def validate(state_dict_path, use_gpu, device): model = UNet(n_channels=1, n_classes=2) model.load_state_dict(torch.load(state_dict_path, map_location='cpu' if not use_gpu else device)) model.to(device) val_transforms = transforms.Compose([ ToTensor(), NormalizeBRATS()]) BraTS_val_ds = BRATS2018('./BRATS2018',\ data_set='val',\ seg_type='et',\ scan_type='t1ce',\ transform=val_transforms) data_loader = DataLoader(BraTS_val_ds, batch_size=2, shuffle=False, num_workers=0) running_dice_score = 0. for batch_ind, batch in enumerate(data_loader): imgs, targets = batch imgs = imgs.to(device) targets = targets.to(device) model.eval() with torch.no_grad(): outputs = model(imgs) preds = torch.argmax(F.softmax(outputs, dim=1), dim=1) running_dice_score += dice_score(preds, targets) * targets.size(0) print('running dice score: {:.6f}'.format(running_dice_score)) dice = running_dice_score / len(BraTS_val_ds) print('mean dice score of the validating set: {:.6f}'.format(dice))
def detect_noise_regions(image, args): # load noise segmentation network (U-Net) unet_model_path = os.path.join(args.checkpoints, 'unet', 'UNet.pth') net = UNet(n_channels=3, n_classes=1).to(device) net.load_state_dict(torch.load(unet_model_path)) net.eval() # predict noise regions predict = predict_img(net, device, image) # search inpaint patches patches, labels, _, absolute_position, relative_position = search_inpaint_area(np.array(image), np.array(predict.convert('RGB'))) # save inpaint patches patches_dir = os.path.join(args.output, 'patches') labels_dir = os.path.join(args.output, 'labels') os.makedirs(patches_dir, exist_ok=True) os.makedirs(labels_dir, exist_ok=True) filename = os.path.basename(args.input).split('.')[0] counter = 0 for patch, label in zip(patches, labels): Image.fromarray(patch).save(os.path.join(patches_dir, '{}-{:0>3d}.png'.format(filename, counter))) Image.fromarray(label).save(os.path.join(labels_dir, '{}-{:0>3d}.png'.format(filename, counter))) counter += 1 return patches_dir, labels_dir, absolute_position, relative_position
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
class Visualizer(object): def __init__(self, input_topic, output_topic, resize_width, resize_height, model_path, force_cpu): self.bridge = CvBridge() self.graph = UNet([3, resize_width, resize_height], 3) self.graph.load_state_dict(torch.load(model_path)) self.force_cpu = force_cpu and torch.cuda.is_available() self.resize_width, self.resize_height = resize_width, resize_height if not self.force_cpu: self.graph.cuda() self.graph.eval() self.to_tensor = transforms.Compose([transforms.ToTensor()]) self.publisher = rospy.Publisher(output_topic, ImMsg, queue_size=1) self.raw_subscriber = rospy.Subscriber(input_topic, CompressedImage, self.image_cb, queue_size=1, buff_size=10**8) def convert_to_tensor(self, image): np_arr = np.fromstring(image.data, np.uint8) image_np = cv2.imdecode(np_arr, cv2.IMREAD_COLOR) image_np = cv2.resize(image_np, dsize=(self.resize_width, self.resize_height)) img_to_tensor = PIL.Image.fromarray(image_np) img_tensor = self.to_tensor(img_to_tensor) if not self.force_cpu: return Variable(img_tensor.unsqueeze(0)).cuda() else: return Variable(img_tensor.unsqueeze(0)) def image_cb(self, image): img_tensor = self.convert_to_tensor(image) # Inference output = self.graph(img_tensor) output_data = output.cpu().data.numpy()[0][0] # # Convert from 32fc1 (0 - 1) to 8uc1 (0 - 255) cv_output = np.uint8(255 * output_data) cv_output = cv2.applyColorMap(cv_output, cv2.COLORMAP_JET) # Convert to ROS message to publish msg_out = self.bridge.cv2_to_imgmsg(cv_output, 'bgr8') msg_out.header.stamp = image.header.stamp self.publisher.publish(msg_out)
