def make_predictions(args: PredictArgs, smiles: List[List[str]] = None,
features: np.ndarray = None,
save_prediction=True) -> List[List[Optional[float]]]:
"""
Loads data and a trained model and uses the model to make predictions on the data.
If SMILES are provided, then makes predictions on smiles.
Otherwise makes predictions on :code:`args.test_data`.
:param args: A :class:`~chemprop.args.PredictArgs` object containing arguments for
loading data and a model and making predictions.
:param smiles: List of list of SMILES to make predictions on.
:return: A list of lists of target predictions.
"""
print('Loading training args')
train_args = load_args(args.checkpoint_paths[0])
num_tasks, task_names = train_args.num_tasks, train_args.task_names
update_prediction_args(predict_args=args, train_args=train_args)
args: Union[PredictArgs, TrainArgs]
if args.atom_descriptors == 'feature':
set_extra_atom_fdim(train_args.atom_features_size)
if args.bond_features_path is not None:
set_extra_bond_fdim(train_args.bond_features_size)
#set explicit H option and reaction option
set_explicit_h(train_args.explicit_h)
set_reaction(train_args.reaction, train_args.reaction_mode)
print('Loading data')
if smiles is not None:
full_data = get_data_from_smiles(
smiles=smiles,
features=features,
skip_invalid_smiles=False,
features_generator=args.features_generator
)
else:
full_data = get_data(path=args.test_path, smiles_columns=args.smiles_columns, target_columns=[], ignore_columns=[],
skip_invalid_smiles=False, args=args, store_row=not args.drop_extra_columns)
print('Validating SMILES')
full_to_valid_indices = {}
valid_index = 0
for full_index in range(len(full_data)):
if all(mol is not None for mol in full_data[full_index].mol):
full_to_valid_indices[full_index] = valid_index
valid_index += 1
test_data = MoleculeDataset([full_data[i] for i in sorted(full_to_valid_indices.keys())])
# Edge case if empty list of smiles is provided
if len(test_data) == 0:
return [None] * len(full_data)
print(f'Test size = {len(test_data):,}')
# Predict with each model individually and sum predictions
if args.dataset_type == 'multiclass':
sum_preds = np.zeros((len(test_data), num_tasks, args.multiclass_num_classes))
else:
sum_preds = np.zeros((len(test_data), num_tasks))
# Create data loader
test_data_loader = MoleculeDataLoader(
dataset=test_data,
batch_size=args.batch_size,
num_workers=args.num_workers
)
# Partial results for variance robust calculation.
if args.ensemble_variance:
all_preds = np.zeros((len(test_data), num_tasks, len(args.checkpoint_paths)))
print(f'Predicting with an ensemble of {len(args.checkpoint_paths)} models')
for index, checkpoint_path in enumerate(tqdm(args.checkpoint_paths, total=len(args.checkpoint_paths))):
# Load model and scalers
model = load_checkpoint(checkpoint_path, device=args.device)
scaler, features_scaler, atom_descriptor_scaler, bond_feature_scaler = load_scalers(checkpoint_path)
# Normalize features
if args.features_scaling or train_args.atom_descriptor_scaling or train_args.bond_feature_scaling:
test_data.reset_features_and_targets()
if args.features_scaling:
test_data.normalize_features(features_scaler)
if train_args.atom_descriptor_scaling and args.atom_descriptors is not None:
test_data.normalize_features(atom_descriptor_scaler, scale_atom_descriptors=True)
if train_args.bond_feature_scaling and args.bond_features_size > 0:
test_data.normalize_features(bond_feature_scaler, scale_bond_features=True)
# Make predictions
model_preds = predict(
model=model,
data_loader=test_data_loader,
scaler=scaler
)
sum_preds += np.array(model_preds)
if args.ensemble_variance:
all_preds[:, :, index] = model_preds
# Ensemble predictions
avg_preds = sum_preds / len(args.checkpoint_paths)
avg_preds = avg_preds.tolist()
if args.ensemble_variance:
