def main(_): config = tf.ConfigProto() config.gpu_options.allow_growth = True with tf.Session(config=config) as sess: model = UNet(args.experiment_dir, batch_size=args.batch_size, experiment_id=args.experiment_id, input_width=args.image_size, output_width=args.image_size, embedding_num=args.embedding_num, embedding_dim=args.embedding_dim, L1_penalty=args.L1_penalty) model.register_session(sess) if args.flip_labels: model.build_model(is_training=True, inst_norm=args.inst_norm, no_target_source=True) else: model.build_model(is_training=True, inst_norm=args.inst_norm) fine_tune_list = None if args.fine_tune: ids = args.fine_tune.split(",") fine_tune_list = set([int(i) for i in ids]) model.train(lr=args.lr, epoch=args.epoch, resume=args.resume, schedule=args.schedule, freeze_encoder=args.freeze_encoder, fine_tune=fine_tune_list, sample_steps=args.sample_steps, checkpoint_steps=args.checkpoint_steps, flip_labels=args.flip_labels)
def load_finetuned_model(args, baseline_model): """ :param args: :param baseline_model: :return: """ # augment_net = Net(0, 0.0, 32, 3, 0.0, num_classes=32**2 * 3, do_res=True) augment_net = UNet(in_channels=3, n_classes=3, depth=1, wf=2, padding=True, batch_norm=False, do_noise_channel=True, up_mode='upsample', use_identity_residual=True) # TODO(PV): Initialize UNet properly # TODO (JON): DEPTH 1 WORKED WELL. Changed upconv to upsample. Use a wf of 2. # This ResNet outputs scalar weights to be applied element-wise to the per-example losses from models.simple_models import CNN, Net imsize, in_channel, num_classes = 32, 3, 10 reweighting_net = Net(0, 0.0, imsize, in_channel, 0.0, num_classes=1) #resnet_cifar.resnet20(num_classes=1) if args.load_finetune_checkpoint: checkpoint = torch.load(args.load_finetune_checkpoint) baseline_model.load_state_dict(checkpoint['elementary_model_state_dict']) augment_net.load_state_dict(checkpoint['augment_model_state_dict']) try: reweighting_net.load_state_dict(checkpoint['reweighting_model_state_dict']) except KeyError: pass augment_net, reweighting_net, baseline_model = augment_net.cuda(), reweighting_net.cuda(), baseline_model.cuda() augment_net.train(), reweighting_net.train(), baseline_model.train() return augment_net, reweighting_net, baseline_model
class EventGANBase(object): def __init__(self, options): self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') self.generator = UNet(num_input_channels=2*options.n_image_channels, num_output_channels=options.n_time_bins * 2, skip_type='concat', activation='relu', num_encoders=4, base_num_channels=32, num_residual_blocks=2, norm='BN', use_upsample_conv=True, with_activation=True, sn=options.sn, multi=False) latest_checkpoint = get_latest_checkpoint(options.checkpoint_dir) checkpoint = torch.load(latest_checkpoint) self.generator.load_state_dict(checkpoint["gen"]) self.generator.to(self.device) def forward(self, images, is_train=False): if len(images.shape) == 3: images = images[None, ...] assert len(images.shape) == 4 and images.shape[1] == 2, \ "Input images must be either 2xHxW or Bx2xHxW." if not is_train: with torch.no_grad(): self.generator.eval() event_volume = self.generator(images) self.generator.train() else: event_volume = self.generator(images) return event_volume
