adam(self.model.parameters(), lr=self.p.learning_rate, betas=self.p.adam[:2], eps=self.p.adam[2]) # Learning rate adjustment self.scheduler = lr_scheduler.ReduceLROnPlateau(self.optim, patience=self.p.nb_epochs/4, factor=0.5, verbose=True) # Loss function if self.p.loss == 'hdr': assert self.is_mc, 'Using HDR loss on non Monte Carlo images' self.loss = HDRLoss() elif self.p.loss == 'l2': self.loss = nn.MSELoss() else: self.loss = nn.L1Loss() # CUDA support self.use_cuda = torch.cuda.is_available() and self.p.cuda if self.use_cuda: self.model = self.model.cuda() if self.trainable: self.loss = self.loss.cuda() def _print_params(self): """Formats parameters to print when training.""" print('Training parameters: ') self.p.cuda = self.use_cuda param_dict = vars(self.p) pretty = lambda x: x.replace('_', ' ').capitalize() print('\n'.join(' {} = {}'.format(pretty(k), str(v)) for k, v in param_dict.items())) print() def save_model(self, epoch, stats, first=False): """Saves model to files; can be overwritten at every epoch to save disk space.""" # Create directory for model checkpoints, if nonexistent if first: if self.p.clean_targets: ckpt_dir_name = f'{datetime.now():{self.p.noise_type}-clean-%H%M}' else: ckpt_dir_name = f'{datetime.now():{self.p.noise_type}-%H%M}' if self.p.ckpt_overwrite: if self.p.clean_targets: ckpt_dir_name = f'{self.p.noise_type}-clean' else: ckpt_dir_name = self.p.noise_type self.ckpt_dir = os.path.join(self.p.ckpt_save_path, ckpt_dir_name) if not os.path.isdir(self.p.ckpt_save_path): os.mkdir(self.p.ckpt_save_path) if not os.path.isdir(self.ckpt_dir): os.mkdir(self.ckpt_dir) # Save checkpoint dictionary if self.p.ckpt_overwrite: fname_unet = '{}/n2n-{}.pt'.format(self.ckpt_dir, self.p.noise_type) else: valid_loss = stats['valid_loss'][epoch] fname_unet = '{}/n2n-epoch{}-{:>1.5f}.pt'.format(self.ckpt_dir, epoch + 1, valid_loss) print('Saving checkpoint to: {}\n'.format(fname_unet)) torch.save(self.model.state_dict(), fname_unet) # Save stats to JSON fname_dict = '{}/n2n-stats.json'.format(self.ckpt_dir) with open(fname_dict, 'w') as fp: json.dump(stats, fp, indent=2) def load_model(self, ckpt_fname): """Loads model from checkpoint file.""" print('Loading checkpoint from: {}'.format(ckpt_fname)) if self.use_cuda: self.model.load_state_dict(torch.load(ckpt_fname)) else: self.model.load_state_dict(torch.load(ckpt_fname, map_location='cpu')) def _on_epoch_end(self, stats, train_loss, epoch, epoch_start, valid_loader): """Tracks and saves starts after each epoch.""" # Evaluate model on validation set print('\rTesting model on validation set... ', end='') epoch_time = time_elapsed_since(epoch_start)[0] valid_loss, valid_time, valid_psnr = self.eval(valid_loader) show_on_epoch_end(epoch_time, valid_time, valid_loss, valid_psnr) # Decrease learning rate if plateau self.scheduler.step(valid_loss) # Save checkpoint stats['train_loss'].append(train_loss) stats['valid_loss'].append(valid_loss) stats['valid_psnr'].append(valid_psnr) self.save_model(epoch, stats, epoch == 0) def test(self, test_loader, show=1): """Evaluates denoiser on test set.""" self.model.train(False) source_imgs = [] denoised_imgs = [] clean_imgs = [] # Create directory for denoised images denoised_dir = os.path.dirname(self.p.data) save_path = os.path.join(denoised_dir, 'denoised') if not os.path.isdir(save_path): os.mkdir(save_path) for batch_idx, (source, target) in enumerate(test_loader): # Only do first <show> images if show == 0 or batch_idx >= show: break source_imgs.append(source) clean_imgs.append(target) if self.use_cuda: source = source.cuda() # Denoise denoised_img = self.model(source).detach() denoised_imgs.append(denoised_img) # Squeeze tensors source_imgs = [t.squeeze(0) for t in source_imgs] denoised_imgs = [t.squeeze(0) for t in denoised_imgs] clean_imgs = [t.squeeze(0) for t in clean_imgs] # Create montage and save images print('Saving images and montages to: {}'.format(save_path)) for i in range(len(source_imgs)): img_name = test_loader.dataset.imgs[i] create_montage(img_name, self.p.noise_type, save_path, source_imgs[i], denoised_imgs[i], clean_imgs[i], show) def eval(self, valid_loader): """Evaluates denoiser on validation set.""" self.model.train(False) valid_start = datetime.now() loss_meter = AvgMeter() psnr_meter = AvgMeter() for batch_idx, (source, target) in enumerate(valid_loader): if self.use_cuda: source = source.cuda() target = target.cuda() # Denoise source_denoised = self.model(source) # Update loss loss = self.loss(source_denoised, target) loss_meter.update(loss.item()) # Compute PSRN if self.is_mc: source_denoised = reinhard_tonemap(source_denoised) # TODO: Find a way to offload to GPU, and deal with uneven batch sizes for i in range(self.p.batch_size): source_denoised = source_denoised.cpu() target = target.cpu() psnr_meter.update(psnr(source_denoised[i], target[i]).item()) valid_loss = loss_meter.avg valid_time = time_elapsed_since(valid_start)[0] psnr_avg = psnr_meter.avg return valid_loss, valid_time, psnr_avg def train(self, train_loader, valid_loader): """Trains denoiser on training set.""" self.model.train(True) self._print_params() num_batches = len(train_loader) assert num_batches % self.p.report_interval == 0, 'Report interval must divide total number of batches' # Dictionaries of tracked stats stats = {'noise_type': self.p.noise_type, 'noise_param': self.p.noise_param, 'train_loss': [], 'valid_loss': [], 'valid_psnr': []} # Main training loop train_start = datetime.now() for epoch in range(self.p.nb_epochs): print('EPOCH {:d} / {:d}'.format(epoch + 1, self.p.nb_epochs)) # Some stats trackers epoch_start = datetime.now() train_loss_meter = AvgMeter() loss_meter = AvgMeter() time_meter = AvgMeter() # Minibatch SGD for batch_idx, (source, target) in enumerate(train_loader): batch_start = datetime.now() progress_bar(batch_idx, num_batches, self.p.report_interval, loss_meter.val) if self.use_cuda: source = source.cuda() target = target.cuda() # Denoise image source_denoised = self.model(source) loss = self.loss(source_denoised, target) loss_meter.update(loss.item()) # Zero gradients, perform a backward pass, and update the weights self.optim.zero_grad() loss.backward() self.optim.step() # Report/update statistics time_meter.update(time_elapsed_since(batch_start)[1]) if (batch_idx + 1) % self.p.report_interval == 0 and batch_idx: show_on_report(batch_idx, num_batches, loss_meter.avg, time_meter.avg) train_loss_meter.update(loss_meter.avg) loss_meter.reset() time_meter.reset() # Epoch end, save and reset tracker self._on_epoch_end(stats, train_loss_meter.avg, epoch, epoch_start, valid_loader) train_loss_meter.reset() train_elapsed = time_elapsed_since(train_start)[0] print('Training done! Total elapsed time: {}\n'.format(train_elapsed))
1, criterion_seg, is_train=False, name='Refine test', epoch=0) save_path = os.path.join(result_path, project_name) if os.path.isdir(save_path) is False: os.makedirs(save_path) all_results = { "project_name": project_name, "project_description": project_description, "epochs": epochs, "num_train": num_train, "num_val": num_val, "num_test": num_test, "learning_rate": learning_rate, "batch_size": batch_size, "train_dice": train_dices, "train_loss": train_losses, "val_dice": val_dices, "val_loss": val_losses, "test_dice": test_dice, "test_loss": test_loss, } np.savez(os.path.join(save_path, 'params'), **all_results) torch.save(refine_net.state_dict(), os.path.join(save_path, 'refiner'))
if os.path.isdir(save_path) is False: os.makedirs(save_path) all_results = { "project_name": project_name, "project_description": project_description, "epochs": epochs, "num_a_train": num_a_train, "num_a_val": num_a_val, "num_a_test": num_a_test, "num_b_train": num_b_train, "num_b_val": num_b_val, "num_b_test": num_b_test, "learning_rate": learning_rate, "batch_size": batch_size, "train_dice": train_dices, "train_loss": train_losses, "val_a_dice": val_a_dices, "val_a_loss": val_a_losses, "val_b_dice": val_b_dices, "val_b_loss": val_b_losses, "test_a_dice": test_a_dice, "test_a_loss": test_a_loss, "test_b_dice": test_b_dice, "test_b_loss": test_b_loss, } np.savez(os.path.join(save_path, 'params'), **all_results) torch.save(ds_unet.state_dict(), os.path.join(save_path, 'ds_unet'))
def train_eval_model(opts): # parse model configuration num_epochs = opts["num_epochs"] train_batch_size = opts["train_batch_size"] val_batch_size = opts["eval_batch_size"] dataset_type = opts["dataset_type"] opti_mode = opts["optimizer"] loss_criterion = opts["loss_criterion"] lr = opts["lr"] lr_decay = opts["lr_decay"] wd = opts["weight_decay"] gpus = opts["gpu_list"].split(',') os.environ['CUDA_VISIBLE_DEVICE'] = opts["gpu_list"] train_dir = opts["log_dir"] train_data_dir = opts["train_data_dir"] eval_data_dir = opts["eval_data_dir"] pretrained = opts["pretrained_model"] resume = opts["resume"] display_iter = opts["display_iter"] save_epoch = opts["save_every_epoch"] show = opts["vis"] # backup train configs log_file = os.path.join(train_dir, "log_file.txt") os.makedirs(train_dir, exist_ok=True) model_dir = os.path.join(train_dir, "code_backup") os.makedirs(model_dir, exist_ok=True) if resume is None and