def main():
parser = get_parser()
args = parser.parse_args()
checked_args = check_argument_parsing(args)
multi = True if args.objective == "multi-metrics" else False
if checked_args.task == "train":
df = load_dataset(args.fname)
# vectorization
if not (args.model_type == "rnn"):
x, y = vectorizes_features(df), vectorizes_label(df, multi)
else:
mol2vec = word2vec.Word2Vec.load('models/model_300dim.pkl')
y = vectorizes_label(df, multi)
x = vec_mol2vec_smile(df["smiles"].tolist(), mol2vec)
# train-test-split
x_train, x_test, y_train, y_test = train_test_split(
x,
y,
stratify=y if not multi else None,
train_size=0.9,
random_state=args.random_state)
if not os.path.isdir(args.output_dir):
os.mkdir(args.output_dir)
# useful for restesting on same array
for arr, arr_name in ((x_test, "x_test.npy"), (y_test, "y_test.npy")):
np.save(os.path.join(args.output_dir, arr_name), arr)
if args.model_type == "dummy":
clf = DummyClassifier(strategy="most_frequent")
clf.fit(x_train, y_train)
score = clf.score(x_test, y_test)
with open(args.model, "wb") as f:
pickle.dump(clf, f)
elif args.model_type == "mlp":
with open(args.hyperparameters) as f:
hp = json.load(f)
clf = get_mlp(1 if args.objective == "single-metrics" else 9,
hp["neurons"], hp["dropout_rate"], hp["activation"])
clf.compile(optimizer=tf.keras.optimizers.Adam(
learning_rate=hp["learning_rate"]),
loss="categorical_crossentropy",
metrics=['accuracy'])
clf.fit(x_train,
y_train,
batch_size=hp["batch_size"],
epochs=hp["epochs"],
validation_data=(x_test, y_test),
callbacks=get_callbacks(args.model))
score = clf.evaluate(x_test, y_test,
batch_size=hp["batch_size"])[1]
elif args.model_type == "rnn":
with open(args.hyperparameters) as f:
hp = json.load(f)
clf = get_rnn(1 if args.objective == "single-metrics" else 9,
hp["neurons"], hp["dropout_rate"])
clf.compile(optimizer=tf.keras.optimizers.Adam(
learning_rate=hp["learning_rate"]),
loss="categorical_crossentropy",
metrics=['accuracy'])
clf.fit(x_train,
y_train,
batch_size=hp["batch_size"],
epochs=hp["epochs"],
validation_data=(x_test, y_test),
callbacks=get_callbacks(args.model))
score = clf.evaluate(x_test, y_test,
batch_size=hp["batch_size"])[1]
results = {
"model-type": args.model_type,
"random_state": args.random_state,
"score": float(score)
}
with open(os.path.join(args.output_dir, "resultats.json"), "w") as f:
json.dump(results, f)
else:
if args.model_type == "dummy":
with open(args.model, "rb") as f:
clf = pickle.load(f)
elif args.model_type == "mlp":
clf = tf.keras.Model.load_model(args.model)
elif args.model_type == "rnn":
clf = tf.keras.load_model(args.model)
mol2vec = word2vec.Word2Vec.load('models/model_300dim.pkl')
if checked_args.task == "predict":
if args.model_type in ["dummy", "mlp"]:
try:
x = vectorizes_smile(checked_args.smile)
print("mol: {}, {}".format(args.smile, clf.predict(x)))
except Exception() as e:
print(e)
else:
x = vec_mol2vec_smile([args.smile], mol2vec)
if checked_args.task == "evaluate":
df = load_dataset(args.fname)
if args.model_type == "rnn":
x_test = vec_mol2vec_smile(df["smiles"].tolist())
else:
x_test = vectorizes_features(df)
y_test = vectorizes_label(df, multi)
score = clf.score(x_test, y_test)
print(score)
# API
elif checked_args.task == "server-start":
app = get_app(clf, args.model_type, args.mol2vec)
app.run()