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()
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))) test_loss += loss.item() # accuracy target, outputs = target.cpu(), torch.squeeze(outputs.cpu(), dim=1) for tar, out in zip(target, outputs):
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
net_is_3d = False if torch.cuda.device_count() > 1: print("Using", torch.cuda.device_count(), "GPUs.") device_ids = [i for i in range(torch.cuda.device_count())] model = nn.DataParallel(model, device_ids=device_ids) model = model.to(device) if experiment == "Unet": model.load_state_dict(torch.load("best_weights.pth")) elif experiment == "DeepLab": model.load_state_dict(torch.load(f"best_weights_{backbone}_deeplab.pth")) model.eval() eval_images, eval_labels, eval_label_corners = batch_generator( eval_image, eval_label, **windowing_params, return_corners=True) eval_dataset = PlateletDataset(eval_images, eval_labels, train=False) prob_maps = stitch(model, eval_images, eval_labels, eval_label.shape, eval_label_corners, windowing_params, net_is_3d, n_classes, device, channels) stitched_classes = np.argmax(prob_maps, axis=0) # A few plots for sanity check for i in [0, 10, 20]:
def inference(): """Support two mode: evaluation (on valid set) or inference mode (on test-set for submission) """ parser = argparse.ArgumentParser(description="Inference mode") parser.add_argument('-testf', "--test-filepath", type=str, default=None, required=True, help="testing dataset filepath.") parser.add_argument("-eval", "--evaluate", action="store_true", default=False, help="Evaluation mode") parser.add_argument("--load-weights", type=str, default=None, help="Load pretrained weights, torch state_dict() (filepath, default: None)") parser.add_argument("--load-model", type=str, default=None, help="Load pretrained model, entire model (filepath, default: None)") parser.add_argument("--save2dir", type=str, default=None, help="save the prediction labels to the directory (default: None)") parser.add_argument("--debug", action="store_true", default=False) parser.add_argument("--batch-size", type=int, default=32, help="Batch size") 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)') args = parser.parse_args() printYellow("="*10 + " Inference mode. "+"="*10) if args.save2dir: os.makedirs(args.save2dir, exist_ok=True) device = torch.device("cuda:{}".format(args.cuda) if torch.cuda.is_available() and (args.cuda >= 0) else "cpu") transform_normalize = transforms.Normalize(mean=[0.5], std=[0.5]) data_transform = transforms.Compose([ transforms.ToTensor(), transform_normalize ]) data_loader_params = {'batch_size': args.batch_size, 'shuffle': False, 'num_workers': args.num_cpu, 'drop_last': False, 'pin_memory': False } test_set = LiTSDataset(args.test_filepath, dtype=np.float32, pixelwise_transform=data_transform, inference_mode=(not args.evaluate), ) dataloader_test = torch.utils.data.DataLoader(test_set, **data_loader_params) # =================== Build model =================== 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=64, inc_rate=2, kernel_size=3, padding=True, batch_norm=True, spec_norm=False, dropout=0.5, up_mode='upconv', include_top=True, include_last_act=False, ) model.load_state_dict(torch.load(args.load_weights)) printYellow("Successfully loaded pretrained weights.") elif args.load_model: # load entire model model = torch.load(args.load_model) printYellow("Successfully loaded pretrained model.") model.eval() model.to(device) # n_batch_per_epoch = len(dataloader_test) sigmoid_act = torch.nn.Sigmoid() st = time.time() volume_start_index = test_set.volume_start_index spacing = test_set.spacing direction = test_set.direction # use it for the submission offset = test_set.offset msk_pred_buffer = [] if args.evaluate: msk_gt_buffer = [] for data_batch in tqdm(dataloader_test): # import ipdb # ipdb.set_trace() if args.evaluate: img, msk_gt = data_batch msk_gt_buffer.append(msk_gt.cpu().detach().numpy()) else: img = data_batch img = img.to(device) with torch.no_grad(): msk_pred = model(img) # shape (N, 3, H, W) msk_pred = sigmoid_act(msk_pred) msk_pred_buffer.append(msk_pred.cpu().detach().numpy()) msk_pred_buffer = np.vstack(msk_pred_buffer) # shape (N, 3, H, W) if args.evaluate: msk_gt_buffer = np.vstack(msk_gt_buffer) results = [] for vol_ind, vol_start_ind in enumerate(volume_start_index): if vol_ind == len(volume_start_index) - 1: volume_msk = msk_pred_buffer[vol_start_ind:] # shape (N, 3, H, W) if args.evaluate: volume_msk_gt = msk_gt_buffer[vol_start_ind:] else: vol_end_ind = volume_start_index[vol_ind+1] volume_msk = msk_pred_buffer[vol_start_ind:vol_end_ind] # shape (N, 3, H, W) if args.evaluate: volume_msk_gt = msk_gt_buffer[vol_start_ind:vol_end_ind] if args.evaluate: # liver liver_scores = get_scores(volume_msk[:, 1] >= 0.5, volume_msk_gt >= 1, spacing[vol_ind]) # tumor lesion_scores = get_scores(volume_msk[:, 2] >= 0.5, volume_msk_gt == 2, spacing[vol_ind]) print("Liver dice", liver_scores['dice'], "Lesion dice", lesion_scores['dice']) results.append([vol_ind, liver_scores, lesion_scores]) # =========================== else: # import ipdb; ipdb.set_trace() if args.save2dir: # reverse the order, because we prioritize tumor, liver then background. msk_pred = (volume_msk >= 0.5)[:, ::-1, ...] # shape (N, 3, H, W) msk_pred = np.argmax(msk_pred, axis=1) # shape (N, H, W) = (z, x, y) msk_pred = np.transpose(msk_pred, axes=(1, 2, 0)) # shape (x, y, z) # remember to correct 'direction' and np.transpose before the submission !!! if direction[vol_ind][0] == -1: # x-axis msk_pred = msk_pred[::-1, ...] if direction[vol_ind][1] == -1: # y-axis msk_pred = msk_pred[:, ::-1, :] if direction[vol_ind][2] == -1: # z-axis msk_pred = msk_pred[..., ::-1] # save medical image header as well # see: http://loli.github.io/medpy/generated/medpy.io.header.Header.html file_header = med_header(spacing=tuple(spacing[vol_ind]), offset=tuple(offset[vol_ind]), direction=np.diag(direction[vol_ind])) # submission guide: # see: https://github.com/PatrickChrist/LITS-CHALLENGE/blob/master/submission-guide.md # test-segmentation-X.nii filepath = os.path.join(args.save2dir, f"test-segmentation-{vol_ind}.nii") med_save(msk_pred, filepath, hdr=file_header) if args.save2dir: # outpath = os.path.join(args.save2dir, "results.csv") outpath = os.path.join(args.save2dir, "results.pkl") with open(outpath, "wb") as file: final_result = {} final_result['liver'] = defaultdict(list) final_result['tumor'] = defaultdict(list) for vol_ind, liver_scores, lesion_scores in results: # [OTC] assuming vol_ind is continuous for key in liver_scores: final_result['liver'][key].append(liver_scores[key]) for key in lesion_scores: final_result['tumor'][key].append(lesion_scores[key]) pickle.dump(final_result, file, protocol=3) # ======== code from official metric ======== # create line