all_epi_uncs = np.var(all_preds, axis=2)
all_epi_uncs = all_epi_uncs.tolist()
# Save predictions
if save_prediction:
print(f'Saving predictions to {args.preds_path}')
assert len(test_data) == len(avg_preds)
if args.ensemble_variance:
assert len(test_data) == len(all_epi_uncs)
makedirs(args.preds_path, isfile=True)
# Get prediction column names
if args.dataset_type == 'multiclass':
task_names = [f'{name}_class_{i}' for name in task_names for i in range(args.multiclass_num_classes)]
else:
task_names = task_names
# Copy predictions over to full_data
for full_index, datapoint in enumerate(full_data):
valid_index = full_to_valid_indices.get(full_index, None)
preds = avg_preds[valid_index] if valid_index is not None else ['Invalid SMILES'] * len(task_names)
if args.ensemble_variance:
epi_uncs = all_epi_uncs[valid_index] if valid_index is not None else ['Invalid SMILES'] * len(task_names)
# If extra columns have been dropped, add back in SMILES columns
if args.drop_extra_columns:
datapoint.row = OrderedDict()
smiles_columns = args.smiles_columns
for column, smiles in zip(smiles_columns, datapoint.smiles):
datapoint.row[column] = smiles
# Add predictions columns
if args.ensemble_variance:
for pred_name, pred, epi_unc in zip(task_names, preds, epi_uncs):
datapoint.row[pred_name] = pred
datapoint.row[pred_name+'_epi_unc'] = epi_unc
else:
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=full_data[0].row.keys())
writer.writeheader()
for datapoint in full_data:
writer.writerow(datapoint.row)
return avg_preds
def run_sklearn(args: SklearnTrainArgs,
data: MoleculeDataset,
logger: Logger = None) -> Dict[str, List[float]]:
"""
Loads data, trains a scikit-learn model, and returns test scores for the model checkpoint with the highest validation score.
:param args: A :class:`~chemprop.args.SklearnTrainArgs` object containing arguments for
loading data and training the scikit-learn model.
:param data: A :class:`~chemprop.data.MoleculeDataset` containing the data.
:param logger: A logger to record output.
:return: A dictionary mapping each metric in :code:`metrics` to a list of values for each task.
"""
if logger is not None:
debug, info = logger.debug, logger.info
else:
debug = info = print
debug(pformat(vars(args)))
debug('Loading data')
data = get_data(path=args.data_path,
smiles_columns=args.smiles_columns,
target_columns=args.target_columns)
args.task_names = get_task_names(path=args.data_path,
smiles_columns=args.smiles_columns,
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,
num_folds=args.num_folds,
args=args)
if args.save_smiles_splits:
save_smiles_splits(
data_path=args.data_path,
save_dir=args.save_dir,
task_names=args.task_names,
features_path=args.features_path,
train_data=train_data,
test_data=test_data,
smiles_columns=args.smiles_columns,
)
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)):
for s in datapoint.smiles:
datapoint.extend_features(
morgan_fingerprint(mol=s,
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,
random_state=args.seed)
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,
metrics=args.metrics,
args=args,
logger=logger)
else:
scores = multi_task_sklearn(model=model,
train_data=train_data,
test_data=test_data,
metrics=args.metrics,
args=args,
logger=logger)
for metric in args.metrics:
info(f'Test {metric} = {np.nanmean(scores[metric])}')
return scores
def cross_validate(args: TrainArgs,
train_func: Callable[[TrainArgs, MoleculeDataset, Logger], Dict[str, List[float]]]
) -> Tuple[float, float]:
"""
Runs k-fold cross-validation.
For each of k splits (folds) of the data, trains and tests a model on that split
and aggregates the performance across folds.
:param args: A :class:`~chemprop.args.TrainArgs` object containing arguments for
loading data and training the Chemprop model.
