import os os.environ["CUDA_VISIBLE_DEVICES"] = "0" import torch import config import monai import warnings import numpy as np import pandas as pd from model.CMTLNet import CMTLNet from torch.nn import CrossEntropyLoss from torch.utils.data import DataLoader from torch.optim.lr_scheduler import ReduceLROnPlateau from dataLoder.loader_S2 import TrainDataLoader_balance as TrainDataLoader from dataLoder.loader_S2 import TestDataLoader from utiles import save_model, get_learning_rate, calculate_classification_metrics, move_tensor_to_device, AverageMeter from survivalstandalone.losses import cox_ph_loss def validate(data_loader, model, mode_flag, ssl_index, class_name): """ Validates the model on a given dataset and computes loss and classification metrics. Parameters: data_loader (DataLoader): DataLoader for the dataset to validate. model (torch.nn.Module): The model to validate. mode_flag (str): Indicates whether the dataset is 'Train' or 'Test'. ssl_index (int): SSL index used for model prediction. class_name (str): Name of the class/category being validated. Returns: tuple: Contains average loss, accuracy, F1 score, precision, and recall. """ model.eval() # Set model to evaluation mode cross_entropy_loss = CrossEntropyLoss(reduction='mean') # Define cross-entropy loss all_targets, all_predictions = [], [] # Lists to store all true labels and predictions loss_meter = AverageMeter() # Utility to track average loss for batch_data in data_loader: images, labels = batch_data images = move_tensor_to_device(images, use_gpu=config.use_cuda) # Move images to GPU if available labels = move_tensor_to_device(labels, use_gpu=config.use_cuda) # Move labels to GPU if available with torch.no_grad(): # Disable gradient computation outputs = model.predict(images, ssl_index) # Get model predictions loss = cross_entropy_loss(outputs, labels) # Compute loss loss_meter.update(loss.item(), 1) # Update loss meter probabilities = torch.nn.functional.softmax(outputs, dim=1) # Compute probabilities predicted_classes = torch.argmax(probabilities, dim=1).cpu().numpy() # Get predicted classes true_labels = labels.cpu().detach().numpy() # Get true labels all_predictions.extend(predicted_classes) # Collect predictions all_targets.extend(true_labels) # Collect true labels average_loss = loss_meter.average # Calculate average loss all_predictions = np.array(all_predictions, dtype=np.int32) all_targets = np.array(all_targets, dtype=np.int32) # Compute classification metrics: Accuracy, F1, Precision, Recall, and Confusion Matrix accuracy, f1, precision, recall, confusion_matrix = calculate_classification_metrics( predictions=all_predictions, targets=all_targets, average_method="macro" ) print(confusion_matrix) # Print confusion matrix # Print metrics based on whether it's training or testing if mode_flag == 'Test': print(f'Test set {class_name} | Accuracy {accuracy:.4f} | F1 {f1:.4f} | Precision {precision:.4f} | Recall {recall:.4f} |') else: print(f'Train set {class_name} | Accuracy {accuracy:.4f} | F1 {f1:.4f} | Precision {precision:.4f} | Recall {recall:.4f} |') model.train() # Set model back to training mode return average_loss, accuracy, f1, precision, recall def valid_metrics_by_class(model, train_loader, test_loader, ssl_index, class_name): """ Computes validation metrics for a specific class on both training and testing datasets. Parameters: model (torch.nn.Module): The model to validate. train_loader (DataLoader): DataLoader for the training dataset. test_loader (DataLoader): DataLoader for the testing dataset. ssl_index (int): SSL index used for model prediction. class_name (str): Name of the class/category being validated. Returns: tuple: Contains training metrics and testing metrics. """ # Validate on training data train_loss, train_accuracy, train_f1, train_precision, train_recall = validate( data_loader=train_loader, model=model, mode_flag='Train', ssl_index=ssl_index, class_name=class_name ) # Validate on testing data test_loss, test_accuracy, test_f1, test_precision, test_recall = validate( data_loader=test_loader, model=model, mode_flag='Test', ssl_index=ssl_index, class_name=class_name ) # Collect training and testing