if torch.cuda.is_available():
args['device'] = torch.device('cuda:0')
else:
args['device'] = torch.device('cpu')
if args['task_names'] is not None:
args['task_names'] = args['task_names'].split(',')
args['node_featurizer'] = CanonicalAtomFeaturizer()
df = pd.read_csv(args['csv_path'])
mkdir_p(args['result_path'])
dataset = MoleculeCSVDataset(df=df,
smiles_to_graph=smiles_to_bigraph,
node_featurizer=args['node_featurizer'],
edge_featurizer=None,
smiles_column=args['smiles_column'],
cache_file_path=args['result_path'] +
'/graph.bin',
task_names=args['task_names'])
args['n_tasks'] = dataset.n_tasks
train_set, val_set, test_set = split_dataset(args, dataset)
exp_config = get_configure(args)
if exp_config is None:
print('Start hyperparameter search with Bayesian optimization')
bayesian_optimization(args, train_set, val_set, test_set)
else:
print('Use the best hyperparameters found before')
args.update(exp_config)
main(args, train_set, val_set, test_set)
args['node_featurizer'] = CanonicalAtomFeaturizer()
df = pd.read_csv(args['csv_path'])
mkdir_p(args['result_path'])
dataset = MoleculeCSVDataset(df=df,
smiles_to_graph=smiles_to_bigraph,
node_featurizer=args['node_featurizer'],
edge_featurizer=None,
smiles_column=args['smiles_column'],
cache_file_path=args['result_path'] + '/graph.bin',
task_names=args['task_names'])
args['n_tasks'] = dataset.n_tasks
train_set, val_set, test_set = split_dataset(args, dataset)
if args['num_evals'] is not None:
assert args['num_evals'] > 0, 'Expect the number of hyperparameter search trials to ' \
'be greater than 0, got {:d}'.format(args['num_evals'])
print('Start hyperparameter search with Bayesian '
'optimization for {:d} trials'.format(args['num_evals']))
trial_path = bayesian_optimization(args, train_set, val_set, test_set)
else:
print('Use the manually specified hyperparameters')
exp_config = get_configure(args['model'])
main(args, exp_config, train_set, val_set, test_set)
trial_path = args['result_path'] + '/1'
# Copy final
copyfile(trial_path + '/model.pth', args['result_path'] + '/model.pth')
copyfile(trial_path + '/configure.json', args['result_path'] + '/configure.json')
copyfile(trial_path + '/eval.txt', args['result_path'] + '/eval.txt')
else:
args['device'] = torch.device('cpu')
args = init_featurizer(args)
mkdir_p(args['result_path'])
if args['dataset'] == 'FreeSolv':
from dgllife.data import FreeSolv
dataset = FreeSolv(smiles_to_graph=partial(smiles_to_bigraph, add_self_loop=True),
node_featurizer=args['node_featurizer'],
edge_featurizer=args['edge_featurizer'],
n_jobs=1 if args['num_workers'] == 0 else args['num_workers'])
elif args['dataset'] == 'Lipophilicity':
from dgllife.data import Lipophilicity
dataset = Lipophilicity(smiles_to_graph=partial(smiles_to_bigraph, add_self_loop=True),
node_featurizer=args['node_featurizer'],
edge_featurizer=args['edge_featurizer'],
n_jobs=1 if args['num_workers'] == 0 else args['num_workers'])
elif args['dataset'] == 'ESOL':
from dgllife.data import ESOL
dataset = ESOL(smiles_to_graph=partial(smiles_to_bigraph, add_self_loop=True),
node_featurizer=args['node_featurizer'],
edge_featurizer=args['edge_featurizer'],
n_jobs=1 if args['num_workers'] == 0 else args['num_workers'])
else:
raise ValueError('Unexpected dataset: {}'.format(args['dataset']))
args['n_tasks'] = dataset.n_tasks
train_set, val_set, test_set = split_dataset(args, dataset)
exp_config = get_configure(args['model'], args['featurizer_type'], args['dataset'])
main(args, exp_config, train_set, val_set, test_set)