for csv file # outstr = str(vol_ind) + ',' # for l in [liver_scores, lesion_scores]: # for k, v in l.items(): # outstr += str(v) + ',' # outstr += '\n' # # create header for csv file if necessary # if not os.path.isfile(outpath): # headerstr = 'Volume,' # for k, v in liver_scores.items(): # headerstr += 'Liver_' + k + ',' # for k, v in liver_scores.items(): # headerstr += 'Lesion_' + k + ',' # headerstr += '\n' # outstr = headerstr + outstr # # write to file # f = open(outpath, 'a+') # f.write(outstr) # f.close() # =========================== printGreen(f"Total elapsed time: {time.time()-st}") return results
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(args): def log_string(str): # logger.info(str) print(str) '''HYPER PARAMETER''' os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu '''CREATE DIR''' timestr = str(datetime.datetime.now().strftime('%Y-%m-%d_%H-%M')) experiment_dir = Path('./log/') experiment_dir.mkdir(exist_ok=True) experiment_dir = experiment_dir.joinpath('part_seg') experiment_dir.mkdir(exist_ok=True) if args.log_dir is None: experiment_dir = experiment_dir.joinpath(timestr) else: experiment_dir = experiment_dir.joinpath(args.log_dir) experiment_dir.mkdir(exist_ok=True) checkpoints_dir = experiment_dir.joinpath('checkpoints/') checkpoints_dir.mkdir(exist_ok=True) log_dir = experiment_dir.joinpath('logs/') log_dir.mkdir(exist_ok=True) '''LOG''' args = parse_args() logger = logging.getLogger("Model") logger.setLevel(logging.INFO) formatter = logging.Formatter( '%(asctime)s - %(name)s - %(levelname)s - %(message)s') file_handler = logging.FileHandler('%s/%s.txt' % (log_dir, args.model)) file_handler.setLevel(logging.INFO) file_handler.setFormatter(formatter) logger.addHandler(file_handler) log_string('PARAMETER ...') log_string(args) root = '/media/feihu/Storage/kitti_point_cloud/semantic_kitti/' # file_list = '/media/feihu/Storage/kitti_point_cloud/semantic_kitti/train2.list' val_list = '/media/feihu/Storage/kitti_point_cloud/semantic_kitti/val2.list' # TRAIN_DATASET = KittiDataset(root = root, file_list=file_list, npoints=args.npoint, training=True, augment=True) # trainDataLoader = torch.utils.data.DataLoader(TRAIN_DATASET, batch_size=args.batch_size, shuffle=True, drop_last=True, num_workers=2) TEST_DATASET = KittiDataset(root=root, file_list=val_list, npoints=args.npoint, training=False, augment=False) testDataLoader = torch.utils.data.DataLoader(TEST_DATASET, batch_size=args.batch_size, shuffle=False, drop_last=True, num_workers=2) # log_string("The number of training data is: %d" % len(TRAIN_DATASET)) log_string("The number of test data is: %d" % len(TEST_DATASET)) # num_classes = 16 num_devices = args.num_gpus #torch.cuda.device_count() # assert num_devices > 1, "Cannot detect more than 1 GPU." # print(num_devices) devices = list(range(num_devices)) target_device = devices[0] # MODEL = importlib.import_module(args.model) net = UNet(4, 20, nPlanes) # net = MODEL.get_model(num_classes, normal_channel=args.normal) net = net.to(target_device) try: checkpoint = torch.load( str(experiment_dir) + '/checkpoints/best_model.pth') start_epoch = checkpoint['epoch'] net.load_state_dict(checkpoint['model_state_dict']) log_string('Use pretrain model') except: log_string('No existing model, starting training from scratch...') quit() if 1: with torch.no_grad(): net.eval() evaluator = iouEval(num_classes, ignore) evaluator.reset() # for iteration, (points, target, ins, mask) in tqdm(enumerate(testDataLoader), total=len(testDataLoader), smoothing=0.9): for