:param train_func: Function which runs training.
:return: A tuple containing the mean and standard deviation performance across folds.
"""
logger = create_logger(name=TRAIN_LOGGER_NAME, save_dir=args.save_dir, quiet=args.quiet)
if logger is not None:
debug, info = logger.debug, logger.info
else:
debug = info = print
# Initialize relevant variables
init_seed = args.seed
save_dir = args.save_dir
args.task_names = get_task_names(path=args.data_path, smiles_columns=args.smiles_columns,
target_columns=args.target_columns, ignore_columns=args.ignore_columns)
# Print command line
debug('Command line')
debug(f'python {" ".join(sys.argv)}')
# Print args
debug('Args')
debug(args)
# Save args
makedirs(args.save_dir)
args.save(os.path.join(args.save_dir, 'args.json'))
#set explicit H option and reaction option
set_explicit_h(args.explicit_h)
set_reaction(args.reaction, args.reaction_mode)
# Get data
debug('Loading data')
data = get_data(
path=args.data_path,
args=args,
smiles_columns=args.smiles_columns,
features_columns=args.features_columns,
logger=logger,
skip_none_targets=True
)
validate_dataset_type(data, dataset_type=args.dataset_type)
args.features_size = data.features_size()
if args.atom_descriptors == 'descriptor':
args.atom_descriptors_size = data.atom_descriptors_size()
args.ffn_hidden_size += args.atom_descriptors_size
elif args.atom_descriptors == 'feature':
args.atom_features_size = data.atom_features_size()
set_extra_atom_fdim(args.atom_features_size)
if args.bond_features_path is not None:
args.bond_features_size = data.bond_features_size()
set_extra_bond_fdim(args.bond_features_size)
debug(f'Number of tasks = {args.num_tasks}')
# Run training on different random seeds for each fold
all_scores = defaultdict(list)
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)
data.reset_features_and_targets()
# If resuming experiment, load results from trained models
test_scores_path = os.path.join(args.save_dir, 'test_scores.json')
if args.resume_experiment and os.path.exists(test_scores_path):
print('Loading scores')
with open(test_scores_path) as f:
model_scores = json.load(f)
# Otherwise, train the models
else:
model_scores = train_func(args, data, logger)
for metric, scores in model_scores.items():
all_scores[metric].append(scores)
all_scores = dict(all_scores)
# Convert scores to numpy arrays
for metric, scores in all_scores.items():
all_scores[metric] = np.array(scores)
# Report results
info(f'{args.num_folds}-fold cross validation')
# Report scores for each fold
for fold_num in range(args.num_folds):
for metric, scores in all_scores.items():
info(f'\tSeed {init_seed + fold_num} ==> test {metric} = {np.nanmean(scores[fold_num]):.6f}')
if args.show_individual_scores:
for task_name, score in zip(args.task_names, scores[fold_num]):
info(f'\t\tSeed {init_seed + fold_num} ==> test {task_name} {metric} = {score:.6f}')
# Report scores across folds
for metric, scores in all_scores.items():
avg_scores = np.nanmean(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 {metric} = {mean_score:.6f} +/- {std_score:.6f}')
if args.show_individual_scores:
for task_num, task_name in enumerate(args.task_names):
info(f'\tOverall test {task_name} {metric} = '
f'{np.nanmean(scores[:, task_num]):.6f} +/- {np.nanstd(scores[:, task_num]):.6f}')
# Save scores
with open(os.path.join(save_dir, TEST_SCORES_FILE_NAME), 'w') as f:
writer = csv.writer(f)
header = ['Task']
for metric in args.metrics:
header += [f'Mean {metric}', f'Standard deviation {metric}'] + \
[f'Fold {i} {metric}' for i in range(args.num_folds)]
writer.writerow(header)
for task_num, task_name in enumerate(args.task_names):
row = [task_name]
for metric, scores in all_scores.items():
task_scores = scores[:, task_num]
mean, std = np.nanmean(task_scores), np.nanstd(task_scores)
row += [mean, std] + task_scores.tolist()
writer.writerow(row)
# Determine mean and std score of main metric
avg_scores = np.nanmean(all_scores[args.metric], axis=1)
mean_score, std_score = np.nanmean(avg_scores), np.nanstd(avg_scores)
# Optionally merge and save test preds
if args.save_preds:
all_preds = pd.concat([pd.read_csv(os.path.join(save_dir, f'fold_{fold_num}', 'test_preds.csv'))
for fold_num in range(args.num_folds)])
all_preds.to_csv(os.path.join(save_dir, 'test_preds.csv'), index=False)
return mean_score, std_score
def run_training(args: TrainArgs,
data: MoleculeDataset,
logger: Logger = None) -> Dict[str, List[float]]:
"""
Loads data, trains a Chemprop model, and returns test scores for the model checkpoint with the highest validation score.
