def class_balance(data_path: str, split_type: str):
# Update args
args.val_fold_index, args.test_fold_index = 1, 2
args.split_type = 'predetermined'
# Load data
data = get_data(path=args.data_path,
smiles_column=args.smiles_column,
target_columns=args.target_columns)
args.task_names = args.target_columns or get_task_names(
path=args.data_path, smiles_column=args.smiles_column)
# Average class sizes
all_class_sizes = {'train': [], 'val': [], 'test': []}
for i in range(10):
print(f'Fold {i}')
# Update args
data_name = os.path.splitext(os.path.basename(data_path))[0]
args.folds_file = f'/data/rsg/chemistry/yangk/lsc_experiments_dump_splits/data/{data_name}/{split_type}/fold_{i}/0/split_indices.pckl'
if not os.path.exists(args.folds_file):
print(f'Fold indices do not exist')
continue
# Split data
train_data, val_data, test_data = split_data(
data=data, split_type=args.split_type, args=args)
# Determine class balance
for data_split, split_name in [(train_data, 'train'),
(val_data, 'val'), (test_data, 'test')]:
class_sizes = get_class_sizes(data_split)
print(f'Class sizes for {split_name}')
for i, task_class_sizes in enumerate(class_sizes):
print(
f'{args.task_names[i]} '
f'{", ".join(f"{cls}: {size * 100:.2f}%" for cls, size in enumerate(task_class_sizes))}'
)
all_class_sizes[split_name].append(class_sizes)
print()
# Mean and std across folds
for split_name in ['train', 'val', 'test']:
print(f'Average class sizes for {split_name}')
mean_class_sizes, std_class_sizes = np.mean(
all_class_sizes[split_name],
axis=0), np.std(all_class_sizes[split_name], axis=0)
for i, (mean_task_class_sizes, std_task_class_sizes) in enumerate(
zip(mean_class_sizes, std_class_sizes)):
print(
f'{args.task_names[i]} '
f'{", ".join(f"{cls}: {mean_size * 100:.2f}% +/- {std_size * 100:.2f}%" for cls, (mean_size, std_size) in enumerate(zip(mean_task_class_sizes, std_task_class_sizes)))}'
)
def cross_validate_sklearn(args: Namespace,
logger: Logger = None) -> Tuple[float, float]:
# TODO: Multi-task might not work with RF.
info = logger.info if logger is not None else print
args.task_names = get_task_names(args.data_path, args.data_format)
init_seed = args.seed
save_dir = args.save_dir
# 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}')
makedirs(args.save_dir)
model_scores = run_sklearn(args, logger)
all_scores.append(model_scores)
all_scores = np.array(all_scores)
# 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}'
)
# Report scores across folds
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}')
return mean_score, std_score
def plot_distribution(data_path: str, save_dir: str, bins: int):
"""
Plots the distribution of values of a dataset.
:param data_path: Path to data CSV file.
:param save_dir: Directory where plot PNGs will be saved.
:param bins: Number of bins in histogram.
"""
# Get values
task_names = get_task_names(data_path)
data = get_data(data_path)
targets = data.targets()
# Arrange values by task
data_size, num_tasks = len(targets), len(task_names)
values = [[targets[i][j] for i in range(data_size)]
for j in range(num_tasks)]
# Plot distributions for each task
data_name = os.path.basename(data_path).replace('.csv', '')
for i in range(num_tasks):
plt.clf()
plt.hist(values[i], bins=bins)
# Save plot
plt.title('{} - {}'.format(data_name, task_names[i]))
plt.xlabel(task_names[i])
plt.ylabel('Frequency')
plt.savefig(
os.path.join(save_dir, '{}_{}.png'.format(data_name,
task_names[i])))
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 cross_validate(args: TrainArgs,
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 = args.target_columns or get_task_names(args.data_path)
# 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}')
makedirs(args.save_dir)
model_scores, uncertainty_estimator = run_training(args, logger)
# Save one model for each fold
if uncertainty_estimator:
process_estimator(uncertainty_estimator, logger, args, fold_num)
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):
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):
info(
f'Overall test {task_name} {args.metric} = '
f'{np.nanmean(all_scores[:, task_num]):.6f} +/- {np.nanstd(all_scores[:, task_num]):.6f}'
)
return mean_score, std_score
def predict_sklearn(args: SklearnPredictArgs):
print('Loading data')
data = get_data(path=args.test_path,
smiles_column=args.smiles_column,
target_columns=[])
print('Computing morgan fingerprints')
morgan_fingerprint = get_features_generator('morgan')
for datapoint in tqdm(data, total=len(data)):
datapoint.set_features(
morgan_fingerprint(mol=datapoint.smiles,
radius=args.radius,
num_bits=args.num_bits))
print(
f'Predicting with an ensemble of {len(args.checkpoint_paths)} models')
sum_preds = np.zeros((len(data), args.num_tasks))
for checkpoint_path in tqdm(args.checkpoint_paths,
total=len(args.checkpoint_paths)):
with open(checkpoint_path, 'rb') as f:
model = pickle.load(f)
model_preds = predict(model=model,
model_type=args.model_type,
dataset_type=args.dataset_type,
features=data.features())
sum_preds += np.array(model_preds)
# Ensemble predictions
avg_preds = sum_preds / len(args.checkpoint_paths)
avg_preds = avg_preds.tolist()
print(f'Saving predictions to {args.preds_path}')
assert len(data) == len(avg_preds)
makedirs(args.preds_path, isfile=True)
# Copy predictions over to data
task_names = get_task_names(path=args.test_path)
for datapoint, preds in zip(data, avg_preds):
for pred_name, pred in zip(task_names, preds):
datapoint.row[pred_name] = pred
# Save
with open(args.preds_path, 'w') as f:
writer = csv.DictWriter(f, fieldnames=data[0].row.keys())
writer.writeheader()
for datapoint in data:
writer.writerow(datapoint.row)
def predict_sklearn(args: Namespace):
print('Loading data')
data = get_data(path=args.test_path)
print('Computing morgan fingerprints')
morgan_fingerprint = get_features_generator('morgan')
for datapoint in tqdm(data, total=len(data)):
datapoint.set_features(morgan_fingerprint(mol=datapoint.smiles, radius=args.radius, num_bits=args.num_bits))
print(f'Predicting with an ensemble of {len(args.checkpoint_paths)} models')
sum_preds = np.zeros((len(data), args.num_tasks))
for checkpoint_path in tqdm(args.checkpoint_paths, total=len(args.checkpoint_paths)):
with open(checkpoint_path, 'rb') as f:
model = pickle.load(f)
model_preds = predict(
model=model,
model_type=args.model_type,
dataset_type=args.dataset_type,
features=data.features()
)
sum_preds += np.array(model_preds)
# Ensemble predictions
avg_preds = sum_preds / len(args.checkpoint_paths)
avg_preds = avg_preds.tolist()
print('Saving predictions')
assert len(data) == len(avg_preds)
makedirs(args.preds_path, isfile=True)
with open(args.preds_path, 'w') as f:
writer = csv.writer(f)
writer.writerow(['smiles'] + get_task_names(args.test_path))
for smiles, pred in zip(data.smiles(), avg_preds):
writer.writerow([smiles] + pred)
def predict_sklearn(args: SklearnPredictArgs):
if args.parcel_size and args.max_data_size:
num_iterations = math.ceil(args.max_data_size / args.parcel_size)
max_data_size = args.parcel_size
else:
num_iterations = 1
max_data_size = args.max_data_size
offset = 0
for iteration in range(num_iterations):
if iteration > 0:
offset = offset + args.parcel_size
max_data_size = max_data_size + args.parcel_size
print('Loading data')
data = get_data(path=args.test_path,
smiles_column=args.smiles_column,
target_columns=[],
max_data_size=max_data_size,
data_offset=offset)
print('Computing morgan fingerprints')
morgan_fingerprint = get_features_generator('morgan')
for datapoint in tqdm(data, total=len(data)):
datapoint.set_features(
morgan_fingerprint(mol=datapoint.smiles,
radius=args.radius,
num_bits=args.num_bits))
print(
f'Predicting with an ensemble of {len(args.checkpoint_paths)} models'
)
sum_preds = np.zeros((len(data), args.num_tasks))
for checkpoint_path in tqdm(args.checkpoint_paths,
total=len(args.checkpoint_paths)):
with open(checkpoint_path, 'rb') as f:
model = pickle.load(f)
model_preds = predict(model=model,
model_type=args.model_type,
dataset_type=args.dataset_type,
features=data.features())
sum_preds += np.array(model_preds)
# Ensemble predictions
avg_preds = sum_preds / len(args.checkpoint_paths)
avg_preds = avg_preds.tolist()
print(f'Saving predictions to {args.preds_path}')
assert len(data) == len(avg_preds)
makedirs(args.preds_path, isfile=True)
# Copy predictions over to data
task_names = get_task_names(path=args.test_path)
for datapoint, preds in zip(data, avg_preds):
for pred_name, pred in zip(task_names, preds):
datapoint.row[pred_name] = pred
# Save
if iteration != 0:
name, ext = os.path.splitext(args.preds_path)
preds_path = "{name}.{it}{ext}".format(name=name,
it=iteration,
ext=ext)
else:
preds_path = args.preds_path
with open(args.preds_path, 'w') as f:
writer = csv.DictWriter(f, fieldnames=data[0].row.keys())
writer.writeheader()
for datapoint in data:
writer.writerow(datapoint.row)
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)
data = get_data(path=args.data_path, args=args, logger=logger)
args.num_tasks = data.num_tasks()
args.features_size = data.features_size()
debug(f'Number of tasks = {args.num_tasks}')
# Split data
debug(f'Splitting data with seed {args.seed}')
if args.separate_test_path:
test_data = get_data(path=args.separate_test_path,
args=args,
features_path=args.separate_test_features_path,
logger=logger)
if args.separate_val_path:
val_data = get_data(path=args.separate_val_path,
args=args,
features_path=args.separate_val_features_path,
logger=logger)
if args.separate_val_path and args.separate_test_path:
train_data = data
elif args.separate_val_path:
train_data, _, test_data = split_data(data=data,
split_type=args.split_type,
sizes=(0.8, 0.2, 0.0),
seed=args.seed,
args=args,
logger=logger)
elif args.separate_test_path:
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:
print('=' * 100)
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)