def load_finetuned_model(self, baseline_model): """ Loads the augmentation net, sample reweighting net, and baseline model Note: sets all these models to train mode """ # augment_net = Net(0, 0.0, 32, 3, 0.0, num_classes=32**2 * 3, do_res=True) if self.args.dataset == DATASET_MNIST: imsize, in_channel, num_classes = 28, 1, 10 else: imsize, in_channel, num_classes = 32, 3, 10 augment_net = UNet( in_channels=in_channel, n_classes=in_channel, depth=2, wf=3, padding=True, batch_norm=False, do_noise_channel=True, up_mode='upconv', use_identity_residual=True) # TODO(PV): Initialize UNet properly # TODO (JON): DEPTH 1 WORKED WELL. Changed upconv to upsample. Use a wf of 2. # This ResNet outputs scalar weights to be applied element-wise to the per-example losses reweighting_net = Net(1, 0.0, imsize, in_channel, 0.0, num_classes=1) # resnet_cifar.resnet20(num_classes=1) if self.args.load_finetune_checkpoint: checkpoint = torch.load(self.args.load_finetune_checkpoint) # temp_baseline_model = baseline_model # baseline_model.load_state_dict(checkpoint['elementary_model_state_dict']) if 'weight_decay' in checkpoint: baseline_model.weight_decay = checkpoint['weight_decay'] # baseline_model.weight_decay = temp_baseline_model.weight_decay # baseline_model.load_state_dict(checkpoint['elementary_model_state_dict']) augment_net.load_state_dict(checkpoint['augment_model_state_dict']) try: reweighting_net.load_state_dict( checkpoint['reweighting_model_state_dict']) except KeyError: pass augment_net, reweighting_net, baseline_model = augment_net.to( self.device), reweighting_net.to(self.device), baseline_model.to( self.device) augment_net.train(), reweighting_net.train(), baseline_model.train() return augment_net, reweighting_net, baseline_model
def train(): # Init data train_dataset, val_dataset = prepare_datasets() train_loader = DataLoader(train_dataset, batch_size=10, shuffle=True) val_loader = DataLoader(val_dataset, batch_size=10, shuffle=True) loaders = dict(train=train_loader, val=val_loader) # Init Model model = UNet().cuda() optimizer = torch.optim.Adam(model.parameters(), lr=1e-3, amsgrad=True) scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer=optimizer, gamma=0.984) loss_fn = nn.BCELoss() epochs = 500 for epoch in range(epochs): for phase in 'train val'.split(): if phase == 'train': model = model.train() torch.set_grad_enabled(True) else: model = model.eval() torch.set_grad_enabled(False) loader = loaders[phase] epoch_losses = dict(train=[], val=[]) running_loss = [] for batch in loader: imgs, masks = batch imgs = imgs.cuda() masks = masks.cuda() outputs = model(imgs) loss = loss_fn(outputs, masks) running_loss.append(loss.item()) if phase == 'train': optimizer.zero_grad() loss.backward() optimizer.step() # End of Epoch print(f'{epoch}) {phase} loss: {np.mean(running_loss)}') visualize_results(loader, model, epoch, phase) epoch_losses[phase].append(np.mean(running_loss)) tensorboard(epoch_losses[phase], phase) if phase == 'train': scheduler.step()
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))
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)
torch.backends.cudnn.benchmark = True train_dataset = Seg_dataset(cfg) train_loader = data.DataLoader(train_dataset, batch_size=cfg.bs, shuffle=True, num_workers=8, pin_memory=True, drop_last=False) if cfg.model == 'unet': model = UNet(input_channels=3).cuda() model.apply(model.weights_init_normal) else: model = DLASeg(cfg).cuda() model.train() if cfg.resume: resume_epoch = int(cfg.resume.split('.')[0].split('_')[1]) + 1 model.load_state_dict(torch.load('weights/' + cfg.resume), strict=True) print(f'Resume training with \'{cfg.resume}\'.') else: resume_epoch = 0 print('Training with ImageNet pre-trained weights.') criterion = nn.CrossEntropyLoss(ignore_index=255).cuda() if cfg.optim == 'sgd': optimizer = torch.optim.SGD(model.optim_parameters(), cfg.lr, cfg.momentum, weight_decay=cfg.decay)