:param args: A :class:`~chemprop.args.TrainArgs` object containing arguments for
loading data and training the Chemprop model.
:param data: A :class:`~chemprop.data.MoleculeDataset` containing the data.
:param logger: A logger to record output.
:return: A dictionary mapping each metric in :code:`args.metrics` to a list of values for each task.
"""
if logger is not None:
debug, info = logger.debug, logger.info
else:
debug = info = print
# Set pytorch seed for random initial weights
torch.manual_seed(args.pytorch_seed)
# 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,
atom_descriptors_path=args.separate_test_atom_descriptors_path,
bond_features_path=args.separate_test_bond_features_path,
smiles_columns=args.smiles_columns,
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,
atom_descriptors_path=args.separate_val_atom_descriptors_path,
bond_features_path=args.separate_val_bond_features_path,
smiles_columns = args.smiles_columns,
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,
num_folds=args.num_folds,
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,
num_folds=args.num_folds,
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,
num_folds=args.num_folds,
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(
data_path=args.data_path,
save_dir=args.save_dir,
task_names=args.task_names,
features_path=args.features_path,
train_data=train_data,
val_data=val_data,
test_data=test_data,
smiles_columns=args.smiles_columns,
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
if args.atom_descriptor_scaling and args.atom_descriptors is not None:
atom_descriptor_scaler = train_data.normalize_features(replace_nan_token=0, scale_atom_descriptors=True)
val_data.normalize_features(atom_descriptor_scaler, scale_atom_descriptors=True)
test_data.normalize_features(atom_descriptor_scaler, scale_atom_descriptors=True)
else:
atom_descriptor_scaler = None
if args.bond_feature_scaling and args.bond_features_size > 0:
bond_feature_scaler = train_data.normalize_features(replace_nan_token=0, scale_bond_features=True)
val_data.normalize_features(bond_feature_scaler, scale_bond_features=True)
test_data.normalize_features(bond_feature_scaler, scale_bond_features=True)
else:
bond_feature_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')
scaler = train_data.normalize_targets()
else:
scaler = None
# Get loss function
loss_func = get_loss_func(args)
# 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:
set_cache_graph(True)
num_workers = 0
else:
set_cache_graph(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,
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
)
test_data_loader = MoleculeDataLoader(
dataset=test_data,
batch_size=args.batch_size,
num_workers=num_workers
)
if args.class_balance:
debug(f'With class_balance, effective train size = {train_data_loader.iter_size:,}')
# 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_FILE_NAME), model, scaler,
features_scaler, atom_descriptor_scaler, bond_feature_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,
metrics=args.metrics,
dataset_type=args.dataset_type,
scaler=scaler,
logger=logger
)
for metric, scores in val_scores.items():
# Average validation score
avg_val_score = np.nanmean(scores)
debug(f'Validation {metric} = {avg_val_score:.6f}')
writer.add_scalar(f'validation_{metric}', avg_val_score, n_iter)
if args.show_individual_scores:
# Individual validation scores
for task_name, val_score in zip(args.task_names, scores):
debug(f'Validation {task_name} {metric} = {val_score:.6f}')
writer.add_scalar(f'validation_{task_name}_{metric}', val_score, n_iter)