###my_code###
train_df = get_data_df(train_data)
train_df.to_csv(
'~/PycharmProjects/CMPNN-master/data/24w_train_df_seed0.csv')
val_df = get_data_df(val_data)
val_df.to_csv(
'~/PycharmProjects/CMPNN-master/data/24w_val_df_seed0.csv')
test_df = get_data_df(test_data)
test_df.to_csv(
'~/PycharmProjects/CMPNN-master/data/24w_test_df_seed0.csv')
##########
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.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)
if args.features_scaling:
features_scaler = train_data.normalize_features(replace_nan_token=0)
val_data.normalize_features(features_scaler)
test_data.normalize_features(features_scaler)
else:
features_scaler = None
args.train_data_size = len(train_data)
debug(
f'Total size = {len(data):,} | '
f'train size = {len(train_data):,} | val size = {len(val_data):,} | test size = {len(test_data):,}'
)
# Initialize scaler and scale training targets by subtracting mean and dividing standard deviation (regression only)
if args.dataset_type == 'regression':
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
# Get loss and metric functions
loss_func = get_loss_func(args)
metric_func = get_metric_func(metric=args.metric)
# Set up test set evaluation
test_smiles, test_targets = test_data.smiles(), test_data.targets()
if args.dataset_type == 'multiclass':
sum_test_preds = np.zeros(
(len(test_smiles), args.num_tasks, args.multiclass_num_classes))
else:
sum_test_preds = np.zeros((len(test_smiles), args.num_tasks))
# 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}')
makedirs(save_dir)
try:
writer = SummaryWriter(log_dir=save_dir)
except:
writer = SummaryWriter(logdir=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)
# 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 range(args.epochs):
debug(f'Epoch {epoch}')
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)
if isinstance(scheduler, ExponentialLR):
scheduler.step()
val_scores = evaluate(model=model,
data=val_data,
num_tasks=args.num_tasks,
metric_func=metric_func,
batch_size=args.batch_size,
dataset_type=args.dataset_type,
scaler=scaler,
logger=logger)
# 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):
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
if args.minimize_score and avg_val_score < best_score or \
not args.minimize_score and avg_val_score > best_score:
best_score, best_epoch = avg_val_score, epoch
save_checkpoint(os.path.join(save_dir, 'model.pt'), model,
scaler, features_scaler, args)
# 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)
test_preds = predict(model=model,
data=test_data,
batch_size=args.batch_size,
scaler=scaler)
test_scores = evaluate_predictions(preds=test_preds,
targets=test_targets,
num_tasks=args.num_tasks,
metric_func=metric_func,
dataset_type=args.dataset_type,
logger=logger)
if len(test_preds) != 0:
sum_test_preds += np.array(test_preds)
# 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):
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
avg_test_preds = (sum_test_preds / args.ensemble_size).tolist()
ensemble_scores = evaluate_predictions(preds=avg_test_preds,
targets=test_targets,
num_tasks=args.num_tasks,
metric_func=metric_func,
dataset_type=args.dataset_type,
logger=logger)
# Average ensemble score
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
def predict_feature(args, logger, model, external_test_path):
if logger is not None:
debug, info = logger.debug, logger.info
else:
debug = info = print
# Print command line
debug('Command line')
debug(f'python {" ".join(sys.argv)}')
# Print args
debug('Args')
debug(args)
# Save args
args.save(os.path.join(args.save_dir, 'args.json'))
# Set pytorch seed for random initial weights
torch.manual_seed(args.pytorch_seed)
# Get data
debug('Loading data')
args.task_names = args.target_columns or get_task_names(args.data_path)
data = get_data(path=args.data_path, args=args, logger=logger)
args.num_tasks = data.num_tasks()
args.features_size = data.features_size()
debug(f'Number of tasks = {args.num_tasks}')
external_test_data = get_data(path=external_test_path,
args=args,
logger=logger)
# Split data
debug(f'Splitting data with seed {args.seed}')
if args.separate_test_path:
test_data = get_data(path=args.separate_test_path,
args=args,
features_path=args.separate_test_features_path,
logger=logger)
if args.separate_val_path:
val_data = get_data(path=args.separate_val_path,
args=args,
features_path=args.separate_val_features_path,
logger=logger)
if args.separate_val_path and args.separate_test_path:
train_data = data
elif args.separate_val_path:
train_data, _, test_data = split_data(data=data,
split_type=args.split_type,
sizes=(0.8, 0.0, 0.2),
seed=args.seed,
args=args,
logger=logger)
elif args.separate_test_path:
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)
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.save_smiles_splits:
save_smiles_splits(train_data=train_data,
val_data=val_data,
test_data=test_data,
data_path=args.data_path,
save_dir=args.save_dir)
if args.features_scaling:
features_scaler = train_data.normalize_features(replace_nan_token=0)
external_test_data.normalize_features(features_scaler)
else:
features_scaler = None
args.train_data_size = len(train_data)
debug(
f'Total size = {len(data):,} | '
f'train size = {len(train_data):,} | val size = {len(val_data):,} | test size = {len(test_data):,}'
)
scaler = None
# Automatically determine whether to cache
if len(data) <= args.cache_cutoff:
cache = True
num_workers = 0
else:
cache = False
num_workers = args.num_workers
# Create data loaders
train_data_loader = MoleculeDataLoader(dataset=train_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache,
class_balance=args.class_balance,
shuffle=False,
seed=args.seed)
external_test_loader = MoleculeDataLoader(dataset=external_test_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache,
class_balance=args.class_balance,
shuffle=False,
seed=args.seed)
external_test_preds, external_test_feature = predict(
model=model, data_loader=external_test_loader, scaler=scaler)
external_test_smiles, external_test_targets = external_test_data.smiles(
), external_test_data.targets()
return external_test_smiles, external_test_feature, external_test_preds, external_test_targets
def cross_validate(args: TrainArgs,
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
args.task_names = get_task_names(path=args.data_path,
smiles_column=args.smiles_column,
target_columns=args.target_columns,
ignore_columns=args.ignore_columns)
# 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}')
makedirs(args.save_dir)
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(args.task_names, scores):
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(args.task_names):
info(
f'Overall test {task_name} {args.metric} = '
f'{np.nanmean(all_scores[:, task_num]):.6f} +/- {np.nanstd(all_scores[:, task_num]):.6f}'
)
# Save scores
with open(os.path.join(save_dir, 'test_scores.csv'), 'w') as f:
writer = csv.writer(f)
writer.writerow([
'Task', f'Mean {args.metric}', f'Standard deviation {args.metric}'
] + [f'Fold {i} {args.metric}' for i in range(args.num_folds)])
for task_num, task_name in enumerate(args.task_names):
task_scores = all_scores[:, task_num]
mean, std = np.nanmean(task_scores), np.nanstd(task_scores)
writer.writerow([task_name, mean, std] + task_scores.tolist())
return mean_score, std_score
def run_sklearn(args: SklearnTrainArgs, logger: Logger = None) -> List[float]:
if logger is not None:
debug, info = logger.debug, logger.info
else:
debug = info = print
debug(pformat(vars(args)))
metric_func = get_metric_func(args.metric)
debug('Loading data')
data = get_data(path=args.data_path,
smiles_column=args.smiles_column,
target_columns=args.target_columns)
args.task_names = get_task_names(path=args.data_path,
smiles_column=args.smiles_column,
target_columns=args.target_columns,
ignore_columns=args.ignore_columns)
if args.model_type == 'svm' and data.num_tasks() != 1:
raise ValueError(
f'SVM can only handle single-task data but found {data.num_tasks()} tasks'
)
debug(f'Splitting data with seed {args.seed}')
# Need to have val set so that train and test sets are the same as when doing MPN
train_data, _, test_data = split_data(data=data,
split_type=args.split_type,
seed=args.seed,
sizes=args.split_sizes,
args=args)
debug(
f'Total size = {len(data):,} | train size = {len(train_data):,} | test size = {len(test_data):,}'
)
debug('Computing morgan fingerprints')
morgan_fingerprint = get_features_generator('morgan')
for dataset in [train_data, test_data]:
for datapoint in tqdm(dataset, total=len(dataset)):
datapoint.set_features(
morgan_fingerprint(mol=datapoint.smiles,
radius=args.radius,
num_bits=args.num_bits))
debug('Building model')
if args.dataset_type == 'regression':
if args.model_type == 'random_forest':
model = RandomForestRegressor(n_estimators=args.num_trees,
n_jobs=-1)
elif args.model_type == 'svm':
model = SVR()
else:
raise ValueError(f'Model type "{args.model_type}" not supported')
elif args.dataset_type == 'classification':
if args.model_type == 'random_forest':
model = RandomForestClassifier(n_estimators=args.num_trees,
n_jobs=-1,
class_weight=args.class_weight)
elif args.model_type == 'svm':
model = SVC()
else:
raise ValueError(f'Model type "{args.model_type}" not supported')
else:
raise ValueError(f'Dataset type "{args.dataset_type}" not supported')
debug(model)
model.train_args = args.as_dict()
debug('Training')
if args.single_task:
scores = single_task_sklearn(model=model,
train_data=train_data,
test_data=test_data,
metric_func=metric_func,
args=args,
logger=logger)
else:
scores = multi_task_sklearn(model=model,
train_data=train_data,
test_data=test_data,
metric_func=metric_func,
args=args,
logger=logger)
info(f'Test {args.metric} = {np.nanmean(scores)}')
return scores
def new_noise(args: TrainArgs, 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.
