def cross_validate(args: Namespace, logger: Logger = None) -> Tuple[float, float]: """k-fold cross validation""" info = logger.info if logger is not None else print # Initialize relevant variables init_seed = args.seed save_dir = args.save_dir task_names = get_task_names(args.data_path) desired_labels = get_desired_labels(args, task_names) # Run training on different random seeds for each fold all_scores = [] for fold_num in range(args.num_folds): info(f'Fold {fold_num}') args.seed = init_seed + fold_num args.save_dir = os.path.join(save_dir, f'fold_{fold_num}') os.makedirs(args.save_dir, exist_ok=True) model_scores = run_training(args, logger) all_scores.append(model_scores) all_scores = np.array(all_scores) # Report results info(f'{args.num_folds}-fold cross validation') # Report scores for each fold for fold_num, scores in enumerate(all_scores): info( f'Seed {init_seed + fold_num} ==> test {args.metric} = {np.nanmean(scores):.6f}' ) if args.show_individual_scores: for task_name, score in zip(task_names, scores): if task_name in desired_labels: info( f'Seed {init_seed + fold_num} ==> test {task_name} {args.metric} = {score:.6f}' ) # Report scores across models avg_scores = np.nanmean( all_scores, axis=1) # average score for each model across tasks mean_score, std_score = np.nanmean(avg_scores), np.nanstd(avg_scores) info(f'Overall test {args.metric} = {mean_score:.6f} +/- {std_score:.6f}') if args.show_individual_scores: for task_num, task_name in enumerate(task_names): if task_name in desired_labels: info( f'Overall test {task_name} {args.metric} = ' f'{np.nanmean(all_scores[:, task_num]):.6f} +/- {np.nanstd(all_scores[:, task_num]):.6f}' ) if args.num_chunks > 1: shutil.rmtree(args.chunk_temp_dir) return mean_score, std_score
def run_training(args: Namespace, logger: Logger = None) -> List[float]: """ Trains a model and returns test scores on the model checkpoint with the highest validation score. :param args: Arguments. :param logger: Logger. :return: A list of ensemble scores for each task. """ if logger is not None: debug, info = logger.debug, logger.info else: debug = info = print # Set GPU if args.gpu is not None: torch.cuda.set_device(args.gpu) # Print args debug(pformat(vars(args))) # Get data debug('Loading data') args.task_names = get_task_names(args.data_path) desired_labels = get_desired_labels(args, args.task_names) data = get_data(path=args.data_path, args=args, logger=logger) args.num_tasks = data.num_tasks() args.features_size = data.features_size() args.real_num_tasks = args.num_tasks - args.features_size if args.predict_features else args.num_tasks debug(f'Number of tasks = {args.num_tasks}') if args.dataset_type == 'bert_pretraining': data.bert_init(args, logger) # Split data if args.dataset_type == 'regression_with_binning': # Note: for now, binning based on whole dataset, not just training set data, bin_predictions, regression_data = data args.bin_predictions = bin_predictions debug(f'Splitting data with seed {args.seed}') train_data, _, _ = split_data(data=data, split_type=args.split_type, sizes=args.split_sizes, seed=args.seed, args=args, logger=logger) _, val_data, test_data = split_data(regression_data, split_type=args.split_type, sizes=args.split_sizes, seed=args.seed, args=args, logger=logger) else: debug(f'Splitting data with seed {args.seed}') if args.separate_test_set: test_data = get_data(path=args.separate_test_set, args=args, features_path=args.separate_test_set_features, logger=logger) if args.separate_val_set: val_data = get_data( path=args.separate_val_set, args=args, features_path=args.separate_val_set_features, logger=logger) train_data = data # nothing to split; we already got our test and val sets else: train_data, val_data, _ = split_data( data=data, split_type=args.split_type, sizes=(0.8, 0.2, 0.0), seed=args.seed, args=args, logger=logger) else: train_data, val_data, test_data = split_data( data=data, split_type=args.split_type, sizes=args.split_sizes, seed=args.seed, args=args, logger=logger) # Optionally replace test data with train or val data if args.test_split == 'train': test_data = train_data elif args.test_split == 'val': test_data = val_data if args.dataset_type == 'classification': class_sizes = get_class_sizes(data) debug('Class sizes') for i, task_class_sizes in enumerate(class_sizes): debug( f'{args.task_names[i]} ' f'{", ".join(f"{cls}: {size * 100:.2f}%" for cls, size in enumerate(task_class_sizes))}' ) if args.class_balance: train_class_sizes = get_class_sizes(train_data) class_batch_counts = torch.Tensor( train_class_sizes) * args.batch_size args.class_weights = 1 / torch.Tensor(class_batch_counts) if args.save_smiles_splits: with open(args.data_path, 'r') as f: reader = csv.reader(f) header = next(reader) lines_by_smiles = {} indices_by_smiles = {} for i, line in enumerate(reader): smiles = line[0] lines_by_smiles[smiles] = line indices_by_smiles[smiles] = i all_split_indices = [] for dataset, name in [(train_data, 'train'), (val_data, 'val'), (test_data, 'test')]: with open(os.path.join(args.save_dir, name + '_smiles.csv'), 'w') as f: writer = csv.writer(f) writer.writerow(['smiles']) for smiles in dataset.smiles(): writer.writerow([smiles]) with open(os.path.join(args.save_dir, name + '_full.csv'), 'w') as f: writer = csv.writer(f) writer.writerow(header) for smiles in dataset.smiles(): writer.writerow(lines_by_smiles[smiles]) split_indices = [] for smiles in dataset.smiles(): split_indices.append(indices_by_smiles[smiles]) split_indices = sorted(split_indices) all_split_indices.append(split_indices) with open(os.path.join(args.save_dir, 'split_indices.pckl'), 'wb') as f: pickle.dump(all_split_indices, f) return [1 for _ in range(args.num_tasks) ] # short circuit out when just generating splits if args.features_scaling: features_scaler = train_data.normalize_features( replace_nan_token=None if args.predict_features else 0) val_data.normalize_features(features_scaler) test_data.normalize_features(features_scaler) else: features_scaler = None args.train_data_size = len( train_data ) if args.prespecified_chunk_dir is None else args.prespecified_chunks_max_examples_per_epoch if args.adversarial or args.moe: val_smiles, test_smiles = val_data.smiles(), test_data.smiles() debug( f'Total size = {len(data):,} | ' f'train size = {len(train_data):,} | val size = {len(val_data):,} | test size = {len(test_data):,}' ) # Optionally truncate outlier values if args.truncate_outliers: print('Truncating outliers in train set') train_data = truncate_outliers(train_data) # Initialize scaler and scale training targets by subtracting mean and dividing standard deviation (regression only) if args.dataset_type == 'regression' and args.target_scaling: debug('Fitting scaler') train_smiles, train_targets = train_data.smiles(), train_data.targets() scaler = StandardScaler().fit(train_targets) scaled_targets = scaler.transform(train_targets).tolist() train_data.set_targets(scaled_targets) else: scaler = None if args.moe: train_data = cluster_split(train_data, args.num_sources, args.cluster_max_ratio, seed=args.cluster_split_seed, logger=logger) # Chunk training data if too large to load in memory all at once if args.num_chunks > 1: os.makedirs(args.chunk_temp_dir, exist_ok=True) train_paths = [] if args.moe: chunked_sources = [td.chunk(args.num_chunks) for td in train_data] chunks = [] for i in range(args.num_chunks): chunks.append([source[i] for source in chunked_sources]) else: chunks = train_data.chunk(args.num_chunks) for i in range(args.num_chunks): chunk_path = os.path.join(args.chunk_temp_dir, str(i) + '.txt') memo_path = os.path.join(args.chunk_temp_dir, 'memo' + str(i) + '.txt') with open(chunk_path, 'wb') as f: pickle.dump(chunks[i], f) train_paths.append((chunk_path, memo_path)) train_data = train_paths # Get loss and metric functions loss_func = get_loss_func(args) metric_func = get_metric_func(metric=args.metric, args=args) # Set up test set evaluation test_smiles, test_targets = test_data.smiles(), test_data.targets() if args.maml: # TODO refactor test_targets = [] for task_idx in range(len(data.data[0].targets)): _, task_test_data, _ = test_data.sample_maml_task(args, seed=0) test_targets += task_test_data.targets() if args.dataset_type == 'bert_pretraining': sum_test_preds = { 'features': np.zeros((len(test_smiles), args.features_size)) if args.features_size is not None else None, 'vocab': np.zeros((len(test_targets['vocab']), args.vocab.output_size)) } elif args.dataset_type == 'kernel': sum_test_preds = np.zeros((len(test_targets), args.num_tasks)) else: sum_test_preds = np.zeros((len(test_smiles), args.num_tasks)) if args.maml: sum_test_preds = None # annoying to determine exact size; will initialize later if args.dataset_type == 'bert_pretraining': # Only predict targets that are masked out test_targets['vocab'] = [ target if mask == 0 else None for target, mask in zip(test_targets['vocab'], test_data.mask()) ] # Train ensemble of models for model_idx in range(args.ensemble_size): # Tensorboard writer save_dir = os.path.join(args.save_dir, f'model_{model_idx}') os.makedirs(save_dir, exist_ok=True) writer = SummaryWriter(log_dir=save_dir) # Load/build model if args.checkpoint_paths is not None: debug( f'Loading model {model_idx} from {args.checkpoint_paths[model_idx]}' ) model = load_checkpoint(args.checkpoint_paths[model_idx], current_args=args, logger=logger) else: debug(f'Building model {model_idx}') model = build_model(args) debug(model) debug(f'Number of parameters = {param_count(model):,}') if args.cuda: debug('Moving model to cuda') model = model.cuda() # Ensure that model is saved in correct location for evaluation if 0 epochs save_checkpoint(os.path.join(save_dir, 'model.pt'), model, scaler, features_scaler, args) if args.adjust_weight_decay: args.pnorm_target = compute_pnorm(model) # Optimizers optimizer = build_optimizer(model, args) # Learning rate schedulers scheduler = build_lr_scheduler(optimizer, args) # Run training best_score = float('inf') if args.minimize_score else -float('inf') best_epoch, n_iter = 0, 0 for epoch in trange(args.epochs): debug(f'Epoch {epoch}') if args.prespecified_chunk_dir is not None: # load some different random chunks each epoch train_data, val_data = load_prespecified_chunks(args, logger) debug('Loaded prespecified chunks for epoch') if args.dataset_type == 'unsupervised': # won't work with moe full_data = MoleculeDataset(train_data.data + val_data.data) generate_unsupervised_cluster_labels( build_model(args), full_data, args) # cluster with a new random init model.create_ffn( args ) # reset the ffn since we're changing targets-- we're just pretraining the encoder. optimizer.param_groups.pop() # remove ffn parameters optimizer.add_param_group({ 'params': model.ffn.parameters(), 'lr': args.init_lr[1], 'weight_decay': args.weight_decay[1] }) if args.cuda: model.ffn.cuda() if args.gradual_unfreezing: if epoch % args.epochs_per_unfreeze == 0: unfroze_layer = model.unfreeze_next( ) # consider just stopping early after we have nothing left to unfreeze? if unfroze_layer: debug('Unfroze last frozen layer') n_iter = train(model=model, data=train_data, loss_func=loss_func, optimizer=optimizer, scheduler=scheduler, args=args, n_iter=n_iter, logger=logger, writer=writer, chunk_names=(args.num_chunks > 1), val_smiles=val_smiles if args.adversarial else None, test_smiles=test_smiles if args.adversarial or args.moe else None) if isinstance(scheduler, ExponentialLR): scheduler.step() val_scores = evaluate(model=model, data=val_data, metric_func=metric_func, args=args, scaler=scaler, logger=logger) if args.dataset_type == 'bert_pretraining': if val_scores['features'] is not None: debug( f'Validation features rmse = {val_scores["features"]:.6f}' ) writer.add_scalar('validation_features_rmse', val_scores['features'], n_iter) val_scores = [val_scores['vocab']] # Average validation score avg_val_score = np.nanmean(val_scores) debug(f'Validation {args.metric} = {avg_val_score:.6f}') writer.add_scalar(f'validation_{args.metric}', avg_val_score, n_iter) if args.show_individual_scores: # Individual validation scores for task_name, val_score in zip(args.task_names, val_scores): if task_name in desired_labels: debug( f'Validation {task_name} {args.metric} = {val_score:.6f}' ) writer.add_scalar( f'validation_{task_name}_{args.metric}', val_score, n_iter) # Save model checkpoint if improved validation score, or always save it if unsupervised if args.minimize_score and avg_val_score < best_score or \ not args.minimize_score and avg_val_score > best_score or \ args.dataset_type == 'unsupervised': best_score, best_epoch = avg_val_score, epoch save_checkpoint(os.path.join(save_dir, 'model.pt'), model, scaler, features_scaler, args) if