os.path.exists(log_file): os.remove(log_file) shutil.copy("./models/unet.py", os.path.join(model_dir, "unet.py")) shutil.copy("./trainer_unet.py", os.path.join(model_dir, "trainer_unet.py")) shutil.copy("./datasets/dataset.py", os.path.join(model_dir, "dataset.py")) ckt_dir = os.path.join(train_dir, "checkpoints") os.makedirs(ckt_dir, exist_ok=True) # format printing configs print("*" * 50) table_key = [] table_value = [] n = 0 for key, value in opts.items(): table_key.append(key) table_value.append(str(value)) n += 1 print_table([table_key, ["="] * n, table_value]) # format gpu list gpu_list = [] for str_id in gpus: id = int(str_id) gpu_list.append(id) # dataloader print("==> Create dataloader") dataloaders_dict = { "train": er_data_loader(train_data_dir, train_batch_size, dataset_type, is_train=True), "eval": er_data_loader(eval_data_dir, val_batch_size, dataset_type, is_train=False) } # define parameters of two networks print("==> Create network") num_channels = 1 num_classes = 1 model = UNet(num_channels, num_classes) init_weights(model) # loss layer criterion = create_criterion(criterion=loss_criterion) best_acc = 0.0 start_epoch = 0 # load pretrained model if pretrained is not None and os.path.isfile(pretrained): print("==> Train from model '{}'".format(pretrained)) checkpoint_gan = torch.load(pretrained) model.load_state_dict(checkpoint_gan['model_state_dict']) print("==> Loaded checkpoint '{}')".format(pretrained)) for param in model.parameters(): param.requires_grad = False # resume training elif resume is not None and os.path.isfile(resume): print("==> Resume from checkpoint '{}'".format(resume)) checkpoint = torch.load(resume) start_epoch = checkpoint['epoch'] + 1 best_acc = checkpoint['best_acc'] model_dict = model.state_dict() pretrained_dict = { k: v for k, v in checkpoint['model_state_dict'].items() if k in model_dict and v.size() == model_dict[k].size() } model_dict.update(pretrained_dict) model.load_state_dict(pretrained_dict) print("==> Loaded checkpoint '{}' (epoch {})".format( resume, checkpoint['epoch'] + 1)) # train from scratch else: print("==> Train from initial or random state.") # define mutiple-gpu mode device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") model.cuda() model = nn.DataParallel(model) # print learnable parameters print("==> List learnable parameters") for name, param in model.named_parameters(): if param.requires_grad == True: print("\t{}, size {}".format(name, param.size())) params_to_update = [{'params': model.parameters()}] # define optimizer print("==> Create optimizer") optimizer = create_optimizer(params_to_update, opti_mode, lr=lr, momentum=0.9, wd=wd) if resume is not None and os.path.isfile(resume): optimizer.load_state_dict(checkpoint['optimizer']) # start training since = time.time() # Each epoch has a training and validation phase print("==> Start training") total_steps = 0 for epoch in range(start_epoch, num_epochs): print('-' * 50) print("==> Epoch {}/{}".format(epoch + 1, num_epochs)) total_steps = train_one_epoch(epoch, total_steps, dataloaders_dict['train'], model, device, criterion, optimizer, lr, lr_decay, display_iter, log_file, show) epoch_acc, epoch_iou, epoch_f1 = eval_one_epoch( epoch, dataloaders_dict['eval'], model, device, log_file) if best_acc < epoch_acc and epoch >= 5: best_acc = epoch_acc torch.save( { 'epoch': epoch, 'model_state_dict': model.module.state_dict(), 'optimizer': optimizer.state_dict(), 'best_acc': best_acc }, os.path.join(ckt_dir, "best.pth")) if (epoch + 1) % save_epoch == 0 and (epoch + 1) >= 20: torch.save( { 'epoch': epoch, 'model_state_dict': model.module.state_dict(), 'optimizer': optimizer.state_dict(), 'best_iou': epoch_iou }, os.path.join(ckt_dir, "checkpoints_" + str(epoch + 1) + ".pth")) time_elapsed = time.time() - since time_message = 'Training complete in {:.0f}m {:.0f}s'.format( time_elapsed // 60, time_elapsed % 60) print(time_message) with open(log_file, "a+") as fid: fid.write('%s\n' % time_message) print('==> Best val Acc: {:4f}'.format(best_acc))
optimizer.zero_grad() outputs = net(inputs) batch_size = outputs.size(0) outputs = outputs.reshape((batch_size, -1)) target = target.reshape((batch_size, -1)) loss = criterion(outputs, target) loss.backward() optimizer.step() running_loss += loss.item() train_info = 'epoch:%d train_loss: %.3f' % (epoch + 1, running_loss / (i + 1)) print(train_info) write_log('weight/train.log', str(datetime.datetime.now())) write_log('weight/train.log', train_info) torch.save(net.state_dict(), 'weight/epoch_{}_{}'.format(epoch, i)) # test phase with torch.no_grad(): net.eval() test_loss = 0.0 accuracy = 0 count = 0 for i, batch in enumerate(test_data_loader): inputs = batch['image'].to(cuda0) target = batch['target'].to(cuda0) outputs = net(inputs) # loss batch_size = outputs.size(0) loss = criterion(outputs.reshape((batch_size, -1)), target.reshape((batch_size, -1)))
