def create_sts_model(train_x, train_y):
"""
Create model using Gaussian Naive Bayes and save this
:param train_x: lagged list, regressor
:param train_y: lagged_list, regressand
:return: trained model
"""
model = GaussianNB()
model.fit(train_x, train_y)
save_model(model, "simple_time_series")
return model
def feed_forward_create_model(ml_df):
ml_df = ml_df.copy()
train_dataset, test_dataset, train_labels, test_labels = preprocess(ml_df)
normed_train_data, normed_test_data = normalize_dataset(
train_dataset, test_dataset
)
history, model = canteen_model(normed_train_data, train_labels)
test_labels["prediction"] = model.predict(normed_test_data).flatten()
mae = mean_absolute_error(
np.asarray(test_labels["Canteen"]),
np.asarray(test_labels["prediction"]),
)
global FEED_MAE
if mae < FEED_MAE:
save_model(history.history, "feed_forward_history")
save_model(history.epoch, "feed_forward_epoch")
save_model(train_dataset, "feed_forward_train_dataset")
save_model(test_labels, "feed_forward_test_set_prediction")
model.save(
"{}/models/saved_models/feed_forward_model.h5".format(ROOT_DIR)
)
FEED_MAE = mae
print("FEED", str(mae))
return model
def linear_create_model(df):
"""
Creates model and saves the model to a sav file.
:param df: dataframe with canteen and date
:return: model, x, y and test data set
"""
dataframe = pd.DataFrame(df.index)
dataframe["Canteen"] = df["Canteen"].values
dataframe["date"] = pd.to_datetime(dataframe["date"])
dataframe["date"] = dataframe["date"].map(dt.datetime.toordinal)
train, test = train_test_split(dataframe, test_size=DATA_SET_TEST_SIZE)
y = np.asarray(train["Canteen"])
x = np.asarray(train["date"]).reshape((-1, 1))
model = LinearRegression(normalize=True).fit(
x, y) # create linear regression object #train model on train data
save_model(model, "linear_regression")
return model, x, y, test
def create_prophet_model(train):
# Create model based on training dataframe
train["floor"] = 0
train["cap"] = 2100
m = Prophet(
growth="logistic",
interval_width=0.95,
seasonality_mode="multiplicative",
changepoint_prior_scale=20,
seasonality_prior_scale=20,
yearly_seasonality=12,
daily_seasonality=False,
)
m.add_seasonality(name="monthly", period=30.5, fourier_order=4)
m.add_country_holidays(country_name="Norway")
m.fit(train)
# m.train_holiday_names
save_model(m, "prophet")
return m
def catboost_create_model(dt_df):
raw_data = dt_df.copy()
df = preprocess_to_catboost(raw_data)
train_dataset, test_dataset, train_labels, test_labels = preprocess(df)
train_dataset_combined = Pool(train_dataset, train_labels)
eval_dataset = Pool(test_dataset, test_labels)
model = CatBoostRegressor(iterations=2000,
learning_rate=0.05,
depth=5,
eval_metric="MAE")
model.fit(train_dataset_combined, eval_set=eval_dataset, verbose=0)
test_labels["prediction"] = model.predict(test_dataset).flatten()
mae = mean_absolute_error(
np.asarray(test_labels["Canteen"]),
np.asarray(test_labels["prediction"]),
)
global CAT_MAE
if mae < CAT_MAE:
save_model(model, "catboost")
save_model(model.get_evals_result(), "catboost_evaluation_result")
save_model(test_labels, "catboost_test_set_prediction")
CAT_MAE = mae
print("CAT", str(mae))
return model
def predict_lstm_with_testset(ml_df, test_period, local_testing=True):
"""
LSTM prediction with existing data in ml_df.csv for training the model
:param ml_df: a dataframe that is non categorical
:param test_period: int for setting the test size
:param local_testing: bool value, will print training data if set to True.
:return: model history and the predicted values
"""
df = ml_df.copy()
model, scaler, test_dataset, test_labels, history = train_lstm(
df, test_period, local_testing)
# From tutorial https://machinelearningmastery.com/multivariate-time-series-forecasting-lstms-keras/
yhat = model.predict(test_dataset)
test_dataset = test_dataset.reshape(
(test_dataset.shape[0], test_dataset.shape[2]))
inv_yhat = concatenate((yhat, test_dataset), axis=1)
inv_yhat = scaler.inverse_transform(inv_yhat)
inv_yhat = inv_yhat[:, 0]
test_y = test_labels.reshape((len(test_labels), 1))
inv_y = concatenate((test_y, test_dataset), axis=1)
inv_y = scaler.inverse_transform(inv_y)
inv_y = inv_y[:, 0]
pred_df = pd.DataFrame({
"prediction": inv_yhat.flatten(),
"Canteen": inv_y.flatten()
})
mae = mean_absolute_error(np.asarray(pred_df["Canteen"]),
np.asarray(pred_df["prediction"]))
global LSTM_MAE
if mae < LSTM_MAE:
model.save("{}/models/saved_models/lstm_model.h5".format(ROOT_DIR))
save_model(history.history, "lstm_history")
save_model(history.epoch, "lstm_epoch")
save_model(pred_df, "lstm_test_set_prediction")
LSTM_MAE = mae
print("LSTM", str(mae))
return history, inv_yhat