metrics metrics_train = [train_loss, train_accuracy, train_f1, train_precision, train_recall] metrics_test = [test_loss, test_accuracy, test_f1, test_precision, test_recall] return metrics_train, metrics_test def get_pandas_series(epoch, train_metrics, test_metrics): """ Creates a Pandas Series for logging metrics. Parameters: epoch (int): Current epoch number. train_metrics (list): List of training metrics. test_metrics (list): List of testing metrics. Returns: pd.Series: Series containing epoch and corresponding metrics. """ series = pd.Series([ epoch, train_metrics[0], # Training loss train_metrics[1], # Training accuracy train_metrics[2], # Training F1 score test_metrics[0], # Testing loss test_metrics[1], # Testing accuracy test_metrics[2], # Testing F1 score ], index=[ 'epoch', 'Train_loss', 'Train_Accuracy', 'Train_F1', 'Val_loss', 'Val_Accuracy', 'Val_F1' ]) return series def get_class_loaders(TrainDataset, TestDataset, transforms, fold, loader_id, pkl_path): """ Creates DataLoaders for training and validation datasets for a specific class. Parameters: TrainDataset (Dataset): Training dataset class. TestDataset (Dataset): Testing dataset class. transforms (Compose): Data transformation pipeline. fold (int): Current fold number for cross-validation. loader_id (str): Identifier for the class. pkl_path (str): Path to the pickle file containing dataset information. Returns: tuple: Contains training DataLoader, validation training DataLoader, and validation testing DataLoader. """ # Create training dataset train_dataset = TrainDataset( mode='train', fold=fold, transform=transforms, pickle_path=pkl_path ) # Create training DataLoader train_loader = DataLoader( train_dataset, batch_size=config.batch_size, shuffle=True, drop_last=True, num_workers=config.n_threads, pin_memory=config.use_cuda ) print(f"Number of train_loader {loader_id} examples {len(train_dataset)}") # Create validation training dataset val_train_dataset = TestDataset( mode='train', fold=fold, transform=transforms, pickle_path=pkl_path ) # Create validation training DataLoader val_train_loader = DataLoader( val_train_dataset, batch_size=1, shuffle=False, drop_last=False, num_workers=config.n_threads, pin_memory=config.use_cuda ) print(f"Number of val_train_loader {loader_id} examples {len(val_train_dataset)}") # Create validation testing dataset val_test_dataset = TestDataset( mode='test', fold=fold, transform=transforms, pickle_path=pkl_path ) # Create validation testing DataLoader val_loader = DataLoader( val_test_dataset, batch_size=1, shuffle=False, drop_last=False, num_workers=config.n_threads, pin_memory=config.use_cuda ) print(f"Number of val_loader {loader_id} examples {len(val_test_dataset)}") return train_loader, val_train_loader, val_loader def reshape_input_data(data): images, labels, times, events = data images = images.view(images.size(0) * images.size(1), images.size(2), images.size(3), images.size(4), images.size(5)) labels = labels.view(-1) times = times.view(-1) events = events.view(-1) images = move_tensor_to_device(images, use_gpu=config.use_cuda) labels = move_tensor_to_device(labels, use_gpu=config.use_cuda) times = move_tensor_to_device(times, use_gpu=config.use_cuda) events = move_tensor_to_device(events, use_gpu=config.use_cuda) return images, labels, times, events def train(fold, pkl_paths): """ Main training function responsible for the entire training process, including data loading, model training, validation, logging, and model saving. Parameters: fold (int): Current fold number for cross-validation. pkl_paths (list): List of paths to pickle files for each class. """ # Define data transformations val_transforms = monai.transforms.Compose([ monai.transforms.ToTensor(), ]) # Get DataLoaders for each class train_loader_1p19q, val_train_loader_1p19q, val_loader_1p19q = get_class_loaders( TrainDataset=TrainDataLoader, TestDataset=TestDataLoader, transforms=val_transforms, fold=fold, loader_id='1p19q', pkl_path=pkl_paths[0] ) train_loader_IDH, val_train_loader_IDH, val_loader_IDH = get_class_loaders( TrainDataset=TrainDataLoader, TestDataset=TestDataLoader, transforms=val_transforms, fold=fold, loader_id='IDH', pkl_path=pkl_paths[1] ) train_loader_LHG, val_train_loader_LHG, val_loader_LHG = get_class_loaders( TrainDataset=TrainDataLoader, TestDataset=TestDataLoader, transforms=val_transforms, fold=fold, loader_id='LHG', pkl_path=pkl_paths[2] ) # Move model to GPU if available if config.use_cuda: model.cuda() cross_entropy_loss = CrossEntropyLoss(reduction='mean') # Define cross-entropy loss # Only optimize parameters that require gradients optim_parameters = filter(lambda p: p.requires_grad, model.parameters()) optimizer = torch.optim.Adam( optim_parameters, lr=config.lr_self, weight_decay=config.weight_decay ) # Using Adam optimizer # Learning rate scheduler scheduler = ReduceLROnPlateau( optimizer=optimizer, factor=config.lr_reduce_factor, patience=config.lr_reduce_patience, mode='max', min_lr=config.min_lr ) global_step = 0 # Initialize global step best_score = -1 # Initialize best score # Initialize DataFrames for logging log_1p19q = pd.DataFrame(columns=[ 'epoch', 'Train_loss', 'Train_Accuracy', 'Train_F1', 'Val_loss', 'Val_Accuracy', 'Val_F1' ]) log_IDH = pd.DataFrame(columns=[ 'epoch', 'Train_loss', 'Train_Accuracy', 'Train_F1', 'Val_loss', 'Val_Accuracy', 'Val_F1' ]) log_LHG = pd.DataFrame(columns=[ 'epoch', 'Train_loss', 'Train_Accuracy', 'Train_F1', 'Val_loss', 'Val_Accuracy', 'Val_F1' ]) log_mean = pd.DataFrame(columns=[ 'epoch', 'Train_loss', 'Train_Accuracy', 'Train_F1', 'Val_loss', 'Val_Accuracy', 'Val_F1' ]) model.train() # Set model to training mode # Initialize iterator indices iter_index_IDH = 0 iter_index_LHG = 0 iter_IDH = iter(train_loader_IDH) iter_LHG = iter(train_loader_LHG) for epoch in range(config.epochs_s2): if epoch == config.save_epoch: best_score = -1 # Reset best score at a specific epoch iter_1p19q = iter(train_loader_1p19q) # Get iterator for 1p19q class for batch_idx in range(len(train_loader_1p19q)): iter_index_IDH += 1 iter_index_LHG += 1 # Reset iterators if they exceed dataset length if iter_index_IDH >= len(train_loader_IDH): iter_IDH = iter(train_loader_IDH) iter_index_IDH = 0 if iter_index_LHG >= len(train_loader_LHG): iter_LHG = iter(train_loader_LHG) iter_index_LHG = 0 # Get data for each class data_1p19q = next(iter_1p19q) data_IDH = next(iter_IDH) data_LHG = next(iter_LHG) # Reshape input data # Reshape input data imgs_1p19q, label_1p19q, time_1p19q, event_1p19q = reshape_input_data(data_1p19q) imgs_IDH, label_IDH, time_IDH, event_IDH = reshape_input_data(data_IDH) imgs_LHG, label_LHG, time_LHG, event_LHG = reshape_input_data(data_LHG) # Combine inputs and targets label_list = [label_1p19q, label_IDH, label_LHG] combined_input = torch.cat((imgs_1p19q, imgs_IDH, imgs_LHG), dim=0) combined_times = torch.cat((time_1p19q, time_IDH, time_LHG), dim=0) combined_events = torch.cat((event_1p19q, event_IDH, event_LHG), dim=0) condition_1p19q = torch.tensor(0, dtype=torch.long, device=label_1p19q.device).repeat(1, label_1p19q.size(0))[0] condition_IDH = torch.tensor(1, dtype=torch.long, device=label_IDH.device).repeat(1, label_IDH.size(0))[0] condition_LHG = torch.tensor(2, dtype=torch.long, device=label_LHG.device).repeat(1, label_LHG.size(0))[0] condition = torch.cat([condition_1p19q, condition_IDH, condition_LHG], dim=0) # Forward pass (now receives risk_scores) (out_cls_list_UFE, out_cls_labels_UFE, out_cls_list_UCFC, out_cls_labels_UCFC, risk_scores) = model(combined_input, label_list, condition) # --- CALCULATE COX LOSS --- # Mask out patients who don't have survival data (time == 0) valid_mask = combined_times > 0 if valid_mask.sum() >= 2: # Cox loss needs at least a pair to rank loss_surv = cox_ph_loss(risk_scores[valid_mask], combined_times[valid_mask], combined_events[valid_mask]) else: loss_surv = torch.tensor(0.0, device=risk_scores.device, requires_grad=True) # Classification losses loss_cls_U1 = cross_entropy_loss(out_cls_list_UFE[0], out_cls_labels_UFE[0]) loss_cls_U2 = cross_entropy_loss(out_cls_list_UFE[1], out_cls_labels_UFE[1]) loss_cls_U3 = cross_entropy_loss(out_cls_list_UFE[2], out_cls_labels_UFE[2]) loss_cls_1 = cross_entropy_loss(out_cls_list_UCFC[0], out_cls_labels_UCFC[0]) loss_cls_2 = cross_entropy_loss(out_cls_list_UCFC[1], out_cls_labels_UCFC[1]) loss_cls_3 = cross_entropy_loss(out_cls_list_UCFC[2], out_cls_labels_UCFC[2]) # Total loss (weighing survival by 0.5 so it doesn't overpower classification) total_loss = loss_cls_U1 + loss_cls_U2 + loss_cls_U3 + loss_cls_1 + loss_cls_2 + loss_cls_3 + (0.5 * loss_surv) # Backward pass optimizer.zero_grad() total_loss.backward() optimizer.step() # Get current learning rate current_lr = get_learning_rate(optimizer) print(f'Global_Step {global_step} | Train Epoch: {epoch} | lr {current_lr:.2e} |') # Define save path save_path = os.path.join(config.data_dir, 'results', str(save_name), f'Fold_{fold}') if not os.path.exists(save_path): os.makedirs(save_path) # Compute metrics for each class metrics_train_1p19q, metrics_test_1p19q = valid_metrics_by_class( model=model, train_loader=val_train_loader_1p19q, test_loader=val_loader_1p19q, ssl_index=0, class_name='1p19q' ) metrics_train_IDH, metrics_test_IDH = valid_metrics_by_class( model=model, train_loader=val_train_loader_IDH, test_loader=val_loader_IDH, ssl_index=1, class_name='IDH' ) metrics_train_LHG, metrics_test_LHG = valid_metrics_by_class( model=model, train_loader=val_train_loader_LHG, test_loader=val_loader_LHG, ssl_index=2, class_name='LHG' ) # Calculate mean metrics across classes metrics_train_mean = (np.array(metrics_train_1p19q) + np.array(metrics_train_IDH) + np.array(metrics_train_LHG)) / 3.0 metrics_test_mean = (np.array(metrics_test_1p19q) + np.array(metrics_test_IDH) + np.array(metrics_test_LHG)) / 3.0 # Create log entries log_entry_1p19q = get_pandas_series(epoch, metrics_train_1p19q, metrics_test_1p19q) log_entry_IDH = get_pandas_series(epoch, metrics_train_IDH, metrics_test_IDH) log_entry_LHG = get_pandas_series(epoch, metrics_train_LHG, metrics_test_LHG) log_entry_mean = get_pandas_series(epoch, metrics_train_mean, metrics_test_mean) # Append logs to DataFrames and save to CSV log_1p19q = log_1p19q.append(log_entry_1p19q, ignore_index=True) log_1p19q.to_csv(os.path.join(save_path, 'log_1p19q.csv'), index=False) log_IDH = log_IDH.append(log_entry_IDH, ignore_index=True) log_IDH.to_csv(os.path.join(save_path, 'log_IDH.csv'), index=False) log_LHG = log_LHG.append(log_entry_LHG, ignore_index=True) log_LHG.to_csv(os.path.join(save_path, 'log_LHG.csv'), index=False) log_mean = log_mean.append(log_entry_mean, ignore_index=True) log_mean.to_csv(os.path.join(save_path, 'log_mean.csv'), index=False) # Use the mean F1 score from the test set as the evaluation metric current_f1 = metrics_test_mean[2] # Save the best model based on F1 score if epoch >= config.save_epoch: if current_f1 > best_score: save_config = { 'name': save_name, 'save_dir': save_path, 'global_step': global_step, 'Test_F1': current_f1 } save_model(model=model, config=save_config) best_score = current_f1 # Update the best score if current F1 is better if current_f1 > best_score: best_score = current_f1 # Step the scheduler based on the best score scheduler.step(best_score) if __name__ == '__main__': # Ignore all warnings warnings.filterwarnings("ignore") # Set random seeds for reproducibility seed = config.random_seed np.random.seed(seed) # Python's random seed os.environ['PYTHONHASHSEED'] = str(seed) # Disable Python hash randomization torch.manual_seed(seed) # Torch CPU random seed torch.cuda.manual_seed(seed) # Torch GPU random seed torch.cuda.manual_seed_all(seed) # Torch multi-GPU random seed torch.backends.cudnn.benchmark = False # Disable cuDNN benchmark torch.backends.cudnn.deterministic = True # Enable deterministic algorithms # Define paths to pickle files for each class pkl_paths = [ './1p19q.pkl', './IDH.pkl', './LHG.pkl' ] Base_CEF_model_paths = [ './Fold_0/CFE.pkl', './Fold_1/CFE.pkl', './Fold_2/CFE.pkl', './Fold_3/CFE.pkl', './Fold_4/CFE.pkl' ] # Perform 5-fold cross-validation training for fold in range(5): model = CMTLNet(in_channels=3, num_classes=2, Arcm=0.5, multi=3) # Initialize the model model_basic_CFE = torch.load(Base_CEF_model_paths[fold], weights_only=False) model.ShareTask.load_state_dict(model_basic_CFE.state_dict()) save_name = model.name # Get the model's name print(f"{save_name}_Fold_{fold}") train(fold, pkl_paths) # Start training