class Trainer: def __init__(self, seq_length, color_channels, unet_path="pretrained/unet.mdl", discrim_path="pretrained/dicrim.mdl", facenet_path="pretrained/facenet.mdl", vgg_path="", embedding_size=1000, unet_depth=3, unet_filts=32, facenet_filts=32, resnet=18): self.color_channels = color_channels self.margin = 0.5 self.writer = SummaryWriter(log_dir="logs") self.unet_path = unet_path self.discrim_path = discrim_path self.facenet_path = facenet_path self.unet = UNet(in_channels=color_channels, out_channels=color_channels, depth=unet_depth, start_filts=unet_filts, up_mode="upsample", merge_mode='concat').to(device) self.discrim = FaceNetModel(embedding_size=embedding_size, start_filts=facenet_filts, in_channels=color_channels, resnet=resnet, pretrained=False).to(device) self.facenet = FaceNetModel(embedding_size=embedding_size, start_filts=facenet_filts, in_channels=color_channels, resnet=resnet, pretrained=False).to(device) if os.path.isfile(unet_path): self.unet.load_state_dict(torch.load(unet_path)) print("unet loaded") if os.path.isfile(discrim_path): self.discrim.load_state_dict(torch.load(discrim_path)) print("discrim loaded") if os.path.isfile(facenet_path): self.facenet.load_state_dict(torch.load(facenet_path)) print("facenet loaded") if os.path.isfile(vgg_path): self.vgg_loss_network = LossNetwork(vgg_face_dag(vgg_path)).to(device) self.vgg_loss_network.eval() print("vgg loaded") self.mse_loss_function = nn.MSELoss().to(device) self.discrim_loss_function = nn.BCELoss().to(device) self.triplet_loss_function = TripletLoss(margin=self.margin) self.unet_optimizer = torch.optim.Adam(self.unet.parameters(), betas=(0.9, 0.999)) self.discrim_optimizer = torch.optim.Adam(self.discrim.parameters(), betas=(0.9, 0.999)) self.facenet_optimizer = torch.optim.Adam(self.facenet.parameters(), betas=(0.9, 0.999)) def test(self, test_loader, epoch=0): X, y = next(iter(test_loader)) B, D, C, W, H = X.shape # X = X.view(B, C * D, W, H) self.unet.eval() self.facenet.eval() self.discrim.eval() with torch.no_grad(): y_ = self.unet(X.to(device)) mse = self.mse_loss_function(y_, y.to(device)) loss_G = self.loss_GAN_generator(btch_X=X.to(device)) loss_D = self.loss_GAN_discrimator(btch_X=X.to(device), btch_y=y.to(device)) loss_facenet, _, n_bad = self.loss_facenet(X.to(device), y.to(device)) plt.title(f"epoch {epoch} mse={mse.item():.4} facenet={loss_facenet.item():.4} bad={n_bad / B ** 2}") i = np.random.randint(0, B) a = np.hstack((y[i].transpose(0, 1).transpose(1, 2), y_[i].transpose(0, 1).transpose(1, 2).to(cpu))) b = np.hstack((X[i][0].transpose(0, 1).transpose(1, 2), X[i][-1].transpose(0, 1).transpose(1, 2))) plt.imshow(np.vstack((a, b))) plt.axis('off') plt.show() self.writer.add_scalar("test bad_percent", n_bad / B ** 2, global_step=epoch) self.writer.add_scalar("test loss", mse.item(), global_step=epoch) # self.writer.add_scalars("test GAN", {"discrim": loss_D.item(), # "gen": loss_G.item()}, global_step=epoch) with torch.no_grad(): n_for_show = 10 y_show_ = y_.to(device) y_show = y.to(device) embeddings_anc, _ = self.facenet(y_show_) embeddings_pos, _ = self.facenet(y_show) embeds = torch.cat((embeddings_anc[:n_for_show], embeddings_pos[:n_for_show])) imgs = torch.cat((y_show_[:n_for_show], y_show[:n_for_show])) names = list(range(n_for_show)) * 2 # print(embeds.shape, imgs.shape, len(names)) # self.writer.add_embedding(mat=embeds, metadata=names, label_img=imgs, tag="embeddings", global_step=epoch) trshs, fprs, tprs = roc_curve(embeddings_anc.detach().to(cpu), embeddings_pos.detach().to(cpu)) rnk1 = rank1(embeddings_anc.detach().to(cpu), embeddings_pos.detach().to(cpu)) plt.step(fprs, tprs) # plt.xlim((1e-4, 1)) plt.yticks(np.arange(0, 1, 0.05)) plt.xticks(np.arange(min(fprs), max(fprs), 10)) plt.xscale('log') plt.title(f"ROC auc={auc(fprs, tprs)} rnk1={rnk1}") self.writer.add_figure("ROC test", plt.gcf(), global_step=epoch) self.writer.add_scalar("auc", auc(fprs, tprs), global_step=epoch) self.writer.add_scalar("rank1", rnk1, global_step=epoch) print(f"\n###### {epoch} TEST mse={mse.item():.4} GAN(G/D)={loss_G.item():.4}/{loss_D.item():.4} " f"facenet={loss_facenet.item():.4} bad={n_bad / B ** 2:.4} auc={auc(fprs, tprs)} rank1={rnk1} #######") def test_test(self, test_loader): X, ys = next(iter(test_loader)) true_idx = 0 x = X[true_idx] D, C, W, H = x.shape # x = x.view(C * D, W, H) dists = list() with torch.no_grad(): y_ = self.unet(x.to(device)) embedding_anc, _ = self.facenet(y_) embeddings_pos, _ = self.facenet(ys) for