def fit_one_category(self, dataset, category=None, verbose=False):
train = self.scale(dataset)
as_windows_train = get_windowed_ts(train, window_size=self.window_size)
train_X, train_y = split_history_and_current(as_windows_train)
train_X = train_X.reshape(train_X.shape[0], train_X.shape[1], 1)
# Get model definition
model = self.lstm_forecast_model(window_size=self.window_size,
verbose=verbose)
# Fit model
start = time.time()
history = model.fit(
train_X,
train_y,
batch_size=max(self.batch_size, len(train_X)),
epochs=self.epochs,
validation_split=0.1,
callbacks=[EarlyStopping(patience=4)],
)
print("> Compilation Time : ", time.time() - start)
self.model[category] = model
def predict_one_category(self, X, category):
test = self.scale(X)
as_windows_test = get_windowed_ts(test, window_size=self.window_size)
test_X, test_y = split_history_and_current(as_windows_test)
test_X = test_X.reshape(test_X.shape[0], test_X.shape[1], 1)
preds = self.model[category].predict(test_X)
from sklearn.metrics import mean_squared_error
# clear session
clear_session()
del self.model
actual = test_y
prediction = preds
mse = mean_squared_error(actual, prediction)
print("MSE=" + str(mse))
results = pd.DataFrame({
"date": X.index[self.window_size:],
"prediction": preds.squeeze(),
"actual": test_y,
"diff": test_y - preds.squeeze(),
})
results["diff_rolling_std"] = (
results["diff"].rolling(self.lookback,
min_periods=self.min_periods).std()
) # naive version where we look at the standard deviation of the
# difference between predicted and actual in a previous window.
results["prediction"] = np.where(
results["diff"] > self.num_of_std * results["diff_rolling_std"], 1,
0)
return results