def run_split_data(args: Args):
# Load raw data
with open(args.data_path) as f:
reader = csv.reader(f)
header = next(reader)
lines = list(reader)
# Load SMILES
smiles = get_smiles(path=args.data_path, smiles_column=args.smiles_column)
# Make sure lines and smiles line up
assert len(lines) == len(smiles)
assert all(smile in line for smile, line in zip(smiles, lines))
# Create data
data = []
for smile, line in tqdm(zip(smiles, lines), total=len(smiles)):
datapoint = MoleculeDatapoint(smiles=smile)
datapoint.line = line
data.append(datapoint)
data = MoleculeDataset(data)
train, val, test = split_data(data=data,
split_type=args.split_type,
sizes=args.split_sizes,
seed=args.seed)
makedirs(args.save_dir)
for name, dataset in [('train', train), ('val', val), ('test', test)]:
with open(os.path.join(args.save_dir, f'{name}.csv'), 'w') as f:
writer = csv.writer(f)
writer.writerow(header)
for datapoint in dataset:
writer.writerow(datapoint.line)
def run_random_forest(args: Namespace, 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)
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)
debug(f'Total size = {len(data):,} | train size = {len(train_data):,} | test size = {len(test_data):,}')
debug('Computing morgan fingerprints')
for dataset in [train_data, test_data]:
for datapoint in tqdm(dataset, total=len(dataset)):
datapoint.set_features(morgan_fingerprint(smiles=datapoint.smiles, radius=args.radius, num_bits=args.num_bits))
debug('Training')
if args.single_task:
scores = single_task_random_forest(train_data, test_data, metric_func, args)
else:
scores = multi_task_random_forest(train_data, test_data, metric_func, args)
info(f'Test {args.metric} = {np.nanmean(scores)}')
return scores
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 run_split_data(data_path: str, split_type: str,
split_sizes: Tuple[int, int,
int], seed: int, save_dir: str):
with open(data_path) as f:
reader = csv.reader(f)
header = next(reader)
lines = list(reader)
data = []
for line in tqdm(lines):
datapoint = MoleculeDatapoint(line=line)
datapoint.line = line
data.append(datapoint)
data = MoleculeDataset(data)
train, dev, test = split_data(data=data,
split_type=split_type,
sizes=split_sizes,
seed=seed)
makedirs(save_dir)
for name, dataset in [('train', train), ('dev', dev), ('test', test)]:
with open(os.path.join(save_dir, f'{name}.csv'), 'w') as f:
writer = csv.writer(f)
writer.writerow(header)
for datapoint in dataset:
writer.writerow(datapoint.line)
def random_split(dataset, seed_val):
# get target names (assuming that 1st column contains molecule names and 2nd column contains smiles and rest of the columns are targets)
df = pd.read_csv(dataset, sep=",", index_col=None, dtype={'RTECS_ID': str})
cols = list(df.columns)
target_names = cols[2:]
mol_dataset = utils.get_data(dataset, use_compound_names=True)
train, valid, test = utils.split_data(mol_dataset,
sizes=(0.7, 0.1, 0.2),
seed=seed_val)
train_df = get_partition_as_df(train, target_names)
train_df = train_df[['RTECS_ID']]
valid_df = get_partition_as_df(valid, target_names)
valid_df = valid_df[['RTECS_ID']]
test_df = get_partition_as_df(test, target_names)
test_df = test_df[['RTECS_ID']]
return train_df, valid_df, test_df
def prepare_data(args):
data = get_data(path=args.data_path, args=args)
source_data = get_data(path=args.source_data_path, args=args)
# split train, val, test
train_data, val_data, test_data = split_data(data=data,
split_type=args.split_type,
sizes=args.split_sizes,
seed=args.seed,
args=args)
args.num_tasks = train_data.num_tasks()
args.features_size = train_data.features_size()
args.train_data_size = len(train_data)
print('source data:', len(source_data))
print('target data:', len(data))
return train_data, val_data, test_data, source_data
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 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 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_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 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 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