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)
train_loss_seg_a = [] train_loss_seg_b = [] train_dice = [] val_loss_a = [] val_dice_a = [] val_loss_b = [] val_dice_b = [] for e in range(epochs): epoch_train_loss_rec = [] epoch_train_loss_seg = [] dice_scores = [] net.train() pseudo.train() print('Epoch ', e) for i, data in enumerate(tqdm.tqdm(train_loader)): iteration += batch_size optimiser_ps.zero_grad() optimiser_net.zero_grad() # either train pseudolabeller or the net # first 10 epochs train the pseudo labeller on edges if e < epochs_pseudo: edges_a = data['A'][2].cuda() target_a = data['A'][1].cuda()
with open(path, mode=mode) as f: f.writelines(context + "\n") cuda0 = torch.device('cuda:0') net = UNet(27).to(cuda0) criterion = nn.MSELoss().to(cuda0) optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9) train_data_laoder, test_data_loader = get_dataloader(batch_size=4) write_log("weight/train.log", str(datetime.datetime.now()), "w") write_log("weight/train.log", "train start") print(net) for epoch in range(300): # train phase running_loss = 0.0 net.train() for i, batch in enumerate(train_data_laoder): inputs = batch['image'].to(cuda0) target = batch['target'].to(cuda0) 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()
def train(model_name=''): # Init data train_dataset, val_dataset = prepare_datasets() train_loader = DataLoader(train_dataset, batch_size=10, shuffle=True) val_loader = DataLoader(val_dataset, batch_size=10, shuffle=True) loaders = dict(train=train_loader, val=val_loader) # Init Model if model_name == '': model = UNet().cuda() else: model = data_utils.load_model(model_name).cuda() optimizer = torch.optim.Adam(model.parameters(), lr=1e-3, amsgrad=True) scheduler = torch.optim.lr_scheduler.ExponentialLR(optimizer=optimizer, gamma=0.984) loss_fn = nn.BCELoss() epochs = 500 epoch_losses = dict(train=[], val=[]) for epoch in range(epochs): for phase in 'train val'.split(): if phase == 'train': model = model.train() torch.set_grad_enabled(True) else: model = model.eval() torch.set_grad_enabled(False) loader = loaders[phase] running_loss = [] for batch in loader: imgs, masks = batch imgs = imgs.cuda() masks = masks.cuda() outputs = model(imgs) loss = loss_fn(outputs, masks) running_loss.append(loss.item()) if phase == 'train': optimizer.zero_grad() loss.backward() optimizer.step() # End of Epoch print(f'{epoch}) {phase} loss: {np.mean(running_loss)}') visualize_results(loader, model, epoch, phase) if epoch % 10 == 0: results_dir = 'weight/' if not os.path.isdir(results_dir): os.makedirs(results_dir) data_utils.save_model(model, results_dir + f'model_{epoch}.pt') epoch_losses[phase].append(np.mean(running_loss)) if phase == 'val': df = pd.DataFrame(data=epoch_losses) df.to_csv('loss.csv') tensorboard(epoch_losses[phase], phase) if phase == 'train': scheduler.step()
adversarial_loss = Adversarial_Loss().cuda() discriminate_loss = Discriminate_Loss().cuda() gradient_loss = Gradient_Loss(3).cuda() flow_loss = Flow_Loss().cuda() intensity_loss = Intensity_Loss().cuda() train_dataset = Dataset.train_dataset(train_cfg) # Remember to set drop_last=True, because we need to use 4 frames to predict one frame. train_dataloader = DataLoader(dataset=train_dataset, batch_size=train_cfg.batch_size, shuffle=True, num_workers=4, drop_last=True) writer = SummaryWriter(f'tensorboard_log/{train_cfg.dataset}_bs{train_cfg.batch_size}') start_iter = int(train_cfg.resume.split('_')[-1].split('.')