# setting up GPU environment
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "1" # set GPU 0
# parser = argparse.ArgumentParser(description='Utility to build Classic ML models')
# parser.add_argument('-i', '--infile', required=True)
# parser.add_argument('-nl', '--n_layer', required=True)
# args = parser.parse_args()
dataset = 'trainingset'
features = 'ECFP6'
n_layer = 3
n_split = 5
X_train, y_train = make_dataset('{}.sdf'.format(dataset), features)
print("Training set includes {} descriptors".format(X_train.shape[1]))
print("Training set includes {}, {} molecules".format(
X_train.shape[0], y_train.shape[0]))
nodes = X_train.shape[1]
num_hidden = [nodes for _ in range(n_layer)]
num_steps = 10001
out_batch = BatchLogger(display=5)
reduce_lr = ReduceLROnPlateau(monitor='loss',
factor=0.9,
patience=50,
min_lr=0.00001,
verbose=1)
seed = 0
env_var = args.env_var
data_dir = os.getenv(env_var)
name_col = args.name_col
endpoint = args.endpoint
threshold = args.threshold
test_set_size = args.test_set_size
# Check to see if necessary directories are present and if not, create them
directory_check(data_dir)
# get training data and split in training, test
# and use a seed for reproducibility
X, y = make_dataset(f'{dataset}.sdf',
data_dir=env_var,
features=features,
name_col=name_col,
endpoint=endpoint,
threshold=threshold)
X_train, y_train_class, X_test, y_test_class = split_train_test(
X, y, n_splits, test_set_size, seed, None)
if len(pd.unique(X.values.ravel('K'))) > 2:
CLASSIFIER_ALGORITHMS.pop(4)
else:
CLASSIFIER_ALGORITHMS.pop(1)
cv = model_selection.StratifiedKFold(shuffle=True,
n_splits=n_splits,
random_state=seed)
threshold = args.threshold
evaluate = args.evaluate
train_name = args.train_name
for alg in ['ada', 'bnb', 'knn', 'nb', 'rf', 'svc']:
for feature in features:
for endpoint in endpoints:
model_name = f'{alg}_{train_name}_{feature}_{endpoint}_{threshold}_pipeline'
model_file_path = os.path.join(data_dir, 'ML_models',
f'{model_name}.pkl')
if os.path.exists(model_file_path):
if evaluate in ['Y', 'y']:
X_pred_set, y = make_dataset(f'{prediction_set}.sdf',
data_dir=env_var,
features=feature,
name_col=name_col,
endpoint=endpoint,
threshold=threshold)
else:
X_pred_set = make_dataset(f'{prediction_set}.sdf',
data_dir=env_var,
features=features,
name_col=name_col,
endpoint=None,
threshold=None,
pred_set=True)
loaded_model = load(model_file_path)
predictions = loaded_model.predict(X_pred_set)
probabilities = loaded_model.predict_proba(X_pred_set)
args = parser.parse_args()
data_dir = os.getenv(args.env_var)
env_var = args.env_var
features = args.features
name_col = args.name_col
prediction_set = args.prediction_set
endpoint = args.endpoint
threshold = args.threshold
train_name = args.train_name
algorithms = args.algorithms.lower().split(',')
X_pred = make_dataset(f'{prediction_set}.sdf',
data_dir=env_var,
features=features,
name_col=name_col,
endpoint=endpoint,
threshold=threshold,
pred_set=True)
algorithms = [alg for alg in algorithms]
preds = []
if len(algorithms) < 1:
raise Exception(
'Please enter at least one algorithm with which to make predictions.')
for alg in algorithms:
model_name = f'{alg}_{train_name}_{features}_{endpoint}_{threshold}_pipeline'
model_file_path = os.path.join(data_dir, 'models', f'{model_name}.pkl')
if os.path.exists(model_file_path):
seed = 0
env_var = args.data_dir
data_dir = os.getenv(env_var)
name_col = args.name_col
endpoint = args.endpoint
test_set_size = args.test_set_size
# Check to see if necessary directories are present and if not, create them
directory_check(data_dir)
# get training data and split in training, test
# and use a seed for reproducibility
X, y_regress, y_class = make_dataset('{}.sdf'.format(dataset),
data_dir=env_var,
features=features,
name_col=name_col,
endpoint=endpoint,
regress=True,
threshold=2000)
X_train, y_train_regress, X_test, y_test_regress = split_train_test(X, y_class, n_splits, test_set_size, seed, None)
if test_set_size != 0:
y_train_regress = y_regress.loc[X_train.index].values.ravel()
y_test_regress = y_regress.loc[X_test.index].values.ravel()
y_train_regress = y_regress.loc[X_train.index]
cv = model_selection.KFold(shuffle=True, n_splits=n_splits, random_state=seed)
for name, clf, params in REGRESSOR_ALGS:
seed = 0
env_var = args.data_dir
data_dir = os.getenv(env_var)
name_col = args.name_col
endpoint = args.endpoint
threshold = args.threshold
test_set_size = args.test_set_size
# Check to see if necessary directories are present and if not, create them
directory_check(data_dir)
# get training data and split in training, test
# and use a seed for reproducibility
X, y = make_dataset('{}.sdf'.format(dataset),
data_dir=env_var,
features=features,
name_col=name_col,
endpoint=endpoint,
threshold=threshold)
X_train, y_train_class, X_test, y_test_class = split_train_test(
X, y, n_splits, test_set_size, seed, None)
cv = model_selection.StratifiedKFold(shuffle=True,
n_splits=n_splits,
random_state=seed)
print("Num Actives: {}".format((y == 1).sum()))
print("Num Inactives: {}".format((y == 0).sum()))
for name, clf, params in CLASSIFIER_ALGS:
pipe = pipeline.Pipeline([('scaler', StandardScaler()), (name, clf)])
grid_search = model_selection.GridSearchCV(pipe,