emb_pos_item in embeddings_pos: dist = l2_dist.forward(embedding_anc, emb_pos_item) dists.append(dist) a_sorted = np.argsort(dists) a = np.hstack((ys[true_idx].transpose(0, 1).transpose(1, 2), y_.transpose(0, 1).transpose(1, 2).to(cpu).numpy(), ys[a_sorted[0]].transpose(0, 1).transpose(1, 2))) b = np.hstack((x[0:3].transpose(0, 1).transpose(1, 2), x[D // 2 * C:D // 2 * C + 3].transpose(0, 1).transpose(1, 2), x[-3:].transpose(0, 1).transpose(1, 2))) b_ = b - np.min(b) b_ = b_ / np.max(b) b_ = equalize_func([(b_ * 255).astype(np.uint8)], use_clahe=True)[0] b = b_.astype(np.float32) / 255 plt.imshow(cv2.cvtColor(np.vstack((a, b)), cv2.COLOR_BGR2RGB)) plt.axis('off') plt.show() def loss_facenet(self, X, y, is_detached=False): B, D, C, W, H = X.shape y_ = self.unet(X) embeddings_anc, D_fake = self.facenet(y_ if not is_detached else y_.detach()) embeddings_pos, D_real = self.facenet(y) target_real = torch.full_like(D_fake, 1) loss_gen = self.discrim_loss_function(D_fake, target_real) pos_dist = l2_dist.forward(embeddings_anc, embeddings_pos) bad_triplets_loss = None n_bad = 0 for shift in range(1, B): embeddings_neg = torch.roll(embeddings_pos, shift, 0) neg_dist = l2_dist.forward(embeddings_anc, embeddings_neg) bad_triplets_idxs = np.where((neg_dist - pos_dist < self.margin).cpu().numpy().flatten())[0] if shift == 1: bad_triplets_loss = self.triplet_loss_function.forward(embeddings_anc[bad_triplets_idxs], embeddings_pos[bad_triplets_idxs], embeddings_neg[bad_triplets_idxs]).to( device) else: bad_triplets_loss += self.triplet_loss_function.forward(embeddings_anc[bad_triplets_idxs], embeddings_pos[bad_triplets_idxs], embeddings_neg[bad_triplets_idxs]).to(device) n_bad += len(bad_triplets_idxs) bad_triplets_loss /= B return bad_triplets_loss, torch.mean(loss_gen), n_bad # def loss_mse(self, btch_X, btch_y): # btch_y_ = self.unet(btch_X) # loss_unet = self.mse_loss_function(btch_y_, btch_y) # # features_target = self.facenet.forward_mse(btch_y) # features = self.facenet.forward_mse(btch_y_) # # loss_first_layer = self.mse_loss_function(features, features_target) # return loss_unet + loss_first_layer def loss_mse_vgg(self, btch_X, btch_y, k_mse, k_vgg): btch_y_ = self.unet(btch_X) # print(btch_y_.shape,btch_y.shape) perceptual_btch_y_ = self.vgg_loss_network(btch_y_) perceptual_btch_y = self.vgg_loss_network(btch_y) perceptual_loss = 0.0 for a, b in zip(perceptual_btch_y_, perceptual_btch_y): perceptual_loss += self.mse_loss_function(a, b) return k_vgg * perceptual_loss + k_mse * self.mse_loss_function(btch_y_, btch_y) def loss_GAN_discrimator(self, btch_X, btch_y): btch_y_ = self.unet(btch_X) _, y_D_fake_ = self.discrim(btch_y_.detach()) _, y_D_real_ = self.discrim(btch_y) target_fake = torch.full_like(y_D_fake_, 0) target_real = torch.full_like(y_D_real_, 1) loss_D_fake_ = self.discrim_loss_function(y_D_fake_, target_fake) loss_D_real_ = self.discrim_loss_function(y_D_real_, target_real) loss_discrim = (loss_D_real_ + loss_D_fake_) return loss_discrim def loss_GAN_generator(self, btch_X): btch_y_ = self.unet(btch_X) _, y_D_fake_ = self.discrim(btch_y_) target_real = torch.full_like(y_D_fake_, 1) loss_gen = self.discrim_loss_function(y_D_fake_, target_real) return loss_gen def relax_discriminator(self, btch_X, btch_y): self.discrim.zero_grad() # train with real y_discrim_real_ = self.discrim(btch_y) y_discrim_real_ = y_discrim_real_.mean() y_discrim_real_.backward(self.mone) # train with fake btch_y_ = self.unet(btch_X) y_discrim_fake_detached_ = self.discrim(btch_y_.detach()) y_discrim_fake_detached_ = y_discrim_fake_detached_.mean() y_discrim_fake_detached_.backward(self.one) # gradient_penalty gradient_penalty = self.discrim_gradient_penalty(btch_y, btch_y_) gradient_penalty.backward() self.discrim_optimizer.step() def relax_generator(self, btch_X): self.unet.zero_grad() btch_y_ = self.unet(btch_X) y_discrim_fake_ = self.discrim(btch_y_) y_discrim_fake_ = y_discrim_fake_.mean() y_discrim_fake_.backward(self.mone) self.unet_optimizer.step() def discrim_gradient_penalty(self, real_y, fake_y): lambd = 10 btch_size = real_y.shape[0] alpha = torch.rand(btch_size, 1, 1, 1).to(device) # print(alpha.shape, real_y.shape) alpha = alpha.expand_as(real_y) interpolates = alpha * real_y + (1 - alpha) * fake_y interpolates = interpolates.to(device) interpolates = autograd.Variable(interpolates, requires_grad=True) interpolates_out = self.discrim(interpolates) gradients = autograd.grad(outputs=interpolates_out, inputs=interpolates, grad_outputs=torch.ones(interpolates_out.size()).to(device), create_graph=True, retain_graph=True, only_inputs=True)[0] gradient_penalty = ((gradients.norm(2, dim=1) - 