"""
debug = info = print
# Get data
args.task_names = args.target_columns or get_task_names(args.data_path)
data = get_data(path=args.data_path, args=args, logger=logger)
args.num_tasks = data.num_tasks()
args.features_size = data.features_size()
# Split data
debug(f'Splitting data with seed {args.seed}')
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)
if args.features_scaling:
features_scaler = train_data.normalize_features(replace_nan_token=0)
val_data.normalize_features(features_scaler)
test_data.normalize_features(features_scaler)
else:
features_scaler = None
args.train_data_size = len(train_data)
# Initialize scaler and scale training targets by subtracting mean and dividing standard deviation (regression only)
if args.dataset_type == 'regression':
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
# Get loss and metric functions
loss_func = neg_log_like
metric_func = get_metric_func(metric=args.metric)
# Set up test set evaluation
test_smiles, test_targets = test_data.smiles(), test_data.targets()
sum_test_preds = np.zeros((len(test_smiles), args.num_tasks))
# Automatically determine whether to cache
if len(data) <= args.cache_cutoff:
cache = True
num_workers = 0
else:
cache = False
num_workers = args.num_workers
# Create data loaders
train_data_loader = MoleculeDataLoader(dataset=train_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache)
val_data_loader = MoleculeDataLoader(dataset=val_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache)
test_data_loader = MoleculeDataLoader(dataset=test_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache)
###########################################
########## Outer loop over ensemble members
###########################################
for model_idx in range(args.ensemble_start_idx,
args.ensemble_start_idx + args.ensemble_size):
# load the model
if (args.method == 'map') or (args.method == 'swag') or (args.method
== 'sgld'):
model = load_checkpoint(args.checkpoint_path +
f'/model_{model_idx}/model.pt',
device=args.device,
logger=logger)
if args.method == 'gp':
args.num_inducing_points = 1200
fake_model = MoleculeModel(args)
fake_model.featurizer = True
feature_extractor = fake_model
inducing_points = initial_inducing_points(train_data_loader,
feature_extractor, args)
gp_layer = GPLayer(inducing_points, args.num_tasks)
model = load_checkpoint(
args.checkpoint_path + f'/model_{model_idx}/DKN_model.pt',
device=args.device,
logger=None,
template=DKLMoleculeModel(MoleculeModel(args, featurizer=True),
gp_layer))
if args.method == 'dropR' or args.method == 'dropA':
model = load_checkpoint(args.checkpoint_path +
f'/model_{model_idx}/model.pt',
device=args.device,
logger=logger)
if args.method == 'bbp':
template = MoleculeModelBBP(args)
for layer in template.children():
if isinstance(layer, BayesLinear):
layer.init_rho(args.rho_min_bbp, args.rho_max_bbp)
for layer in template.encoder.encoder.children():
if isinstance(layer, BayesLinear):
layer.init_rho(args.rho_min_bbp, args.rho_max_bbp)
model = load_checkpoint(args.checkpoint_path +
f'/model_{model_idx}/model_bbp.pt',
device=args.device,
logger=None,
template=template)
if args.method == 'dun':
args.prior_sig_dun = 0.05
args.depth_min = 1
args.depth_max = 5
args.rho_min_dun = -5.5
args.rho_max_dun = -5
args.log_cat_init = 0
template = MoleculeModelDUN(args)
for layer in template.children():
if isinstance(layer, BayesLinear):
layer.init_rho(args.rho_min_dun, args.rho_max_dun)
for layer in template.encoder.encoder.children():
if isinstance(layer, BayesLinear):
layer.init_rho(args.rho_min_dun, args.rho_max_dun)
template.create_log_cat(args)
model = load_checkpoint(args.checkpoint_path +
f'/model_{model_idx}/model_dun.pt',
device=args.device,
logger=None,
template=template)
# make results_dir
results_dir = os.path.join(args.results_dir, f'model_{model_idx}')
makedirs(results_dir)
# train_preds, train_targets
train_preds = predict(model=model,
data_loader=train_data_loader,
args=args,
scaler=scaler,
test_data=False,
bbp_sample=False)
train_preds = np.array(train_preds)
train_targets = np.array(train_targets)
# compute tstats
tstats = np.ones((12, 3))
for task in range(12):
resid = train_preds[:, task] - train_targets[:, task]
tstats[task] = np.array(stats.t.fit(resid, floc=0.0))
##################################
########## Inner loop over samples
##################################
for sample_idx in range(args.samples):
# save down
np.savez(os.path.join(results_dir, f'tstats_{sample_idx}'), tstats)
print('done one')
def run_meta_training(args: TrainArgs, 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
# Print command line
debug('Command line')
debug(f'python {" ".join(sys.argv)}')
# Print args
debug('Args')
debug(args)
# Save args
args.save(os.path.join(args.save_dir, 'args.json'))
# Set pytorch seed for random initial weights
torch.manual_seed(args.pytorch_seed)
# Get data
debug('Loading data')
args.task_names = args.target_columns or get_task_names(args.data_path)
data = get_data(path=args.data_path, args=args, logger=logger)
args.num_tasks = data.num_tasks()
args.features_size = data.features_size()
debug(f'Number of tasks = {args.num_tasks}')
# Split data
# debug(f'Splitting data with seed {args.seed}')
# if args.separate_test_path:
# test_data = get_data(path=args.separate_test_path, args=args, features_path=args.separate_test_features_path, logger=logger)
# if args.separate_val_path:
# val_data = get_data(path=args.separate_val_path, args=args, features_path=args.separate_val_features_path, logger=logger)
# if args.separate_val_path and args.separate_test_path:
# train_data = data
# elif args.separate_val_path:
# train_data, _, test_data = split_data(data=data, split_type=args.split_type, sizes=(0.8, 0.0, 0.2), seed=args.seed, args=args, logger=logger)
# elif args.separate_test_path:
# 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)
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.save_smiles_splits:
# save_smiles_splits(
# train_data=train_data,
# val_data=val_data,
# test_data=test_data,
# data_path=args.data_path,
# save_dir=args.save_dir
# )
# If this happens, then need to move this logic into the task data loader
# when it creates the datasets!
# if args.features_scaling:
# features_scaler = train_data.normalize_features(replace_nan_token=0)
# val_data.normalize_features(features_scaler)
# test_data.normalize_features(features_scaler)
# else:
# features_scaler = None
# args.train_data_size = len(train_data)
# debug(f'Total size = {len(data):,} | '
# f'train size = {len(train_data):,} | val size = {len(val_data):,} | test size = {len(test_data):,}')
# Initialize scaler and scale training targets by subtracting mean and dividing standard deviation (regression only)
# if args.dataset_type == 'regression':
# 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
# Get loss and metric functions
loss_func = get_loss_func(args)
metric_func = get_metric_func(metric=args.metric)
# Set up test set evaluation
# test_smiles, test_targets = test_data.smiles(), test_data.targets()
# if args.dataset_type == 'multiclass':
# sum_test_preds = np.zeros((len(test_smiles), args.num_tasks, args.multiclass_num_classes))
# else:
# sum_test_preds = np.zeros((len(test_smiles), args.num_tasks))
# Automatically determine whether to cache
if len(data) <= args.cache_cutoff:
cache = True
num_workers = 0
else:
cache = False
num_workers = args.num_workers
# Set up MetaTaskDataLoaders, which takes care of task splits under the hood
# Set up task splits into T_tr, T_val, T_test
assert args.chembl_assay_metadata_pickle_path is not None
with open(args.chembl_assay_metadata_pickle_path +
'chembl_128_assay_type_to_names.pickle', 'rb') as handle:
chembl_128_assay_type_to_names = pickle.load(handle)
with open(args.chembl_assay_metadata_pickle_path +
'chembl_128_assay_name_to_type.pickle', 'rb') as handle:
chembl_128_assay_name_to_type = pickle.load(handle)
"""
Copy GSK implementation of task split
We have 5 Task types remaining
ADME (A)
Toxicity (T)
Unassigned (U)
Binding (B)
Functional (F)
resulting in 902 tasks.
For T_val, randomly select 10 B and F tasks
For T_test, select another 10 B and F tasks and allocate all A, T, and U
tasks to the test split.
For T_train, allocate the remaining B and F tasks.