iteration, (points, target, ins, mask) in enumerate(testDataLoader): evaone = iouEval(num_classes, ignore) evaone.reset() cur_batch_size, NUM_POINT, _ = points.size() if iteration > 128: break inputs, targets, masks = [], [], [] coords = [] for i in range(num_devices): start = int(i * (cur_batch_size / num_devices)) end = int((i + 1) * (cur_batch_size / num_devices)) with torch.cuda.device(devices[i]): pc = points[start:end, :, :].to(devices[i]) #feas = points[start:end,:,3:].to(devices[i]) targeti = target[start:end, :].to(devices[i]) maski = mask[start:end, :].to(devices[i]) locs, feas, label, maski, offsets = input_layer( pc, targeti, maski, scale.to(devices[i]), spatialSize.to(devices[i]), True) # print(locs.size(), feas.size(), label.size(), maski.size(), offsets.size()) org_coords = locs[1] label = Variable(label, requires_grad=False) inputi = ME.SparseTensor(feas.cpu(), locs[0].cpu()) inputs.append([inputi.to(devices[i]), org_coords]) targets.append(label) masks.append(maski) replicas = parallel.replicate(net, devices) outputs = parallel.parallel_apply(replicas, inputs, devices=devices) seg_pred = outputs[0].cpu() mask = masks[0].cpu() target = targets[0].cpu() loc = locs[0].cpu() for i in range(1, num_devices): seg_pred = torch.cat((seg_pred, outputs[i].cpu()), 0) mask = torch.cat((mask, masks[i].cpu()), 0) target = torch.cat((target, targets[i].cpu()), 0) seg_pred = seg_pred[target > 0, :] target = target[target > 0] _, seg_pred = seg_pred.data.max(1) #[1] target = target.data.numpy() evaluator.addBatch(seg_pred, target) evaone.addBatch(seg_pred, target) cur_accuracy = evaone.getacc() cur_jaccard, class_jaccard = evaone.getIoU() print('%.4f %.4f' % (cur_accuracy, cur_jaccard)) m_accuracy = evaluator.getacc() m_jaccard, class_jaccard = evaluator.getIoU() log_string('Validation set:\n' 'Acc avg {m_accuracy:.3f}\n' 'IoU avg {m_jaccard:.3f}'.format(m_accuracy=m_accuracy, m_jaccard=m_jaccard)) # print also classwise for i, jacc in enumerate(class_jaccard): if i not in ignore: log_string( 'IoU class {i:} [{class_str:}] = {jacc:.3f}'.format( i=i, class_str=class_strings[class_inv_remap[i]], jacc=jacc))
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
# 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() res_pseudo = pseudo.downsample(edges_a) pred_seg_a = pseudo.upsample(*res_pseudo) # pred_seg_a = pseudo(edges_a) loss_seg_a = criterion(pred_seg_a, target_a) loss_seg_a.backward() optimiser_ps.step() else: pseudo.eval() image_a = data['A'][0].cuda() target_a = data['A'][1].cuda() image_b = data['B'][0].cuda() edges_b = data['B'][2].cuda() pseudo_b = pseudo.downsample(edges_b) pred_pseudo_b = pseudo.upsample(*pseudo_b) # pred_pseudo_b = pseudo(edges_b) target_b = torch.round(pred_pseudo_b).detach().cuda() net.set_domain(DOMAIN_A) res_a = net.downsample(image_a) pred_seg_a = net.upsample(*res_a) net.set_domain(DOMAIN_B)
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 test_UNet(): cfg = UnetConfig() transform = transforms.Compose([ GrayscaleNormalization(mean=0.5, std=0.5), ToTensor(), ]) RESULTS_DIR = os.path.join(ROOT_DIR, 'test_results/unet') if not os.path.exists(RESULTS_DIR): os.makedirs(RESULTS_DIR) label_save_path = os.path.join(RESULTS_DIR, 'label') output_save_path = os.path.join(RESULTS_DIR, 'output') if not os.path.exists(label_save_path): os.makedirs(label_save_path, exist_ok=True) if not os.path.exists(output_save_path): os.makedirs(output_save_path, exist_ok=True) test_dataset = Dataset(imgs_dir=TEST_IMGS_DIR, labels_dir=TEST_LABELS_DIR, transform=transform) test_loader = DataLoader(test_dataset, batch_size=cfg.BATCH_SIZE, shuffle=False, num_workers=0) test_data_num = len(test_dataset) test_batch_num = int(np.ceil(test_data_num / cfg.BATCH_SIZE)) # Network net = UNet().to(device) # Loss Function loss_fn = nn.BCEWithLogitsLoss().to(device) # Optimizer optim = torch.optim.Adam(params=net.parameters(), lr=cfg.LEARNING_RATE) start_epoch = 0 # Load Checkpoint File if os.listdir(CKPT_DIR): net, optim, _ = load_net(ckpt_dir=os.path.join(CKPT_DIR, 'unet'), net=net, optim=optim) # Evaluation with torch.no_grad(): net.eval() loss_arr = list() for batch_idx, data in enumerate(test_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) loss_arr.append(loss.item()) print_form = '[Test] | Batch: {:0>4d} / {:0>4d} | Loss: {:.4f}' print(print_form.format(batch_idx, test_batch_num, loss_arr[-1])) label = to_numpy(label) output = to_numpy(classify_class(output)) for j in range(label.shape[0]): crt_id = int(test_batch_num * (batch_idx - 1) + j) plt.imsave(os.path.join(label_save_path, f'{crt_id:04}.png'), label[j].squeeze(), cmap='gray') plt.imsave(os.path.join(output_save_path, f'{crt_id:04}.png'), output[j].squeeze(), cmap='gray') unet_acc(output_save_path, label_save_path)
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()
def validate(): device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # Setup Dataloader # data_loader = get_loader(cfg["data"]["dataset"]) model_id = "20200404_00_UNet" checkpoint_path = "../checkpoints/{}/checkpoint.pth.tar".format(model_id) base_path = os.path.dirname(os.path.dirname(os.path.abspath(__file__))) dataroot = os.path.join(os.path.dirname(base_path), "datasets") if not os.path.exists(dataroot): os.mkdir(dataroot) n_classes = 21 model = UNet(n_channels=3, n_classes=21).to(device) # .cuda() datasets = torchvision.datasets.VOCSegmentation( dataroot, year='2012', image_set='train', download=False, transform=original_transform, target_transform=teacher_transform) # valloader = data.DataLoader(loader, batch_size=cfg["training"]["batch_size"], num_workers=8) valloader = torch.utils.data.DataLoader(datasets, batch_size=1, shuffle=False) running_metrics = runningScore(n_classes) # Setup Model model.load_state_dict(torch.load(checkpoint_path)) model.eval() # model.to(device) flag = False for i, (images, labels) in enumerate(valloader): start_time = timeit.default_timer() images = images.to(device) if flag: outputs = model(images) # Flip images in numpy (not support in tensor) outputs = outputs.data.cpu().numpy() flipped_images = np.copy(images.data.cpu().numpy()[:, :, :, ::-1]) flipped_images = torch.from_numpy(flipped_images).float().to( device) outputs_flipped = model(flipped_images) outputs_flipped = outputs_flipped.data.cpu().numpy() outputs = (outputs + outputs_flipped[:, :, :, ::-1]) / 2.0 pred = np.argmax(outputs, axis=1) else: outputs = model(images) pred = outputs.data.max(1)[1].cpu().numpy() gt = labels.numpy() if True: elapsed_time = timeit.default_timer() - start_time print("Inference time \ (iter {0:5d}): {1:3.5f} fps".format( i + 1, pred.shape[0] / elapsed_time)) running_metrics.update(gt, pred) score, class_iou = running_metrics.get_scores() for k, v in score.items(): print(k, v) for i in range(n_classes): print(i, class_iou[i])
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 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)
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)