[0]) if train_cfg.resume else 0 training = True generator = generator.train() discriminator = discriminator.train() try: step = start_iter while training: for indice, clips, flow_strs in train_dataloader: input_frames = clips[:, 0:12, :, :].cuda() # (n, 12, 256, 256) target_frame = clips[:, 12:15, :, :].cuda() # (n, 3, 256, 256) input_last = input_frames[:, 9:12, :, :].cuda() # use for flow_loss # pop() the used frame index, this can't work in train_dataset.__getitem__ because of multiprocessing. for index in indice: train_dataset.all_seqs[index].pop() if len(train_dataset.all_seqs[index]) == 0: train_dataset.all_seqs[index] = list(range(len(train_dataset.videos[index]) - 4))
def main(): """Create the model and start the training.""" args = get_arguments() cudnn.enabled = True n_discriminators = 5 # create teacher & student student_net = UNet(3, n_classes=args.num_classes) teacher_net = UNet(3, n_classes=args.num_classes) student_params = list(student_net.parameters()) # teacher doesn't need gradient as it's just a EMA of the student teacher_params = list(teacher_net.parameters()) for param in teacher_params: param.requires_grad = False student_net.train() student_net.cuda(args.gpu) teacher_net.train() teacher_net.cuda(args.gpu) cudnn.benchmark = True unsup_weights = [ args.unsup_weight5, args.unsup_weight6, args.unsup_weight7, args.unsup_weight8, args.unsup_weight9 ] lambda_adv_tgts = [ args.lambda_adv_tgt5, args.lambda_adv_tgt6, args.lambda_adv_tgt7, args.lambda_adv_tgt8, args.lambda_adv_tgt9 ] # create a list of discriminators discriminators = [] for dis_idx in range(n_discriminators): discriminators.append(FCDiscriminator(num_classes=args.num_classes)) discriminators[dis_idx].train() discriminators[dis_idx].cuda(args.gpu) if not os.path.exists(args.snapshot_dir): os.makedirs(args.snapshot_dir) max_iters = args.num_steps * args.iter_size * args.batch_size src_set = REFUGE(True, domain='REFUGE_SRC', is_transform=True, augmentations=aug_student, aug_for_target=aug_teacher, max_iters=max_iters) src_loader = data.DataLoader(src_set, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True) src_loader_iter = enumerate(src_loader) tgt_set = REFUGE(True, domain='REFUGE_DST', is_transform=True, augmentations=aug_student, aug_for_target=aug_teacher, max_iters=max_iters) tgt_loader = data.DataLoader(tgt_set, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers, pin_memory=True) tgt_loader_iter = enumerate(tgt_loader) student_optimizer = optim.SGD(student_params, lr=args.learning_rate, momentum=args.momentum, weight_decay=args.weight_decay) teacher_optimizer = optim_weight_ema.WeightEMA(teacher_params, student_params, alpha=args.teacher_alpha) d_optimizers = [] for idx in range(n_discriminators): optimizer = optim.Adam(discriminators[idx].parameters(), lr=args.learning_rate_D, betas=(0.9, 0.99)) d_optimizers.append(optimizer) calc_bce_loss = torch.nn.BCEWithLogitsLoss() # labels for adversarial training source_label, tgt_label = 0, 1 for i_iter in range(args.num_steps): total_seg_loss = 0 seg_loss_vals = [0] * n_discriminators adv_tgt_loss_vals = [0] * n_discriminators d_loss_vals = [0] * n_discriminators unsup_loss_vals = [0] * n_discriminators for d_optimizer in d_optimizers: d_optimizer.zero_grad() adjust_learning_rate_D(d_optimizer, i_iter, args) student_optimizer.zero_grad() adjust_learning_rate(student_optimizer, i_iter, args) for sub_i in range(args.iter_size): # ******** Optimize source network with segmentation loss ******** # As we don't change the discriminators, their parameters are fixed for discriminator in discriminators: for param in discriminator.parameters(): param.requires_grad = False _, src_batch = src_loader_iter.