"""
import pdb; pdb.set_trace()
T_val_num_BF_tasks = args.meta_split_sizes_BF[0]
T_test_num_BF_tasks = args.meta_split_sizes_BF[1]
T_val_idx = T_val_num_BF_tasks
T_test_idx = T_val_num_BF_tasks + T_test_num_BF_tasks
chembl_id_to_idx = {chembl_id: idx for idx, chembl_id in enumerate(args.task_names)}
# Shuffle B and F tasks
randomized_B_tasks = np.copy(chembl_128_assay_type_to_names['B'])
np.random.shuffle(randomized_B_tasks)
randomized_B_task_indices = [chembl_id_to_idx[assay] for assay in
randomized_B_tasks]
randomized_F_tasks = np.copy(chembl_128_assay_type_to_names['F'])
np.random.shuffle(randomized_F_tasks)
randomized_F_task_indices = [chembl_id_to_idx[assay] for assay in
randomized_F_tasks]
# Grab B and F indices for T_val
T_val_B_task_indices = randomized_B_task_indices[:T_val_idx]
T_val_F_task_indices = randomized_F_task_indices[:T_val_idx]
# Grab B and F indices for T_test
T_test_B_task_indices = randomized_B_task_indices[T_val_idx:T_test_idx]
T_test_F_task_indices = randomized_F_task_indices[T_val_idx:T_test_idx]
# Grab all A, T and U indices for T_test
T_test_A_task_indices = [chembl_id_to_idx[assay] for assay in chembl_128_assay_type_to_names['A']]
T_test_T_task_indices = [chembl_id_to_idx[assay] for assay in chembl_128_assay_type_to_names['T']]
T_test_U_task_indices = [chembl_id_to_idx[assay] for assay in chembl_128_assay_type_to_names['U']]
# Slot remaining BF tasks into T_tr
T_tr_B_task_indices = randomized_B_task_indices[T_test_idx:]
T_tr_F_task_indices = randomized_F_task_indices[T_test_idx:]
T_tr = [0] * len(args.task_names)
T_val = [0] * len(args.task_names)
T_test = [0] * len(args.task_names)
# Now make task bit vectors
for idx_list in (T_tr_B_task_indices, T_tr_F_task_indices):
for idx in idx_list:
T_tr[idx] = 1
for idx_list in (T_val_B_task_indices, T_val_F_task_indices):
for idx in idx_list:
T_val[idx] = 1
for idx_list in (T_test_B_task_indices, T_test_F_task_indices, T_test_A_task_indices, T_test_T_task_indices, T_test_U_task_indices):
for idx in idx_list:
T_test[idx] = 1
"""
Random task split for testing
task_indices = list(range(len(args.task_names)))
np.random.shuffle(task_indices)
train_task_split, val_task_split, test_task_split = 0.9, 0, 0.1
train_task_cutoff = int(len(task_indices) * train_task_split)
train_task_idxs, test_task_idxs = [0] * len(task_indices), [0] * len(task_indices)
for idx in task_indices[:train_task_cutoff]:
train_task_idxs[idx] = 1
for idx in task_indices[train_task_cutoff:]:
test_task_idxs[idx] = 1
"""
train_meta_task_data_loader = MetaTaskDataLoader(
dataset=data,
tasks=T_tr,
sizes=args.meta_train_split_sizes,
args=args,
logger=logger)
val_meta_task_data_loader = MetaTaskDataLoader(
dataset=data,
tasks=T_val,
sizes=args.meta_test_split_sizes,
args=args,
logger=logger)
test_meta_task_data_loader = MetaTaskDataLoader(
dataset=data,
tasks=T_test,
sizes=args.meta_test_split_sizes,
args=args,
logger=logger)
import pdb; pdb.set_trace()
for meta_train_batch in train_meta_task_data_loader.tasks():
for train_task in meta_train_batch:
print('In inner loop')
continue
# 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}')
makedirs(save_dir)
try:
writer = SummaryWriter(log_dir=save_dir)
except:
writer = SummaryWriter(logdir=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], logger=logger)
else:
debug(f'Building model {model_idx}')
model = MoleculeModel(args)
debug(model)
debug(f'Number of parameters = {param_count(model):,}')
if args.cuda:
debug('Moving model to cuda')
model = model.to(args.device)
# 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)
# 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}')
n_iter = train(
model=model,
data_loader=train_data_loader,
loss_func=loss_func,
optimizer=optimizer,
scheduler=scheduler,
args=args,
n_iter=n_iter,
logger=logger,
writer=writer
)
if isinstance(scheduler, ExponentialLR):
scheduler.step()
val_scores = evaluate(
model=model,
data_loader=val_data_loader,
num_tasks=args.num_tasks,
metric_func=metric_func,
dataset_type=args.dataset_type,
scaler=scaler,
logger=logger
)
# 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):
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
if args.minimize_score and avg_val_score < best_score or \
not args.minimize_score and avg_val_score > best_score:
best_score, best_epoch = avg_val_score, epoch
save_checkpoint(os.path.join(save_dir, 'model.pt'), model, scaler, features_scaler, args)
# 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'), device=args.device, logger=logger)
test_preds = predict(
model=model,
data_loader=test_data_loader,
scaler=scaler
)
test_scores = evaluate_predictions(
preds=test_preds,
targets=test_targets,
num_tasks=args.num_tasks,
metric_func=metric_func,
dataset_type=args.dataset_type,
logger=logger
)
if len(test_preds) != 0:
sum_test_preds += np.array(test_preds)
# 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):
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)
writer.close()
# Evaluate ensemble on test set
avg_test_preds = (sum_test_preds / args.ensemble_size).tolist()
ensemble_scores = evaluate_predictions(
preds=avg_test_preds,
targets=test_targets,
num_tasks=args.num_tasks,
metric_func=metric_func,
dataset_type=args.dataset_type,
logger=logger
)
# Average ensemble score
avg_ensemble_test_score = np.nanmean(ensemble_scores)
info(f'Ensemble test {args.metric} = {avg_ensemble_test_score:.6f}')
# 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
def save_test_data(args: TrainArgs, 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.
"""
debug = info = print
# Print command line and args
debug('Command line')
debug(f'python {" ".join(sys.argv)}')
debug('Args')
debug(args)
# Get data
debug('Loading data')
args.task_names = args.target_columns or get_task_names(args.data_path)
data = get_data(path=args.data_path, args=args, logger=logger)
args.num_tasks = data.num_tasks()
args.features_size = data.features_size()
debug(f'Number of tasks = {args.num_tasks}')
# Split data
debug(f'Splitting data with seed {args.seed}')
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)
if args.features_scaling:
features_scaler = train_data.normalize_features(replace_nan_token=0)
val_data.normalize_features(features_scaler)
test_data.normalize_features(features_scaler)
else:
features_scaler = None
args.train_data_size = len(train_data)
debug(
f'Total size = {len(data):,} | '
f'train size = {len(train_data):,} | val size = {len(val_data):,} | test size = {len(test_data):,}'
)
# Initialize scaler and scale training targets by subtracting mean and dividing standard deviation (regression only)
if args.dataset_type == 'regression':
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
# Get loss and metric functions
loss_func = neg_log_like
metric_func = get_metric_func(metric=args.metric)
# Set up test set evaluation
test_smiles, test_targets = test_data.smiles(), test_data.targets()
sum_test_preds = np.zeros((len(test_smiles), args.num_tasks))
# save down test targets
np.savez('/home/willlamb/results/test_targets', np.array(test_targets))
return None
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
def cross_validate(args: TrainArgs,
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 = args.target_columns or get_task_names(args.data_path)
# Run training on different random seeds for each fold
all_scores = []
if args.seeds and (len(args.seeds) != args.num_folds):
raise ValueError(
"Num seeds provided must match the number of folds required")
for fold_num in range(args.num_folds):
info(f'Fold {fold_num}')
if args.seeds:
args.seed = args.seeds[fold_num]
else:
args.seed = init_seed + fold_num
args.save_dir = os.path.join(save_dir, f'fold_{fold_num}')
makedirs(args.save_dir)
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):
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}')
wandb.log({
'test_{}_mean'.format(args.metric): mean_score,
'test_{}_std'.format(args.metric): std_score
})
# save results
save_results(all_scores, [], task_names, args)
if args.show_individual_scores:
for task_num, task_name in enumerate(task_names):
info(
f'Overall test {task_name} {args.metric} = '
f'{np.nanmean(all_scores[:, task_num]):.6f} +/- {np.nanstd(all_scores[:, task_num]):.6f}'
)
return mean_score, std_score
def run_training(args: TrainArgs, 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.
"""
debug = info = print
# Print command line and args
debug('Command line')
debug(f'python {" ".join(sys.argv)}')
debug('Args')
debug(args)
# Save args
args.save(os.path.join(args.save_dir, 'args.json'))
# Get data
debug('Loading data')
args.task_names = args.target_columns or get_task_names(args.data_path)
data = get_data(path=args.data_path, args=args, logger=logger)
args.num_tasks = data.num_tasks()
args.features_size = data.features_size()
debug(f'Number of tasks = {args.num_tasks}')
# Split data
debug(f'Splitting data with seed {args.seed}')
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)
if args.features_scaling:
features_scaler = train_data.normalize_features(replace_nan_token=0)
val_data.normalize_features(features_scaler)
test_data.normalize_features(features_scaler)
else:
features_scaler = None
args.train_data_size = len(train_data)
debug(
f'Total size = {len(data):,} | '
f'train size = {len(train_data):,} | val size = {len(val_data):,} | test size = {len(test_data):,}'
)
# Initialize scaler and scale training targets by subtracting mean and dividing standard deviation (regression only)
if args.dataset_type == 'regression':