1) ** 2).mean() * lambd return gradient_penalty def train(self, train_loader, test_loader, batch_size=2, epochs=30, k_gen=1, k_discrim=1, k_mse=1, k_facenet=1, k_facenet_back=1, k_vgg=1): """ :param X: np.array shape=(n_videos, n_frames, h, w) :param y: np.array shape=(n_videos, h, w) :param epochs: int """ print("\nSTART TRAINING\n") for epoch in range(epochs): self.test(test_loader, epoch) self.unet.train() self.facenet.train() self.discrim.train() # train by batches for idx, (btch_X, btch_y) in enumerate(train_loader): B, D, C, W, H = btch_X.shape # btch_X = btch_X.view(B, C * D, W, H) btch_X = btch_X.to(device) btch_y = btch_y.to(device) # Mse loss self.unet.zero_grad() mse = self.loss_mse_vgg(btch_X, btch_y, k_mse, k_vgg) mse.backward() self.unet_optimizer.step() # facenet_backup = deepcopy(self.facenet.state_dict()) # for i in range(unrolled_iterations): self.discrim.zero_grad() loss_D = self.loss_GAN_discrimator(btch_X, btch_y) loss_D = k_discrim * loss_D loss_D.backward() self.discrim_optimizer.step() self.discrim.zero_grad() self.unet.zero_grad() loss_G = self.loss_GAN_generator(btch_X) loss_G = k_gen * loss_G loss_G.backward() self.unet_optimizer.step() # Facenet self.unet.zero_grad() self.facenet.zero_grad() facenet_loss, _, n_bad = self.loss_facenet(btch_X, btch_y) facenet_loss = k_facenet * facenet_loss facenet_loss.backward() self.facenet_optimizer.step() self.unet.zero_grad() self.facenet.zero_grad() facenet_back_loss, _, n_bad = self.loss_facenet(btch_X, btch_y) facenet_back_loss = k_facenet_back * facenet_back_loss facenet_back_loss.backward() self.unet_optimizer.step() print(f"btch {idx * batch_size} mse={mse.item():.4} GAN(G/D)={loss_G.item():.4}/{loss_D.item():.4} " f"facenet={facenet_loss.item():.4} bad={n_bad / B ** 2:.4}") global_step = epoch * len(train_loader.dataset) // batch_size + idx self.writer.add_scalar("train bad_percent", n_bad / B ** 2, global_step=global_step) self.writer.add_scalar("train loss", mse.item(), global_step=global_step) # self.writer.add_scalars("train GAN", {"discrim": loss_D.item(), # "gen": loss_G.item()}, global_step=global_step) torch.save(self.unet.state_dict(), self.unet_path) torch.save(self.discrim.state_dict(), self.discrim_path) torch.save(self.facenet.state_dict(), self.facenet_path)
def train(frame_num, layer_nums, input_channels, output_channels, discriminator_num_filters, bn=False, pretrain=False, generator_pretrain_path=None, discriminator_pretrain_path=None): generator = UNet(n_channels=input_channels, layer_nums=layer_nums, output_channel=output_channels, bn=bn) discriminator = PixelDiscriminator(output_channels, discriminator_num_filters, use_norm=False) generator = generator.cuda() discriminator = discriminator.cuda() flow_network = Network() flow_network.load_state_dict(torch.load(lite_flow_model_path)) flow_network.cuda().eval() adversarial_loss = Adversarial_Loss().cuda() discriminate_loss = Discriminate_Loss().cuda() gd_loss = Gradient_Loss(alpha, num_channels).cuda() op_loss = Flow_Loss().cuda() int_loss = Intensity_Loss(l_num).cuda() step = 0 if not pretrain: generator.apply(weights_init_normal) discriminator.apply(weights_init_normal) else: assert (generator_pretrain_path != None and discriminator_pretrain_path != None) generator.load_state_dict(torch.load(generator_pretrain_path)) discriminator.load_state_dict(torch.load(discriminator_pretrain_path)) step = int(generator_pretrain_path.split('-')[-1]) print('pretrained model loaded!') print('initializing the model with Generator-Unet {} layers,' 'PixelDiscriminator with filters {} '.format( layer_nums, discriminator_num_filters)) optimizer_G = torch.optim.Adam(generator.parameters(), lr=g_lr) optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=d_lr) writer = SummaryWriter(writer_path) dataset = img_dataset.ano_pred_Dataset(training_data_folder, frame_num) dataset_loader = DataLoader(dataset=dataset, batch_size=batch_size, shuffle=True, num_workers=1, drop_last=True) test_dataset = img_dataset.ano_pred_Dataset(testing_data_folder, frame_num) test_dataloader = DataLoader(dataset=test_dataset, batch_size=batch_size, shuffle=True, num_workers=1, drop_last=True) for epoch in range(epochs): for (input, _), (test_input, _) in zip(dataset_loader, test_dataloader): # generator = generator.train() # discriminator = discriminator.train() target = input[:, -1, :, :, :].cuda() input = input[:, :-1, ] input_last = input[:, -1, ].cuda() input = input.view(input.shape[0], -1, input.shape[-2], input.shape[-1]).cuda() test_target = test_input[:, -1, ].cuda() test_input = test_input[:, :-1].view(test_input.shape[0], -1, test_input.shape[-2], test_input.shape[-1]).cuda() #------- update