__next__() _, _, src_images, src_labels, _ = src_batch src_images = Variable(src_images).cuda(args.gpu) # calculate the segmentation losses sup_preds = list(student_net(src_images)) seg_losses, total_seg_loss = [], 0 for idx, sup_pred in enumerate(sup_preds): sup_interp_pred = (sup_pred) # you also can use dice loss like: dice_loss(src_labels, sup_interp_pred) seg_loss = Weighted_Jaccard_loss(src_labels, sup_interp_pred, args.class_weights, args.gpu) seg_losses.append(seg_loss) total_seg_loss += seg_loss * unsup_weights[idx] seg_loss_vals[idx] += seg_loss.item() / args.iter_size _, tgt_batch = tgt_loader_iter.__next__() tgt_images0, tgt_lbl0, tgt_images1, tgt_lbl1, _ = tgt_batch tgt_images0 = Variable(tgt_images0).cuda(args.gpu) tgt_images1 = Variable(tgt_images1).cuda(args.gpu) # calculate ensemble losses stu_unsup_preds = list(student_net(tgt_images1)) tea_unsup_preds = teacher_net(tgt_images0) total_mse_loss = 0 for idx in range(n_discriminators): stu_unsup_probs = F.softmax(stu_unsup_preds[idx], dim=-1) tea_unsup_probs = F.softmax(tea_unsup_preds[idx], dim=-1) unsup_loss = calc_mse_loss(stu_unsup_probs, tea_unsup_probs, args.batch_size) unsup_loss_vals[idx] += unsup_loss.item() / args.iter_size total_mse_loss += unsup_loss * unsup_weights[idx] total_mse_loss = total_mse_loss / args.iter_size # As the requires_grad is set to False in the discriminator, the # gradients are only accumulated in the generator, the target # student network is optimized to make the outputs of target domain # images close to the outputs of source domain images stu_unsup_preds = list(student_net(tgt_images0)) d_outs, total_adv_loss = [], 0 for idx in range(n_discriminators): stu_unsup_interp_pred = (stu_unsup_preds[idx]) d_outs.append(discriminators[idx](stu_unsup_interp_pred)) label_size = d_outs[idx].data.size() labels = torch.FloatTensor(label_size).fill_(source_label) labels = Variable(labels).cuda(args.gpu) adv_tgt_loss = calc_bce_loss(d_outs[idx], labels) total_adv_loss += lambda_adv_tgts[idx] * adv_tgt_loss adv_tgt_loss_vals[idx] += adv_tgt_loss.item() / args.iter_size total_adv_loss = total_adv_loss / args.iter_size # requires_grad is set to True in the discriminator, we only # accumulate gradients in the discriminators, the discriminators are # optimized to make true predictions d_losses = [] for idx in range(n_discriminators): discriminator = discriminators[idx] for param in discriminator.parameters(): param.requires_grad = True sup_preds[idx] = sup_preds[idx].detach() d_outs[idx] = discriminators[idx](sup_preds[idx]) label_size = d_outs[idx].data.size() labels = torch.FloatTensor(label_size).fill_(source_label) labels = Variable(labels).cuda(args.gpu) d_losses.append(calc_bce_loss(d_outs[idx], labels)) d_losses[idx] = d_losses[idx] / args.iter_size / 2 d_losses[idx].backward() d_loss_vals[idx] += d_losses[idx].item() for idx in range(n_discriminators): stu_unsup_preds[idx] = stu_unsup_preds[idx].detach() d_outs[idx] = discriminators[idx](stu_unsup_preds[idx]) label_size = d_outs[idx].data.size() labels = torch.FloatTensor(label_size).fill_(tgt_label) labels = Variable(labels).cuda(args.gpu) d_losses[idx] = calc_bce_loss(d_outs[idx], labels) d_losses[idx] = d_losses[idx] / args.iter_size / 2 d_losses[idx].backward() d_loss_vals[idx] += d_losses[idx].item() for