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
# Get loss and metric functions
loss_func = neg_log_like
metric_func = get_metric_func(metric=args.metric)
# Set up test set evaluation
test_smiles, test_targets = test_data.smiles(), test_data.targets()
sum_test_preds = np.zeros((len(test_smiles), args.num_tasks))
# Automatically determine whether to cache
if len(data) <= args.cache_cutoff:
cache = True
num_workers = 0
else:
cache = False
num_workers = args.num_workers
# Create data loaders
train_data_loader = MoleculeDataLoader(dataset=train_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache,
class_balance=args.class_balance,
shuffle=True,
seed=args.seed)
val_data_loader = MoleculeDataLoader(dataset=val_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache)
test_data_loader = MoleculeDataLoader(dataset=test_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache)
###########################################
########## Outer loop over ensemble members
###########################################
for model_idx in range(args.ensemble_start_idx,
args.ensemble_start_idx + args.ensemble_size):
# Set pytorch seed for random initial weights
torch.manual_seed(args.pytorch_seeds[model_idx])
######## set up all logging ########
# make save_dir
save_dir = os.path.join(args.save_dir, f'model_{model_idx}')
makedirs(save_dir)
# make results_dir
results_dir = os.path.join(args.results_dir, f'model_{model_idx}')
makedirs(results_dir)
# initialise wandb
os.environ['WANDB_MODE'] = 'dryrun'
wandb.init(name=args.wandb_name + '_' + str(model_idx),
project=args.wandb_proj,
reinit=True)
print('WANDB directory is:')
print(wandb.run.dir)
####################################
# Load/build model
if args.checkpoint_path is not None:
debug(f'Loading model {model_idx} from {args.checkpoint_path}')
model = load_checkpoint(args.checkpoint_path +
f'/model_{model_idx}/model.pt',
device=args.device,
logger=logger)
else:
debug(f'Building model {model_idx}')
model = MoleculeModel(args)
debug(model)
debug(f'Number of parameters = {param_count(model):,}')
if args.cuda:
debug('Moving model to cuda')
model = model.to(args.device)
# 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)
# Optimizer
optimizer = Adam([{
'params': model.encoder.parameters()
}, {
'params': model.ffn.parameters()
}, {
'params': model.log_noise,
'weight_decay': 0
}],
lr=args.init_lr,
weight_decay=args.weight_decay)
# Learning rate scheduler
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 range(args.epochs):
debug(f'Epoch {epoch}')
n_iter = train(model=model,
data_loader=train_data_loader,
loss_func=loss_func,
optimizer=optimizer,
scheduler=scheduler,
args=args,
n_iter=n_iter,
logger=logger)
val_scores = evaluate(model=model,
data_loader=val_data_loader,
args=args,
num_tasks=args.num_tasks,
metric_func=metric_func,
dataset_type=args.dataset_type,
scaler=scaler,
logger=logger)
# Average validation score
avg_val_score = np.nanmean(val_scores)
debug(f'Validation {args.metric} = {avg_val_score:.6f}')
wandb.log({"Validation MAE": avg_val_score})
# Save model checkpoint if improved validation score
if args.minimize_score and avg_val_score < best_score or \
not args.minimize_score and avg_val_score > best_score:
best_score, best_epoch = avg_val_score, epoch
save_checkpoint(os.path.join(save_dir, 'model.pt'), model,
scaler, features_scaler, args)
if epoch == args.noam_epochs - 1:
optimizer = Adam([{
'params': model.encoder.parameters()
}, {
'params': model.ffn.parameters()
}, {
'params': model.log_noise,
'weight_decay': 0
}],
lr=args.final_lr,
weight_decay=args.weight_decay)
scheduler = scheduler_const([args.final_lr])
# load 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'),
device=args.device,
logger=logger)
# SWAG training loop, returns swag_model
if args.swag:
model = train_swag(model, train_data, val_data, num_workers, cache,
loss_func, metric_func, scaler, features_scaler,
args, save_dir)
# SGLD loop, which saves nets
if args.sgld:
model = train_sgld(model, train_data, val_data, num_workers, cache,
loss_func, metric_func, scaler, features_scaler,
args, save_dir)
# GP loop
if args.gp:
model, likelihood = train_gp(model, train_data, val_data,
num_workers, cache, metric_func,
scaler, features_scaler, args,
save_dir)
# BBP
if args.bbp:
model = train_bbp(model, train_data, val_data, num_workers, cache,
loss_func, metric_func, scaler, features_scaler,
args, save_dir)
# DUN
if args.dun:
model = train_dun(model, train_data, val_data, num_workers, cache,
loss_func, metric_func, scaler, features_scaler,
args, save_dir)
##################################
########## Inner loop over samples
##################################
for sample_idx in range(args.samples):
# draw model from SWAG posterior
if args.swag:
model.sample(scale=1.0, cov=args.cov_mat, block=args.block)
# draw model from collected SGLD models
if args.sgld:
model = load_checkpoint(os.path.join(save_dir,
f'model_{sample_idx}.pt'),
device=args.device,
logger=logger)
# make predictions
test_preds = predict(model=model,
data_loader=test_data_loader,
args=args,
scaler=scaler,
test_data=True,
bbp_sample=True)
#######################################################################
#######################################################################
##### SAVING STUFF DOWN
if args.gp:
# get test_preds_std (scaled back to original data)
test_preds_std = predict_std_gp(model=model,
data_loader=test_data_loader,
args=args,
scaler=scaler,
likelihood=likelihood)
# 1 - MEANS
np.savez(os.path.join(results_dir, f'preds_{sample_idx}'),
np.array(test_preds))
# 2 - STD, combined aleatoric and epistemic (we save down the stds, always)
np.savez(os.path.join(results_dir, f'predsSTDEV_{sample_idx}'),
np.array(test_preds_std))
else:
# save test_preds and aleatoric uncertainties
if args.dun:
log_cat = model.log_cat.detach().cpu().numpy()
cat = np.exp(log_cat) / np.sum(np.exp(log_cat))
np.savez(os.path.join(results_dir, f'cat_{sample_idx}'),
cat)
# samples from categorical dist and saves a depth MC sample
depth_sample = np.random.multinomial(1,
cat).nonzero()[0][0]
test_preds_MCdepth = predict_MCdepth(
model=model,
data_loader=test_data_loader,
args=args,
scaler=scaler,
d=depth_sample)
np.savez(
os.path.join(results_dir,
f'predsMCDEPTH_{sample_idx}'),
np.array(test_preds_MCdepth))
if args.swag:
log_noise = model.base.log_noise
else:
log_noise = model.log_noise
noise = np.exp(log_noise.detach().cpu().numpy()) * np.array(
scaler.stds)
np.savez(os.path.join(results_dir, f'preds_{sample_idx}'),
np.array(test_preds))
np.savez(os.path.join(results_dir, f'noise_{sample_idx}'),
noise)
#######################################################################
#######################################################################
# add predictions to sum_test_preds
if len(test_preds) != 0:
sum_test_preds += np.array(test_preds)
# evaluate predictions using metric function
test_scores = evaluate_predictions(preds=test_preds,
targets=test_targets,
num_tasks=args.num_tasks,
metric_func=metric_func,
dataset_type=args.dataset_type,
logger=logger)
# compute average test score
avg_test_score = np.nanmean(test_scores)
info(
f'Model {model_idx}, sample {sample_idx} test {args.metric} = {avg_test_score:.6f}'
)
#################################
########## Bayesian Model Average
#################################
# note: this is an average over Bayesian samples AND components in an ensemble
# compute number of prediction iterations
pred_iterations = args.ensemble_size * args.samples
# average predictions across iterations
avg_test_preds = (sum_test_preds / pred_iterations).tolist()
# evaluate
BMA_scores = evaluate_predictions(preds=avg_test_preds,
targets=test_targets,
num_tasks=args.num_tasks,
metric_func=metric_func,
dataset_type=args.dataset_type,
logger=logger)
# average scores across tasks
avg_BMA_test_score = np.nanmean(BMA_scores)
info(f'BMA test {args.metric} = {avg_BMA_test_score:.6f}')
return BMA_scores
def pdts(args: TrainArgs, model_idx):
"""
preliminary experiment with PDTS (approximate BO)
we use a data set size of 50k and run until we have trained with 15k data points
our batch size is 50
we initialise with 1000 data points
"""
######## set up all logging ########
logger = None
# make save_dir
save_dir = os.path.join(args.save_dir, f'model_{model_idx}')
makedirs(save_dir)
# make results_dir
results_dir = args.results_dir
makedirs(results_dir)
# initialise wandb
#os.environ['WANDB_MODE'] = 'dryrun'
wandb.init(name=args.wandb_name + '_' + str(model_idx),
project=args.wandb_proj,
reinit=True)
#print('WANDB directory is:')
#print(wandb.run.dir)
####################################
########## get data
args.task_names = args.target_columns or get_task_names(args.data_path)
data = get_data(path=args.data_path, args=args, logger=logger)
args.num_tasks = data.num_tasks()
args.features_size = data.features_size()
########## SMILES of top 1%
top1p = np.array(MoleculeDataset(data).targets())
top1p_idx = np.argsort(-top1p[:, 0])[:int(args.max_data_size * 0.01)]
SMILES = np.array(MoleculeDataset(data).smiles())[top1p_idx]
########## initial data splits
args.seed = args.data_seeds[model_idx]
data.shuffle(seed=args.seed)
sizes = args.split_sizes
train_size = int(sizes[0] * len(data))
train_orig = data[:train_size]
test_orig = data[train_size:]
train_data, test_data = copy.deepcopy(
MoleculeDataset(train_orig)), copy.deepcopy(MoleculeDataset(test_orig))
args.train_data_size = len(train_data)
########## standardising
# features (train and test)
features_scaler = train_data.normalize_features(replace_nan_token=0)