# Save model checkpoint if improved validation score
avg_val_score = np.nanmean(val_scores[args.metric])
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_FILE_NAME), model, scaler, features_scaler,
atom_descriptor_scaler, bond_feature_scaler, args)
if args.split_sizes[1] == 0.0:
save_checkpoint(os.path.join(save_dir, MODEL_FILE_NAME), model,
scaler, features_scaler,
atom_descriptor_scaler, bond_feature_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_FILE_NAME), 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,
metrics=args.metrics,
dataset_type=args.dataset_type,
logger=logger
)
if len(test_preds) != 0:
sum_test_preds += np.array(test_preds)
# Average test score
for metric, scores in test_scores.items():
avg_test_score = np.nanmean(scores)
info(f'Model {model_idx} test {metric} = {avg_test_score:.6f}')
writer.add_scalar(f'test_{metric}', avg_test_score, 0)
if args.show_individual_scores:
# Individual test scores
for task_name, test_score in zip(args.task_names, scores):
info(f'Model {model_idx} test {task_name} {metric} = {test_score:.6f}')
writer.add_scalar(f'test_{task_name}_{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,
metrics=args.metrics,
dataset_type=args.dataset_type,
logger=logger
)
for metric, scores in ensemble_scores.items():
# Average ensemble score
avg_ensemble_test_score = np.nanmean(scores)
info(f'Ensemble test {metric} = {avg_ensemble_test_score:.6f}')
# Individual ensemble scores
if args.show_individual_scores:
for task_name, ensemble_score in zip(args.task_names, scores):
info(f'Ensemble test {task_name} {metric} = {ensemble_score:.6f}')
# Save scores
with open(os.path.join(args.save_dir, 'test_scores.json'), 'w') as f:
json.dump(ensemble_scores, f, indent=4, sort_keys=True)
# Optionally save test preds
if args.save_preds:
test_preds_dataframe = pd.DataFrame(data={'smiles': test_data.smiles()})
for i, task_name in enumerate(args.task_names):
test_preds_dataframe[task_name] = [pred[i] for pred in avg_test_preds]
test_preds_dataframe.to_csv(os.path.join(args.save_dir, 'test_preds.csv'), index=False)
return ensemble_scores
def predict_sklearn(args: SklearnPredictArgs) -> None:
"""
Loads data and a trained scikit-learn model and uses the model to make predictions on the data.
:param args: A :class:`~chemprop.args.SklearnPredictArgs` object containing arguments for
loading data, loading a trained scikit-learn model, and making predictions with the model.
"""
print('Loading data')
data = get_data(path=args.test_path,
smiles_columns=args.smiles_columns,
target_columns=[],
ignore_columns=[],
store_row=True)
print('Loading training arguments')
with open(args.checkpoint_paths[0], 'rb') as f:
model = pickle.load(f)
train_args: SklearnTrainArgs = SklearnTrainArgs().from_dict(
model.train_args, skip_unsettable=True)
print('Computing morgan fingerprints')
morgan_fingerprint = get_features_generator('morgan')
for datapoint in tqdm(data, total=len(data)):
for s in datapoint.smiles:
datapoint.extend_features(
morgan_fingerprint(mol=s,
radius=train_args.radius,
num_bits=train_args.num_bits))
print(
f'Predicting with an ensemble of {len(args.checkpoint_paths)} models')
sum_preds = np.zeros((len(data), train_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=train_args.model_type,
dataset_type=train_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
for datapoint, preds in zip(data, avg_preds):
for pred_name, pred in zip(train_args.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 molecule_fingerprint(
args: PredictArgs,
smiles: List[List[str]] = None) -> List[List[Optional[float]]]:
"""
Loads data and a trained model and uses the model to encode fingerprint vectors for the data.