d_optimizer in d_optimizers: d_optimizer.step() total_loss = total_seg_loss + total_adv_loss + total_mse_loss total_loss.backward() student_optimizer.step() teacher_optimizer.step() log_str = 'iter = {0:7d}/{1:7d}'.format(i_iter, args.num_steps) log_str += ', total_seg_loss = {0:.3f} '.format(total_seg_loss) templ = 'seg_losses = [' + ', '.join(['%.2f'] * len(seg_loss_vals)) log_str += templ % tuple(seg_loss_vals) + '] ' templ = 'ens_losses = [' + ', '.join(['%.5f'] * len(unsup_loss_vals)) log_str += templ % tuple(unsup_loss_vals) + '] ' templ = 'adv_losses = [' + ', '.join(['%.2f'] * len(adv_tgt_loss_vals)) log_str += templ % tuple(adv_tgt_loss_vals) + '] ' templ = 'd_losses = [' + ', '.join(['%.2f'] * len(d_loss_vals)) log_str += templ % tuple(d_loss_vals) + '] ' print(log_str) if i_iter >= args.num_steps_stop - 1: print('save model ...') filename = 'UNet' + str( args.num_steps_stop) + '_v18_weightedclass.pth' torch.save(teacher_net.cpu().state_dict(), os.path.join(args.snapshot_dir, filename)) break if i_iter % args.save_pred_every == 0 and i_iter != 0: print('taking snapshot ...') filename = 'UNet' + str(i_iter) + '_v18_weightedclass.pth' torch.save(teacher_net.cpu().state_dict(), os.path.join(args.snapshot_dir, filename)) teacher_net.cuda(args.gpu)
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)
def main(): """Create the model and start the evaluation process.""" args = get_arguments() gpu0 = args.gpu if not os.path.exists(args.save): os.makedirs(args.save) model = UNet(3, n_classes=args.num_classes) saved_state_dict = torch.load(args.restore_from) model.load_state_dict(saved_state_dict) model.cuda(gpu0) model.train() testloader = data.DataLoader(REFUGE(False, domain='REFUGE_TEST', is_transform=True), batch_size=args.batch_size, shuffle=False, pin_memory=True) if version.parse(torch.__version__) >= version.parse('0.4.0'): interp = nn.Upsample(size=(460, 460), mode='bilinear', align_corners=True) else: interp = nn.Upsample(size=(460, 460), mode='bilinear') for index, batch in enumerate(testloader): if index % 100 == 0: print('%d processd' % index) image, label, _, _, name = batch if args.model == 'Unet': _, _, _, _, output2 = model( Variable(image, volatile=True).cuda(gpu0)) output = interp(output2).cpu().data.numpy() for idx, one_name in enumerate(name): pred = output[idx] pred = pred.transpose(1, 2, 0) pred = np.asarray(np.argmax(pred, axis=2), dtype=np.uint8) output_col = colorize_mask(pred) if is_polar: # plt.imshow(output_col) # plt.show() output_col = np.array(output_col) output_col[output_col == 0] = 0 output_col[output_col == 1] = 128 output_col[output_col == 2] = 255 # plt.imshow(output_col) # plt.show() output_col = cv2.linearPolar( rotate(output_col, 90), (args.ROI_size / 2, args.ROI_size / 2), args.ROI_size / 2, cv2.WARP_FILL_OUTLIERS + cv2.WARP_INVERSE_MAP) # plt.imshow(output_col) # plt.show() output_col = np.array(output_col * 255, dtype=np.uint8) output_col[output_col > 200] = 210 output_col[output_col == 0] = 255 output_col[output_col == 210] = 0 output_col[(output_col > 0) & (output_col < 255)] = 128 output_col = Image.fromarray(output_col) # plt.imshow(output_col) # plt.show() one_name = one_name.split('/')[-1] if len(one_name.split('_')) > 0: one_name = one_name[:-4] #pred.save('%s/%s.bmp' % (args.save, one_name)) output_col = output_col.convert('L') print(output_col.size) output_col.save('%s/%s.bmp' % (args.save, one_name.split('.')[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