test_data.normalize_features(features_scaler)
# targets (train)
train_targets = train_data.targets()
test_targets = test_data.targets()
scaler = StandardScaler().fit(train_targets)
scaled_targets = scaler.transform(train_targets).tolist()
train_data.set_targets(scaled_targets)
########## loss, metric functions
loss_func = neg_log_like
metric_func = get_metric_func(metric=args.metric)
########## data loaders
if len(data) <= args.cache_cutoff:
cache = True
num_workers = 0
else:
cache = False
num_workers = args.num_workers
train_data_loader = MoleculeDataLoader(dataset=train_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache,
class_balance=args.class_balance,
shuffle=True,
seed=args.seed)
test_data_loader = MoleculeDataLoader(dataset=test_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache)
########## instantiating model, optimiser, scheduler (MAP)
# set pytorch seed for random initial weights
torch.manual_seed(args.pytorch_seeds[model_idx])
# build model
print(f'Building model {model_idx}')
model = MoleculeModel(args)
print(model)
print(f'Number of parameters = {param_count(model):,}')
if args.cuda:
print('Moving model to cuda')
model = model.to(args.device)
# optimizer
optimizer = Adam([{
'params': model.encoder.parameters()
}, {
'params': model.ffn.parameters()
}, {
'params': model.log_noise,
'weight_decay': 0
}],
lr=args.lr,
weight_decay=args.weight_decay)
# learning rate scheduler
scheduler = scheduler_const([args.lr])
####################################################################
####################################################################
# FIRST THOMPSON ITERATION
### scores array
ptds_scores = np.ones(args.pdts_batches + 1)
batch_no = 0
### fill for batch 0
SMILES_train = np.array(train_data.smiles())
SMILES_stack = np.hstack((SMILES, SMILES_train))
overlap = len(SMILES_stack) - len(np.unique(SMILES_stack))
prop = overlap / len(SMILES)
ptds_scores[batch_no] = prop
wandb.log({
"Proportion of top 1%": prop,
"batch_no": batch_no
},
commit=False)
### train MAP posterior
gp_switch = False
likelihood = None
bbp_switch = None
n_iter = 0
for epoch in range(args.epochs_init_map):
n_iter = train(model=model,
data_loader=train_data_loader,
loss_func=loss_func,
optimizer=optimizer,
scheduler=scheduler,
args=args,
n_iter=n_iter,
bbp_switch=bbp_switch)
# save to save_dir
#if epoch == args.epochs_init_map - 1:
#save_checkpoint(os.path.join(save_dir, f'model_{batch_no}.pt'), model, scaler, features_scaler, args)
# if X load from checkpoint path
if args.bbp or args.gp or args.swag or args.sgld:
model = load_checkpoint(args.checkpoint_path +
f'/model_{model_idx}/model_{batch_no}.pt',
device=args.device,
logger=None)
########## BBP
if args.bbp:
model_bbp = MoleculeModelBBP(
args) # instantiate with bayesian linear layers
for (_, param_bbp), (_, param_pre) in zip(model_bbp.named_parameters(),
model.named_parameters()):
param_bbp.data = copy.deepcopy(
param_pre.data.T) # copy over parameters
# instantiate rhos
for layer in model_bbp.children():
if isinstance(layer, BayesLinear):
layer.init_rho(args.rho_min_bbp, args.rho_max_bbp)
for layer in model_bbp.encoder.encoder.children():
if isinstance(layer, BayesLinear):
layer.init_rho(args.rho_min_bbp, args.rho_max_bbp)
model = model_bbp # name back
# move to cuda
if args.cuda:
print('Moving bbp model to cuda')
model = model.to(args.device)
# optimiser and scheduler
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
scheduler = scheduler_const([args.lr])
bbp_switch = 2
n_iter = 0
for epoch in range(args.epochs_init):
n_iter = train(model=model,
data_loader=train_data_loader,
loss_func=loss_func,
optimizer=optimizer,
scheduler=scheduler,
args=args,
n_iter=n_iter,
bbp_switch=bbp_switch)
########## GP
if args.gp:
# feature_extractor
model.featurizer = True
feature_extractor = model
# inducing points
inducing_points = initial_inducing_points(train_data_loader,
feature_extractor, args)
# GP layer
gp_layer = GPLayer(inducing_points, args.num_tasks)
# full DKL model
model = copy.deepcopy(DKLMoleculeModel(feature_extractor, gp_layer))
# likelihood (rank 0 restricts to diagonal matrix)
likelihood = gpytorch.likelihoods.MultitaskGaussianLikelihood(
num_tasks=12, rank=0)
# model and likelihood to CUDA
if args.cuda:
model.cuda()
likelihood.cuda()
# loss object
loss_func = gpytorch.mlls.VariationalELBO(
likelihood, model.gp_layer, num_data=args.train_data_size)
# optimiser and scheduler
params_list = [
{
'params': model.feature_extractor.parameters(),
'weight_decay': args.weight_decay_gp
},
{
'params': model.gp_layer.hyperparameters()
},
{
'params': model.gp_layer.variational_parameters()
},
{
'params': likelihood.parameters()
},
]
optimizer = torch.optim.Adam(params_list, lr=args.lr)
scheduler = scheduler_const([args.lr])
gp_switch = True
n_iter = 0
for epoch in range(args.epochs_init):
n_iter = train(model=model,
data_loader=train_data_loader,
loss_func=loss_func,
optimizer=optimizer,
scheduler=scheduler,
args=args,
n_iter=n_iter,
gp_switch=gp_switch,
likelihood=likelihood)
########## SWAG
if args.swag:
model_core = copy.deepcopy(model)
model = train_swag_pdts(model_core, train_data_loader, loss_func,
scaler, features_scaler, args, save_dir,
batch_no)
########## SGLD
if args.sgld:
model = train_sgld_pdts(model, train_data_loader, loss_func, scaler,
features_scaler, args, save_dir, batch_no)
### find top_idx
top_idx = [] # need for thom
sum_test_preds = np.zeros(
(len(test_orig), args.num_tasks)) # need for greedy
for sample in range(args.samples):
# draw model from SWAG posterior
if args.swag:
model.sample(scale=1.0, cov=args.cov_mat, block=args.block)
# retrieve sgld sample
if args.sgld:
model = load_checkpoint(
args.save_dir +
f'/model_{model_idx}/model_{batch_no}/model_{sample}.pt',
device=args.device,
logger=logger)
test_preds = predict(model=model,
data_loader=test_data_loader,
args=args,
scaler=scaler,
test_data=True,
gp_sample=args.thompson,
bbp_sample=True)
test_preds = np.array(test_preds)
# thompson bit
rank = 0
# base length
if args.sgld:
base_length = 5 * sample + 4
else:
base_length = sample
while args.thompson and (len(top_idx) <= base_length):
top_unique_molecule = np.argsort(-test_preds[:, 0])[rank]
rank += 1
if top_unique_molecule not in top_idx:
top_idx.append(top_unique_molecule)
# add to sum_test_preds
sum_test_preds += test_preds
# print
print('done sample ' + str(sample))
# final top_idx
if args.thompson:
top_idx = np.array(top_idx)
else:
sum_test_preds /= args.samples
top_idx = np.argsort(-sum_test_preds[:, 0])[:50]
### transfer from test to train
top_idx = -np.sort(-top_idx)
for idx in top_idx:
train_orig.append(test_orig.pop(idx))
train_data, test_data = copy.deepcopy(
MoleculeDataset(train_orig)), copy.deepcopy(MoleculeDataset(test_orig))
args.train_data_size = len(train_data)
if args.gp:
loss_func = gpytorch.mlls.VariationalELBO(
likelihood, model.gp_layer, num_data=args.train_data_size)
print(args.train_data_size)
### standardise features (train and test; using original features_scaler)
train_data.normalize_features(features_scaler)
test_data.normalize_features(features_scaler)
### standardise targets (train only; using original scaler)
train_targets = train_data.targets()
scaled_targets_tr = scaler.transform(train_targets).tolist()
train_data.set_targets(scaled_targets_tr)
### create data loaders
train_data_loader = MoleculeDataLoader(dataset=train_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache,
class_balance=args.class_balance,
shuffle=True,
seed=args.seed)
test_data_loader = MoleculeDataLoader(dataset=test_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache)
####################################################################
####################################################################
##################################
########## thompson sampling loop
##################################
for batch_no in range(1, args.pdts_batches + 1):
### fill in ptds_scores
SMILES_train = np.array(train_data.smiles())
SMILES_stack = np.hstack((SMILES, SMILES_train))
overlap = len(SMILES_stack) - len(np.unique(SMILES_stack))
prop = overlap / len(SMILES)
ptds_scores[batch_no] = prop
wandb.log({
"Proportion of top 1%": prop,
"batch_no": batch_no
},
commit=False)
### train posterior
n_iter = 0
for epoch in range(args.epochs):
n_iter = train(model=model,
data_loader=train_data_loader,
loss_func=loss_func,
optimizer=optimizer,
scheduler=scheduler,
args=args,
n_iter=n_iter,
gp_switch=gp_switch,
likelihood=likelihood,
bbp_switch=bbp_switch)
# save to save_dir
#if epoch == args.epochs - 1:
#save_checkpoint(os.path.join(save_dir, f'model_{batch_no}.pt'), model, scaler, features_scaler, args)
# if swag, load checkpoint
if args.swag:
model_core = load_checkpoint(
args.checkpoint_path +
f'/model_{model_idx}/model_{batch_no}.pt',
device=args.device,
logger=None)
########## SWAG
if args.swag:
model = train_swag_pdts(model_core, train_data_loader, loss_func,
scaler, features_scaler, args, save_dir,
batch_no)
########## SGLD
if args.sgld:
model = train_sgld_pdts(model, train_data_loader, loss_func,
scaler, features_scaler, args, save_dir,
batch_no)
### find top_idx
top_idx = [] # need for thom
sum_test_preds = np.zeros(
(len(test_orig), args.num_tasks)) # need for greedy
for sample in range(args.samples):
# draw model from SWAG posterior
if args.swag:
model.sample(scale=1.0, cov=args.cov_mat, block=args.block)
# retrieve sgld sample
if args.sgld:
model = load_checkpoint(
args.save_dir +