args.dataset_type == 'unsupervised': return [0] # rest of this is meaningless when unsupervised # Evaluate on test set using model with best validation score info( f'Model {model_idx} best validation {args.metric} = {best_score:.6f} on epoch {best_epoch}' ) model = load_checkpoint(os.path.join(save_dir, 'model.pt'), cuda=args.cuda, logger=logger) if args.split_test_by_overlap_dataset is not None: overlap_data = get_data(path=args.split_test_by_overlap_dataset, logger=logger) overlap_smiles = set(overlap_data.smiles()) test_data_intersect, test_data_nonintersect = [], [] for d in test_data.data: if d.smiles in overlap_smiles: test_data_intersect.append(d) else: test_data_nonintersect.append(d) test_data_intersect, test_data_nonintersect = MoleculeDataset( test_data_intersect), MoleculeDataset(test_data_nonintersect) for name, td in [('Intersect', test_data_intersect), ('Nonintersect', test_data_nonintersect)]: test_preds = predict(model=model, data=td, args=args, scaler=scaler, logger=logger) test_scores = evaluate_predictions( preds=test_preds, targets=td.targets(), metric_func=metric_func, dataset_type=args.dataset_type, args=args, logger=logger) avg_test_score = np.nanmean(test_scores) info( f'Model {model_idx} test {args.metric} for {name} = {avg_test_score:.6f}' ) if len( test_data ) == 0: # just get some garbage results without crashing; in this case we didn't care anyway test_preds, test_scores = sum_test_preds, [ 0 for _ in range(len(args.task_names)) ] else: test_preds = predict(model=model, data=test_data, args=args, scaler=scaler, logger=logger) test_scores = evaluate_predictions(preds=test_preds, targets=test_targets, metric_func=metric_func, dataset_type=args.dataset_type, args=args, logger=logger) if args.maml: if sum_test_preds is None: sum_test_preds = np.zeros(np.array(test_preds).shape) if args.dataset_type == 'bert_pretraining': if test_preds['features'] is not None: sum_test_preds['features'] += np.array(test_preds['features']) sum_test_preds['vocab'] += np.array(test_preds['vocab']) else: sum_test_preds += np.array(test_preds) if args.dataset_type == 'bert_pretraining': if test_preds['features'] is not None: debug( f'Model {model_idx} test features rmse = {test_scores["features"]:.6f}' ) writer.add_scalar('test_features_rmse', test_scores['features'], 0) test_scores = [test_scores['vocab']] # Average test score avg_test_score = np.nanmean(test_scores) info(f'Model {model_idx} test {args.metric} = {avg_test_score:.6f}') writer.add_scalar(f'test_{args.metric}', avg_test_score, 0) if args.show_individual_scores: # Individual test scores for task_name, test_score in zip(args.task_names, test_scores): if task_name in desired_labels: info( f'Model {model_idx} test {task_name} {args.metric} = {test_score:.6f}' ) writer.add_scalar(f'test_{task_name}_{args.metric}', test_score, n_iter) # Evaluate ensemble on test set if args.dataset_type == 'bert_pretraining': avg_test_preds = { 'features': (sum_test_preds['features'] / args.ensemble_size).tolist() if sum_test_preds['features'] is not None else None, 'vocab': (sum_test_preds['vocab'] / args.ensemble_size).tolist() } else: avg_test_preds = (sum_test_preds / args.ensemble_size).tolist() if len(test_data ) == 0: # just return some garbage when we didn't want test data ensemble_scores = test_scores else: ensemble_scores = evaluate_predictions(preds=avg_test_preds, targets=test_targets, metric_func=metric_func, dataset_type=args.dataset_type, args=args, logger=logger) # Average ensemble score if args.dataset_type == 'bert_pretraining': if ensemble_scores['features'] is not None: info( f'Ensemble test features rmse = {ensemble_scores["features"]:.6f}' ) writer.add_scalar('ensemble_test_features_rmse', ensemble_scores['features'], 0) ensemble_scores = [ensemble_scores['vocab']] avg_ensemble_test_score = np.nanmean(ensemble_scores) info(f'Ensemble test {args.metric} = {avg_ensemble_test_score:.6f}') writer.add_scalar(f'ensemble_test_{args.metric}', avg_ensemble_test_score, 0) # Individual ensemble scores if args.show_individual_scores: for task_name, ensemble_score in zip(args.task_names, ensemble_scores): info( f'Ensemble test {task_name} {args.metric} = {ensemble_score:.6f}' ) return ensemble_scores