def train_UNet(): cfg = UnetConfig() train_transform = transforms.Compose([ GrayscaleNormalization(mean=0.5, std=0.5), RandomRotation(), RandomFlip(), ToTensor(), ]) val_transform = transforms.Compose([ GrayscaleNormalization(mean=0.5, std=0.5), ToTensor(), ]) # Set Dataset train_dataset = Dataset(imgs_dir=TRAIN_IMGS_DIR, labels_dir=TRAIN_LABELS_DIR, transform=train_transform) train_loader = DataLoader(train_dataset, batch_size=cfg.BATCH_SIZE, shuffle=True, num_workers=0) val_dataset = Dataset(imgs_dir=VAL_IMGS_DIR, labels_dir=VAL_LABELS_DIR, transform=val_transform) val_loader = DataLoader(val_dataset, batch_size=cfg.BATCH_SIZE, shuffle=False, num_workers=0) train_data_num = len(train_dataset) val_data_num = len(val_dataset) train_batch_num = int(np.ceil(train_data_num / cfg.BATCH_SIZE)) # np.ceil val_batch_num = int(np.ceil(val_data_num / cfg.BATCH_SIZE)) # Network net = UNet().to(device) print(count_parameters(net)) # Loss Function loss_fn = nn.BCEWithLogitsLoss().to(device) # Optimizer optim = torch.optim.Adam(params=net.parameters(), lr=cfg.LEARNING_RATE) # Tensorboard # train_writer = SummaryWriter(log_dir=TRAIN_LOG_DIR) # val_writer = SummaryWriter(log_dir=VAL_LOG_DIR) # Training start_epoch = 0 # Load Checkpoint File if os.listdir(os.path.join(CKPT_DIR, 'unet')): net, optim, start_epoch = load_net(ckpt_dir=os.path.join( CKPT_DIR, 'unet'), net=net, optim=optim) else: print('* Training from scratch') num_epochs = cfg.NUM_EPOCHS for epoch in range(start_epoch + 1, num_epochs + 1): net.train() train_loss_arr = list() for batch_idx, data in enumerate(train_loader, 1): # Forward Propagation img = data['img'].to(device) label = data['label'].to(device) output = net(img) # Backward Propagation optim.zero_grad() loss = loss_fn(output, label) loss.backward() optim.step() # Calc Loss Function train_loss_arr.append(loss.item()) print_form = '[Train] | Epoch: {:0>4d} / {:0>4d} | Batch: {:0>4d} / {:0>4d} | Loss: {:.4f}' print( print_form.format(epoch, num_epochs, batch_idx, train_batch_num, train_loss_arr[-1])) train_loss_avg = np.mean(train_loss_arr) # train_writer.add_scalar(tag='loss', scalar_value=train_loss_avg, global_step=epoch) # Validation (No Back Propagation) with torch.no_grad(): net.eval() # Evaluation Mode val_loss_arr = list() for batch_idx, data in enumerate(val_loader, 1): # Forward Propagation img = data['img'].to(device) label = data['label'].to(device) output = net(img) # Calc Loss Function loss = loss_fn(output, label) val_loss_arr.append(loss.item()) print_form = '[Validation] | Epoch: {:0>4d} / {:0>4d} | Batch: {:0>4d} / {:0>4d} | Loss: {:.4f}' print( print_form.format(epoch, num_epochs, batch_idx, val_batch_num, val_loss_arr[-1])) val_loss_avg = np.mean(val_loss_arr) # val_writer.add_scalar(tag='loss', scalar_value=val_loss_avg, global_step=epoch) print_form = '[Epoch {:0>4d}] Training Avg Loss: {:.4f} | Validation Avg Loss: {:.4f}' print(print_form.format(epoch, train_loss_avg, val_loss_avg)) if epoch % 10 == 0: save_net(ckpt_dir=os.path.join(CKPT_DIR, 'unet'), net=net, optim=optim, epoch=epoch)