f'/model_{model_idx}/model_{batch_no}/model_{sample}.pt',
device=args.device,
logger=logger)
test_preds = predict(model=model,
data_loader=test_data_loader,
args=args,
scaler=scaler,
test_data=True,
gp_sample=args.thompson,
bbp_sample=True)
test_preds = np.array(test_preds)
# thompson bit
rank = 0
# base length
if args.sgld:
base_length = 5 * sample + 4
else:
base_length = sample
while args.thompson and (len(top_idx) <= base_length):
top_unique_molecule = np.argsort(-test_preds[:, 0])[rank]
rank += 1
if top_unique_molecule not in top_idx:
top_idx.append(top_unique_molecule)
# add to sum_test_preds
sum_test_preds += test_preds
# print
print('done sample ' + str(sample))
# final top_idx
if args.thompson:
top_idx = np.array(top_idx)
else:
sum_test_preds /= args.samples
top_idx = np.argsort(-sum_test_preds[:, 0])[:50]
### transfer from test to train
top_idx = -np.sort(-top_idx)
for idx in top_idx:
train_orig.append(test_orig.pop(idx))
train_data, test_data = copy.deepcopy(
MoleculeDataset(train_orig)), copy.deepcopy(
MoleculeDataset(test_orig))
args.train_data_size = len(train_data)
if args.gp:
loss_func = gpytorch.mlls.VariationalELBO(
likelihood, model.gp_layer, num_data=args.train_data_size)
print(args.train_data_size)
### standardise features (train and test; using original features_scaler)
train_data.normalize_features(features_scaler)
test_data.normalize_features(features_scaler)
### standardise targets (train only; using original scaler)
train_targets = train_data.targets()
scaled_targets_tr = scaler.transform(train_targets).tolist()
train_data.set_targets(scaled_targets_tr)
### create data loaders
train_data_loader = MoleculeDataLoader(
dataset=train_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache,
class_balance=args.class_balance,
shuffle=True,
seed=args.seed)
test_data_loader = MoleculeDataLoader(dataset=test_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache)
# save scores
np.savez(os.path.join(results_dir, f'ptds_{model_idx}'), ptds_scores)
def run_training_gnn_xgb(args: TrainArgs,
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
# Print command line
debug('Command line')
debug(f'python {" ".join(sys.argv)}')
# Print args
debug('Args')
debug(args)
# Save args
args.save(os.path.join(args.save_dir, 'args.json'))
# Set pytorch seed for random initial weights
torch.manual_seed(args.pytorch_seed)
# Get data
debug('Loading data')
args.task_names = args.target_columns or get_task_names(args.data_path)
data = get_data(path=args.data_path, args=args, logger=logger)
args.num_tasks = data.num_tasks()
args.features_size = data.features_size()
debug(f'Number of tasks = {args.num_tasks}')
# Split data
debug(f'Splitting data with seed {args.seed}')
if args.separate_test_path:
test_data = get_data(path=args.separate_test_path,
args=args,
features_path=args.separate_test_features_path,
logger=logger)
if args.separate_val_path:
val_data = get_data(path=args.separate_val_path,
args=args,
features_path=args.separate_val_features_path,
logger=logger)
if args.separate_val_path and args.separate_test_path:
train_data = data
elif args.separate_val_path:
train_data, _, test_data = split_data(data=data,
split_type=args.split_type,
sizes=(0.8, 0.0, 0.2),
seed=args.seed,
args=args,
logger=logger)
elif args.separate_test_path:
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)
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.save_smiles_splits:
save_smiles_splits(train_data=train_data,
val_data=val_data,
test_data=test_data,
data_path=args.data_path,
save_dir=args.save_dir)
if args.features_scaling:
features_scaler = train_data.normalize_features(replace_nan_token=0)
val_data.normalize_features(features_scaler)
test_data.normalize_features(features_scaler)
else:
features_scaler = None
args.train_data_size = len(train_data)
debug(
f'Total size = {len(data):,} | '
f'train size = {len(train_data):,} | val size = {len(val_data):,} | test size = {len(test_data):,}'
)
# Initialize scaler and scale training targets by subtracting mean and dividing standard deviation (regression only)
if args.dataset_type == 'regression':
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
# Get loss and metric functions
loss_func = get_loss_func(args)
metric_func = get_metric_func(metric=args.metric)
# Set up test set evaluation
val_smiles, val_targets = val_data.smiles(), val_data.targets()
test_smiles, test_targets = test_data.smiles(), test_data.targets()
if args.dataset_type == 'multiclass':
sum_test_preds = np.zeros(
(len(test_smiles), args.num_tasks, args.multiclass_num_classes))
else:
sum_test_preds = np.zeros((len(test_smiles), args.num_tasks))
# Automatically determine whether to cache
if len(data) <= args.cache_cutoff:
cache = True
num_workers = 0
else:
cache = False
num_workers = args.num_workers
# Create data loaders
train_data_loader = MoleculeDataLoader(dataset=train_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache,
class_balance=args.class_balance,
shuffle=True,
seed=args.seed)
val_data_loader = MoleculeDataLoader(dataset=val_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache)
test_data_loader = MoleculeDataLoader(dataset=test_data,
batch_size=args.batch_size,
num_workers=num_workers,
cache=cache)
# 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}')
makedirs(save_dir)
try:
writer = SummaryWriter(log_dir=save_dir)
except:
writer = SummaryWriter(logdir=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],
logger=logger)
else:
debug(f'Building model {model_idx}')
model = MoleculeModel(args)
debug(model)
debug(f'Number of parameters = {param_count(model):,}')
if args.cuda:
debug('Moving model to cuda')
model = model.to(args.device)
# 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)
# 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}')
n_iter = train(model=model,
data_loader=train_data_loader,
loss_func=loss_func,
optimizer=optimizer,
scheduler=scheduler,
args=args,
n_iter=n_iter,
logger=logger,
writer=writer)
if isinstance(scheduler, ExponentialLR):
scheduler.step()
val_scores = evaluate(model=model,
data_loader=val_data_loader,
num_tasks=args.num_tasks,
metric_func=metric_func,
dataset_type=args.dataset_type,
scaler=scaler,
logger=logger)
# 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):
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
if args.minimize_score and avg_val_score < best_score or \
not args.minimize_score and avg_val_score > best_score:
best_score, best_epoch = avg_val_score, epoch
save_checkpoint(os.path.join(save_dir, 'model.pt'), model,
scaler, features_scaler, args)
# 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'),
device=args.device,
logger=logger)
test_preds, _ = predict(model=model,
data_loader=test_data_loader,
scaler=scaler)
test_scores = evaluate_predictions(preds=test_preds,
targets=test_targets,
num_tasks=args.num_tasks,
metric_func=metric_func,
dataset_type=args.dataset_type,
logger=logger)
if len(test_preds) != 0:
sum_test_preds += np.array(test_preds)
# 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):
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)
writer.close()
# Evaluate ensemble on test set
avg_test_preds = (sum_test_preds / args.ensemble_size).tolist()
ensemble_scores = evaluate_predictions(preds=avg_test_preds,
targets=test_targets,
num_tasks=args.num_tasks,
metric_func=metric_func,
dataset_type=args.dataset_type,
logger=logger)
# Average ensemble score
avg_ensemble_test_score = np.nanmean(ensemble_scores)
info(f'Ensemble test {args.metric} = {avg_ensemble_test_score:.6f}')
# 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}'
)
_, train_feature = predict(model=model,
data_loader=train_data_loader,
scaler=scaler)
_, val_feature = predict(model=model,
data_loader=val_data_loader,
scaler=scaler)
_, test_feature = predict(model=model,
data_loader=test_data_loader,
scaler=scaler)
return ensemble_scores, train_feature, val_feature, test_feature, train_targets, val_targets, test_targets
def cross_validate(args: TrainArgs,
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 = args.target_columns or get_task_names(args.data_path)
# Run training on different random seeds for each fold
dmpnn_scores = []
dmpnn_xgb_scores = []
morgan_scores = []
dmpnn_morgan_scores = []
for fold_num in range(args.num_folds):
if args.dataset_type == 'classification':
args.data_path = 'molnet_benchmark/molnet_random_' + args.protein + '_c/seed' + str(
fold_num + 1) + '/train.csv'
args.separate_test_path = 'molnet_benchmark/molnet_random_' + args.protein + '_c/seed' + str(
fold_num + 1) + '/val.csv'
args.separate_val_path = 'molnet_benchmark/molnet_random_' + args.protein + '_c/seed' + str(
fold_num + 1) + '/test.csv'
elif args.dataset_type == 'regression':
args.data_path = 'molnet_benchmark/molnet_random_' + args.protein + '_r/seed' + str(
fold_num + 1) + '/train.csv'
args.separate_test_path = 'molnet_benchmark/molnet_random_' + args.protein + '_r/seed' + str(
fold_num + 1) + '/val.csv'
args.separate_val_path = 'molnet_benchmark/molnet_random_' + args.protein + '_r/seed' + str(
fold_num + 1) + '/test.csv'
info(f'Fold {fold_num}')
args.seed = init_seed + fold_num
args.save_dir = os.path.join(save_dir, f'fold_{fold_num}')
makedirs(args.save_dir)
model_scores, model, scaler = run_training(args, logger)
dmpnn_scores.append(model_scores)
train_target, train_feature, val_target, val_feature, test_target, test_feature, train_smiles, val_smiles, test_smiles = get_xgboost_feature(
args, logger, model)
train_target = pd.DataFrame(train_target)