optim_G -------------- G_output = generator(input) pred_flow_esti_tensor = torch.cat([input_last, G_output], 1) gt_flow_esti_tensor = torch.cat([input_last, target], 1) flow_gt = batch_estimate(gt_flow_esti_tensor, flow_network) flow_pred = batch_estimate(pred_flow_esti_tensor, flow_network) g_adv_loss = adversarial_loss(discriminator(G_output)) g_op_loss = op_loss(flow_pred, flow_gt) g_int_loss = int_loss(G_output, target) g_gd_loss = gd_loss(G_output, target) g_loss = lam_adv * g_adv_loss + lam_gd * g_gd_loss + lam_op * g_op_loss + lam_int * g_int_loss optimizer_G.zero_grad() g_loss.backward() optimizer_G.step() train_psnr = psnr_error(G_output, target) #----------- update optim_D ------- optimizer_D.zero_grad() d_loss = discriminate_loss(discriminator(target), discriminator(G_output.detach())) #d_loss.requires_grad=True d_loss.backward() optimizer_D.step() #----------- cal psnr -------------- test_generator = generator.eval() test_output = test_generator(test_input) test_psnr = psnr_error(test_output, test_target).cuda() if step % 10 == 0: print("[{}/{}]: g_loss: {} d_loss {}".format( step, epoch, g_loss, d_loss)) print('\t gd_loss {}, op_loss {}, int_loss {} ,'.format( g_gd_loss, g_op_loss, g_int_loss)) print('\t train psnr{},test_psnr {}'.format( train_psnr, test_psnr)) writer.add_scalar('psnr/train_psnr', train_psnr, global_step=step) writer.add_scalar('psnr/test_psnr', test_psnr, global_step=step) writer.add_scalar('total_loss/g_loss', g_loss, global_step=step) writer.add_scalar('total_loss/d_loss', d_loss, global_step=step) writer.add_scalar('g_loss/adv_loss', g_adv_loss, global_step=step) writer.add_scalar('g_loss/op_loss', g_op_loss, global_step=step) writer.add_scalar('g_loss/int_loss', g_int_loss, global_step=step) writer.add_scalar('g_loss/gd_loss', g_gd_loss, global_step=step) writer.add_image('image/train_target', target[0], global_step=step) writer.add_image('image/train_output', G_output[0], global_step=step) writer.add_image('image/test_target', test_target[0], global_step=step) writer.add_image('image/test_output', test_output[0], global_step=step) step += 1 if step % 500 == 0: utils.saver(generator.state_dict(), model_generator_save_path, step, max_to_save=10) utils.saver(discriminator.state_dict(), model_discriminator_save_path, step, max_to_save=10) if step >= 2000: print('==== begin evaluate the model of {} ===='.format( model_generator_save_path + '-' + str(step))) auc = evaluate(frame_num=5, layer_nums=4, input_channels=12, output_channels=3, model_path=model_generator_save_path + '-' + str(step), evaluate_name='compute_auc') writer.add_scalar('results/auc', auc, global_step=step)
def evaluate_performance(args, gridargs, logger): ''' -------------------------Hyperparameters-------------------------- ''' EPOCHS = args.epochs ITER = args.iterations # per epoch LR = gridargs['lr'] MOM = gridargs['mom'] # LOGInterval = args.log_interval BATCHSIZE = args.batch_size NUMBER_OF_WORKERS = args.workers DATA_FOLDER = args.data ROOT = gridargs['run'] CUSTOM_LOG_DIR = os.path.join(ROOT, "additionalLOGS") # check existance of data if not os.path.isdir(DATA_FOLDER): print("data folder not existant or in wrong layout.\n\t", DATA_FOLDER) exit(0) ''' ---------------------------preparations--------------------------- ''' # CUDA for PyTorch use_cuda = torch.cuda.is_available() device = torch.device("cuda:0" if use_cuda else "cpu") print("using device: ", str(device)) ''' ---------------------------loading dataset and normalizing--------------------------- ''' # Dataloader Parameters train_params = {'batch_size': BATCHSIZE, 'shuffle': True, 'num_workers': NUMBER_OF_WORKERS} test_params = {'batch_size': BATCHSIZE, 'shuffle': False, 'num_workers': NUMBER_OF_WORKERS} # create a folder for the weights and custom logs if not os.path.isdir(CUSTOM_LOG_DIR): os.makedirs(CUSTOM_LOG_DIR) traindata = CONRADataset(DATA_FOLDER, True, device=device, precompute=True, transform=None) testdata = CONRADataset(DATA_FOLDER, False, device=device, precompute=True, transform=None) trainingset = DataLoader(traindata, **train_params) testset = DataLoader(testdata, **test_params) if args.model == "unet": m = UNet(2, 2).to(device) else: m = simpleConvNet(2, 2).to(device) o = optim.SGD(m.parameters(), lr=LR, momentum=MOM) loss_fn = nn.MSELoss() test_loss = None train_loss = None ''' -----------------------------training----------------------------- ''' global_step = 0 # calculating initial loss if test_loss is None or train_loss is None: print("calculating initial loss") m.eval() print("testset...") test_loss = calculate_loss(set=testset, loss_fn=loss_fn, length_set=len(testdata), dev=device, model=m) print("trainset...") train_loss = calculate_loss(set=trainingset, loss_fn=loss_fn, length_set=len(traindata), dev=device, model=m) # printing runtime information print("starting training at {} for {} epochs {} iterations each\n\t{} total".format(0, EPOCHS, ITER, EPOCHS * ITER)) print("\tbatchsize: {}\n\tloss: {}\n".format(BATCHSIZE, train_loss)) print("\tmodel: {}\n\tlearningrate: {}\n\tmomentum: {}\n\tnorming output space: {}".format(args.model, LR, MOM, False)) #start actual training loops for e in range(0, EPOCHS): # iterations will not be interupted with validation and metrics for i in range(ITER): global_step = (e * ITER) + i # training m.train() iteration_loss = 0 for x, y in tqdm(trainingset): x, y = x.to(device=device, dtype=torch.float), y.to(device=device, dtype=torch.float) pred = m(x) loss = loss_fn(pred, y) iteration_loss += loss.item() o.zero_grad() loss.backward() o.step() print("\niteration {}: --accumulated loss {}".format(global_step, iteration_loss)) if not np.isfinite(iteration_loss): print("EXPLODING OR VANISHING GRADIENT at lr: {} mom: {} step: {}".format(LR, MOM, global_step)) return # validation, saving and logging print("\nvalidating") m.eval() # disable dropout batchnorm etc print("testset...") test_loss = calculate_loss(set=testset, loss_fn=loss_fn, length_set=len(testdata), dev=device, model=m) print("trainset...") train_loss = calculate_loss(set=trainingset, loss_fn=loss_fn, length_set=len(traindata), dev=device, model=m) print("calculating performace...") currSSIM, currR = performance(set=testset, dev=device, model=m, bs=BATCHSIZE) print("SSIM (iod/water): {}/{}\nR (iod/water): {}/{}".format(currSSIM[0], currSSIM[1], currR[0], currR[1])) #f.write("num, lr, mom, step, ssimIOD, ssimWAT, rIOD, rWAT, trainLOSS, testLOSS\n") with open(gridargs['stats'], 'a') as f: newCSVline = "{}, {}, {}, {}, {}, {}, {}, {}, {}, {}\n".format(gridargs['runnum'], LR, MOM, global_step, currSSIM[0], currSSIM[1], currR[0], currR[1], train_loss, test_loss) f.write(newCSVline) print("wrote new line to csv:\n\t{}".format(newCSVline)) print("advanced metrics") with torch.no_grad(): for x, y in testset: # x, y in shape[2,2,480,620] [b,c,h,w] x, y = x.to(device=device, dtype=torch.float), y.to(device=device, dtype=torch.float) pred = m(x) iod = pred.cpu().numpy()[0, 0, :, :] water = pred.cpu().numpy()[0, 1, :, :] gtiod = y.cpu().numpy()[0, 0, :, :] gtwater = y.cpu().numpy()[0, 1, :, :] IMAGE_LOG_DIR = os.path.join(CUSTOM_LOG_DIR, str(global_step)) if not os.path.isdir(IMAGE_LOG_DIR): os.makedirs(IMAGE_LOG_DIR) plt.imsave(os.path.join(IMAGE_LOG_DIR, 'iod' + str(global_step) + '.png'), iod, cmap='gray') plt.imsave(os.path.join(IMAGE_LOG_DIR, 'water' + str(global_step) + '.png'), water, cmap='gray') plt.imsave(os.path.join(IMAGE_LOG_DIR, 'gtiod' + str(global_step) + '.png'), gtiod, cmap='gray') plt.imsave(os.path.join(IMAGE_LOG_DIR, 'gtwater' + str(global_step) + '.png'), gtwater, cmap='gray') print("creating and saving profile plot at 240") fig2, (ax1, ax2) = plt.subplots(nrows=2, ncols=1) # plot water and iodine in one plot ax1.plot(iod[240]) ax1.plot(gtiod[240]) ax1.title.set_text("iodine horizontal profile") ax1.set_ylabel("mm iodine") ax1.set_ylim([np.min(gtiod), np.max(gtiod)]) print("max value in gtiod is {}".format(np.max(gtiod))) ax2.plot(water[240]) ax2.plot(gtwater[240]) ax2.title.set_text("water horizontal profile") ax2.set_ylabel("mm water") ax2.set_ylim([np.min(gtwater), np.max(gtwater)]) plt.subplots_adjust(wspace=0.3) plt.savefig(os.path.join(IMAGE_LOG_DIR, 'ProfilePlots' + str(global_step) + '.png')) break if logger is not None and train_loss is not None: logger.add_scalar('test_loss', test_loss, global_step=global_step) logger.add_scalar('train_loss', train_loss, global_step=global_step) logger.add_image("iodine-prediction", iod.reshape(1, 480, 620), global_step=global_step) logger.add_image("ground-truth", gtiod.reshape(1, 480, 620), global_step=global_step) # logger.add_image("water-prediction", wat) print("\ttensorboard updated with test/train loss and a sample image") # saving final results CHECKPOINT = os.path.join(ROOT, "finalWeights.pt") print("saving upon exit") torch.save({ 'epoch': EPOCHS, 'iterations': ITER, 'model_state_dict': m.state_dict(), 'optimizer_state_dict': o.state_dict(), 'train_loss': train_loss, 'test_loss': test_loss}, CHECKPOINT) print('\tsaved progress to: ', CHECKPOINT) if logger is not None and train_loss is not None: logger.add_scalar('test_loss', test_loss, global_step=global_step) logger.add_scalar('train_loss', train_loss, global_step=global_step)