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("--load-weights", type=str, default=None, help="load pretrained weights") parser.add_argument("--load-model", type=str, default=None, help="load pretrained model, entire model (filepath, default: None)") parser.add_argument("--debug", action="store_true", default=False) parser.add_argument('--epochs', type=int, default=30, 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 ? # https://pytorch.org/tutorials/beginner/former_torchies/parallelism_tutorial.html 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: DONE. # transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2), # Use OpenCVRotation, OpenCVXXX, ... (our implementation) # OpenCVRotation((-10, 10)), # angles (in degree) transforms.ToTensor(), # already done in the dataloader transform_normalize ]) geo_transform = GeoCompose([ OpenCVRotation(angles=(-10, 10), scales=(0.9, 1.1), centers=(-0.05, 0.05)), # TODO add more data augmentation here ]) def worker_init_fn(worker_id): # WARNING spawn start method is used, # worker_init_fn cannot be an unpicklable object, e.g., a lambda function. # A work-around for issue #5059: https://github.com/pytorch/pytorch/issues/5059 np.random.seed() data_loader_train = {'batch_size': args.batch_size, 'shuffle': args.shuffle, 'num_workers': args.num_cpu, # 'sampler': balanced_sampler, 'drop_last': True, # for GAN-like 'pin_memory': False, 'worker_init_fn': worker_init_fn, } data_loader_valid = {'batch_size': args.batch_size, 'shuffle': False, 'num_workers': args.num_cpu, 'drop_last': False, 'pin_memory': False, } train_set = LiTSDataset(args.train_filepath, dtype=np.float32, geometry_transform=geo_transform, # TODO enable data augmentation pixelwise_transform=data_transform, ) valid_set = LiTSDataset(args.val_filepath, dtype=np.float32, pixelwise_transform=data_transform, ) dataloader_train = torch.utils.data.DataLoader(train_set, **data_loader_train) dataloader_valid = torch.utils.data.DataLoader(valid_set, **data_loader_valid) # =================== Build model =================== # TODO: control the model by bash command if args.load_weights: model = UNet(in_ch=1, out_ch=3, # there are 3 classes: 0: background, 1: liver, 2: tumor depth=4, start_ch=32, # 64 inc_rate=2, kernel_size=5, # 3 padding=True, batch_norm=True, spec_norm=False, dropout=0.5, up_mode='upconv', include_top=True, include_last_act=False, ) printYellow(f"Loading pretrained weights from: {args.load_weights}...") model.load_state_dict(torch.load(args.load_weights)) printYellow("+ Done.") elif args.load_model: # load entire model model = torch.load(args.load_model) printYellow("Successfully loaded pretrained model.") model.to(device) optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, betas=(0.9, 0.95)) # TODO best_valid_loss = float('inf') # TODO TODO: add learning decay 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: use torch.utils.tensorboard !! 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) # shape (N, 3, 512, 512) # label_weights is determined according the liver_ratio & tumor_ratio # loss = CrossEntropyLoss(msk_pred, msk, label_weights=[1., 10., 100.], device=device) loss = DiceLoss(msk_pred, msk, label_weights=[1., 20., 50.], device=device) # loss = DiceLoss(msk_pred, msk, label_weights=[1., 20., 500.], 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) # save the entire model torch.save(model, best_model_path) return best_valid_loss
train_loss_seg_a = [] train_loss_seg_b = [] train_dice = [] val_loss_a = [] val_dice_a = [] val_loss_b = [] val_dice_b = [] for e in range(epochs): epoch_train_loss_rec = [] epoch_train_loss_seg = [] dice_scores = [] net.train() pseudo.train() print('Epoch ', e) for i, data in enumerate(tqdm.tqdm(train_loader)): iteration += batch_size optimiser_ps.zero_grad() optimiser_net.zero_grad() # either train pseudolabeller or the net # first 10 epochs train the pseudo labeller on edges if e < epochs_pseudo: edges_a = data['A'][2].cuda() target_a = data['A'][1].cuda()