train_feature = pd.DataFrame(train_feature)
val_target = pd.DataFrame(val_target)
val_feature = pd.DataFrame(val_feature)
test_target = pd.DataFrame(test_target)
test_feature = pd.DataFrame(test_feature)
train_morgan_feature = get_morgan_feature(train_smiles)
val_morgan_feature = get_morgan_feature(val_smiles)
test_morgan_feature = get_morgan_feature(test_smiles)
if args.dataset_type == 'classification':
if test_target.shape[1] == 1:
xgb_gbc = xgb.XGBClassifier(base_score=0.5,
booster='gbtree',
colsample_bylevel=1,
colsample_bytree=1,
gamma=1,
learning_rate=0.1,
max_delta_step=0,
max_depth=4,
min_child_weight=8,
missing=None,
n_estimators=2000,
n_jobs=1,
nthread=None,
objective='binary:logistic',
random_state=0,
reg_alpha=0,
reg_lambda=1,
scale_pos_weight=1,
seed=None,
silent=True,
subsample=0.8,
tree_method='gpu_hist',
n_gpus=-1)
xgb_gbc.fit(train_feature,
train_target,
eval_set=[(val_feature, val_target)],
eval_metric='auc',
early_stopping_rounds=200)
pre_pro = xgb_gbc.predict_proba(test_feature)[:, 1]
fpr, tpr, threshold = roc_curve(test_target, pre_pro)
AUC = auc(fpr, tpr)
pre_pro = [1 if i > 0.5 else 0 for i in pre_pro]
tn, fp, fn, tp = confusion_matrix(test_target, pre_pro).ravel()
# Sn = TP /（TP + FN） Sp = TN / (TN+FP)
Sn = tp / (tp + fn)
Sp = tn / (tn + fp)
acc = accuracy_score(test_target, pre_pro)
dmpnn_xgb_scores.append([AUC, Sn, Sp, acc])
joblib.dump(xgb_gbc, 'external_test/dmpnn_xgb.model')
xgb_gbc = xgb.XGBClassifier(base_score=0.5,
booster='gbtree',
colsample_bylevel=1,
colsample_bytree=1,
gamma=1,
learning_rate=0.1,
max_delta_step=0,
max_depth=4,
min_child_weight=8,
missing=None,
n_estimators=2000,
n_jobs=1,
nthread=None,
objective='binary:logistic',
random_state=0,
reg_alpha=0,
reg_lambda=1,
scale_pos_weight=1,
seed=None,
silent=True,
subsample=0.8,
tree_method='gpu_hist',
n_gpus=-1)
xgb_gbc.fit(train_morgan_feature,
train_target,
eval_set=[(val_morgan_feature, val_target)],
eval_metric='auc',
early_stopping_rounds=200)
pre_pro = xgb_gbc.predict_proba(test_morgan_feature)[:, 1]
fpr, tpr, threshold = roc_curve(test_target, pre_pro)
AUC = auc(fpr, tpr)
pre_pro = [1 if i > 0.5 else 0 for i in pre_pro]
tn, fp, fn, tp = confusion_matrix(test_target, pre_pro).ravel()
# Sn = TP /（TP + FN） Sp = TN / (TN+FP)
Sn = tp / (tp + fn)
Sp = tn / (tn + fp)
acc = accuracy_score(test_target, pre_pro)
morgan_scores.append([AUC, Sn, Sp, acc])
joblib.dump(xgb_gbc, 'external_test/morgan_xgb.model')
train_gcn_mor_feature = pd.concat(
[train_feature, train_morgan_feature], axis=1)
val_gcn_mor_feature = pd.concat(
[val_feature, val_morgan_feature], axis=1)
test_gcn_mor_feature = pd.concat(
[test_feature, test_morgan_feature], axis=1)
train_gcn_mor_feature.columns = val_gcn_mor_feature.columns = test_gcn_mor_feature.columns = range(
train_gcn_mor_feature.shape[1])
xgb_gbc = xgb.XGBClassifier(base_score=0.5,
booster='gbtree',
colsample_bylevel=1,
colsample_bytree=1,
gamma=1,
learning_rate=0.1,
max_delta_step=0,
max_depth=4,
min_child_weight=8,
missing=None,
n_estimators=2000,
n_jobs=1,
nthread=None,
objective='binary:logistic',
random_state=0,
reg_alpha=0,
reg_lambda=1,
scale_pos_weight=1,
seed=None,
silent=True,
subsample=0.8,
tree_method='gpu_hist',
n_gpus=-1)
xgb_gbc.fit(train_gcn_mor_feature,
train_target,
eval_set=[(val_gcn_mor_feature, val_target)],
eval_metric='auc',
early_stopping_rounds=200)
pre_pro = xgb_gbc.predict_proba(test_gcn_mor_feature)[:, 1]
fpr, tpr, threshold = roc_curve(test_target, pre_pro)
AUC = auc(fpr, tpr)
pre_pro = [1 if i > 0.5 else 0 for i in pre_pro]
tn, fp, fn, tp = confusion_matrix(test_target, pre_pro).ravel()
Sn = tp / (tp + fn)
Sp = tn / (tn + fp)
acc = accuracy_score(test_target, pre_pro)
dmpnn_morgan_scores.append([AUC, Sn, Sp, acc])
joblib.dump(xgb_gbc, 'external_test/dmpnn_morgan_xgb.model')
else:
aucs = []
for i in range(test_target.shape[1]):
xgb_gbc = xgb.XGBClassifier(base_score=0.5,
booster='gbtree',
colsample_bylevel=1,
colsample_bytree=1,
gamma=1,
learning_rate=0.1,
max_delta_step=0,
max_depth=4,
min_child_weight=8,
missing=None,
n_estimators=2000,
n_jobs=1,
nthread=None,
objective='binary:logistic',
random_state=0,
reg_alpha=0,
reg_lambda=1,
scale_pos_weight=1,
seed=None,
silent=True,
subsample=0.8,
tree_method='gpu_hist',
n_gpus=-1)
if max(val_target[i]) == 0 or max(
train_target[i]) == 0 or max(test_target[i]) == 0:
continue
xgb_gbc.fit(train_feature,
train_target[i],
eval_set=[(val_feature, val_target[i])],
eval_metric='auc',
early_stopping_rounds=100)
pre_pro = xgb_gbc.predict_proba(test_feature)[:, 1]
fpr, tpr, threshold = roc_curve(test_target[i], pre_pro)
AUC = auc(fpr, tpr)
if args.metric == "prc-auc":
precision, recall, _ = precision_recall_curve(
test_target[i], pre_pro)
AUC = auc(recall, precision)
pre_pro = [1 if i > 0.5 else 0 for i in pre_pro]
tn, fp, fn, tp = confusion_matrix(test_target,
pre_pro).ravel()
Sn = tp / (tp + fn)
Sp = tn / (tn + fp)
acc = accuracy_score(test_target, pre_pro)
aucs.append([AUC, Sn, Sp, acc])
dmpnn_xgb_scores.append([np.mean(aucs)])
elif args.dataset_type == 'regression':
if test_target.shape[1] == 1:
xgb_gbc = xgb.XGBRegressor(learn_rate=0.1,
max_depth=4,
min_child_weight=10,
gamma=1,
subsample=0.8,
colsample_bytree=0.8,
reg_alpha=0.8,
objective='reg:linear',
n_estimators=2000,
tree_method='gpu_hist',
n_gpus=-1)
xgb_gbc.fit(train_feature,
train_target,
eval_set=[(val_feature, val_target)],
eval_metric='rmse',
early_stopping_rounds=200)
y_pred = xgb_gbc.predict(test_feature)
y_pred = scaler.inverse_transform(y_pred)
y_test = test_target.astype('float')
MSE = mean_squared_error(y_test, y_pred)
RMSE = MSE**0.5
MAE = median_absolute_error(y_test, y_pred)
dmpnn_xgb_scores.append([RMSE, MAE])
joblib.dump(xgb_gbc, 'external_test/dmpnn_xgb.model')
xgb_gbc = xgb.XGBRegressor(learn_rate=0.1,
max_depth=4,
min_child_weight=10,
gamma=1,
subsample=0.8,
colsample_bytree=0.8,
reg_alpha=0.8,
objective='reg:linear',
n_estimators=2000,
tree_method='gpu_hist',
n_gpus=-1)
xgb_gbc.fit(train_morgan_feature,
train_target,
eval_set=[(val_morgan_feature, val_target)],
eval_metric='rmse',
early_stopping_rounds=200)
y_pred = xgb_gbc.predict(test_morgan_feature)
y_pred = scaler.inverse_transform(y_pred)
MSE = mean_squared_error(y_test, y_pred)
RMSE = MSE**0.5
MAE = median_absolute_error(y_test, y_pred)
morgan_scores.append([RMSE, MAE])
joblib.dump(xgb_gbc, 'external_test/morgan_xgb.model')
train_gcn_mor_feature = pd.concat(
[train_feature, train_morgan_feature], axis=1)
val_gcn_mor_feature = pd.concat(
[val_feature, val_morgan_feature], axis=1)
test_gcn_mor_feature = pd.concat(
[test_feature, test_morgan_feature], axis=1)
train_gcn_mor_feature.columns = val_gcn_mor_feature.columns = test_gcn_mor_feature.columns = range(
train_gcn_mor_feature.shape[1])
xgb_gbc = xgb.XGBRegressor(learn_rate=0.1,
max_depth=4,
min_child_weight=10,
gamma=1,
subsample=0.8,
colsample_bytree=0.8,
reg_alpha=0.8,
objective='reg:linear',
n_estimators=2000,
tree_method='gpu_hist',
n_gpus=-1)
xgb_gbc.fit(train_gcn_mor_feature,
train_target,
eval_set=[(val_gcn_mor_feature, val_target)],
eval_metric='rmse',
early_stopping_rounds=200)
y_pred = xgb_gbc.predict(test_gcn_mor_feature)
y_pred = scaler.inverse_transform(y_pred)
MSE = mean_squared_error(y_test, y_pred)
RMSE = MSE**0.5
MAE = median_absolute_error(y_test, y_pred)
dmpnn_morgan_scores.append([RMSE, MAE])
joblib.dump(xgb_gbc, 'external_test/dmpnn_morgan_xgb.model')
else:
MAEs = []
for i in range(test_target.shape[1]):
xgb_gbc = xgb.XGBRegressor(learn_rate=0.1,
max_depth=4,
min_child_weight=10,
gamma=1,
subsample=0.8,
colsample_bytree=0.8,
reg_alpha=0.8,
objective='reg:linear',
n_estimators=2000,
tree_method='gpu_hist',
n_gpus=-1)
xgb_gbc.fit(train_feature,
train_target[i],
eval_set=[(val_feature, val_target[i])],
eval_metric='rmse',
early_stopping_rounds=200)
y_pred = xgb_gbc.predict(test_feature)
y_test = test_target[i].astype('float')
MSE = mean_squared_error(y_test, y_pred)
RMSE = MSE**0.5
MAE = median_absolute_error(y_test, y_pred)
MAEs.append([MAE, RMSE])
dmpnn_xgb_scores.append([np.mean(MAEs)])
dmpnn_scores = np.array(dmpnn_scores)
# Report scores across models
dmpnn_scores = np.nanmean(
dmpnn_scores, axis=1) # average score for each model across tasks
dmpnn_mean_score, dmpnn_std_score = np.nanmean(dmpnn_scores), np.nanstd(
dmpnn_scores)
print('three dmpnn test = ', dmpnn_scores)
info(
f'Overall dmpnn test {args.metric} = {dmpnn_mean_score:.6f} +/- {dmpnn_std_score:.6f}'
)
dmpnn_xgb_scores = np.nanmean(
dmpnn_xgb_scores, axis=1) # average score for each model across tasks
dmpnn_xgb_mean_score, dmpnn_xgb_std_score = np.nanmean(
dmpnn_xgb_scores), np.nanstd(dmpnn_xgb_scores)
print('three dmpnn_xgb_test = ', dmpnn_xgb_scores)
info(
f'Overall dmpnn_xgb_test {args.metric} = {dmpnn_xgb_mean_score:.6f} +/- {dmpnn_xgb_std_score:.6f}'
)
morgan_scores = np.nanmean(
morgan_scores, axis=1) # average score for each model across tasks
morgan_mean_score, morgan_std_score = np.nanmean(morgan_scores), np.nanstd(
morgan_scores)
print('three morgen_test = ', morgan_scores)
info(
f'Overall morgen_test {args.metric} = {morgan_mean_score:.6f} +/- {morgan_std_score:.6f}'
)
dmpnn_morgan_scores = np.nanmean(
dmpnn_morgan_scores,
axis=1) # average score for each model across tasks
dmpnn_morgan_mean_score, dmpnn_morgan_std_score = np.nanmean(
dmpnn_morgan_scores), np.nanstd(dmpnn_morgan_scores)
print('three dmpnn_morgan_scores = ', dmpnn_morgan_scores)
info(
f'Overall dmpnn_morgen_test {args.metric} = {dmpnn_morgan_mean_score:.6f} +/- {dmpnn_morgan_std_score:.6f}'
)
return model