:param args: A :class:`~chemprop.args.PredictArgs` object containing arguments for
loading data and a model and making predictions.
:param smiles: List of list of SMILES to make predictions on.
:return: A list of fingerprint vectors (list of floats)
"""
print('Loading training args')
train_args = load_args(args.checkpoint_paths[0])
# Update args with training arguments
update_prediction_args(predict_args=args,
train_args=train_args,
validate_feature_sources=False)
args: Union[PredictArgs, TrainArgs]
#set explicit H option and reaction option
set_explicit_h(train_args.explicit_h)
set_reaction(train_args.reaction, train_args.reaction_mode)
print('Loading data')
if smiles is not None:
full_data = get_data_from_smiles(
smiles=smiles,
skip_invalid_smiles=False,
features_generator=args.features_generator)
else:
full_data = get_data(path=args.test_path,
smiles_columns=args.smiles_columns,
target_columns=[],
ignore_columns=[],
skip_invalid_smiles=False,
args=args,
store_row=True)
print('Validating SMILES')
full_to_valid_indices = {}
valid_index = 0
for full_index in range(len(full_data)):
if all(mol is not None for mol in full_data[full_index].mol):
full_to_valid_indices[full_index] = valid_index
valid_index += 1
test_data = MoleculeDataset(
[full_data[i] for i in sorted(full_to_valid_indices.keys())])
# Edge case if empty list of smiles is provided
if len(test_data) == 0:
return [None] * len(full_data)
print(f'Test size = {len(test_data):,}')
# Create data loader
test_data_loader = MoleculeDataLoader(dataset=test_data,
batch_size=args.batch_size,
num_workers=args.num_workers)
# Load model
print(f'Encoding smiles into a fingerprint vector from a single model')
if len(args.checkpoint_paths) != 1:
raise ValueError(
"Fingerprint generation only supports one model, cannot use an ensemble"
)
model = load_checkpoint(args.checkpoint_paths[0], device=args.device)
scaler, features_scaler, atom_descriptor_scaler, bond_feature_scaler = load_scalers(
args.checkpoint_paths[0])
# Normalize features
if args.features_scaling or train_args.atom_descriptor_scaling or train_args.bond_feature_scaling:
test_data.reset_features_and_targets()
if args.features_scaling:
test_data.normalize_features(features_scaler)
if train_args.atom_descriptor_scaling and args.atom_descriptors is not None:
test_data.normalize_features(atom_descriptor_scaler,
scale_atom_descriptors=True)
if train_args.bond_feature_scaling and args.bond_features_size > 0:
test_data.normalize_features(bond_feature_scaler,
scale_bond_features=True)
# Make fingerprints
model_preds = model_fingerprint(model=model, data_loader=test_data_loader)
# Save predictions
print(f'Saving predictions to {args.preds_path}')
assert len(test_data) == len(model_preds)
makedirs(args.preds_path, isfile=True)
# Copy predictions over to full_data
total_hidden_size = args.hidden_size * args.number_of_molecules
for full_index, datapoint in enumerate(full_data):
valid_index = full_to_valid_indices.get(full_index, None)
preds = model_preds[valid_index] if valid_index is not None else [
'Invalid SMILES'
] * total_hidden_size
fingerprint_columns = [f'fp_{i}' for i in range(total_hidden_size)]
for i in range(len(fingerprint_columns)):
datapoint.row[fingerprint_columns[i]] = preds[i]
# Write predictions
with open(args.preds_path, 'w') as f:
writer = csv.DictWriter(f,
fieldnames=args.smiles_columns +
fingerprint_columns,
extrasaction='ignore')
writer.writeheader()
for datapoint in full_data:
writer.writerow(datapoint.row)
return model_preds