def main():
from src.utils import data_utils
from keras import layers
from matplotlib import pyplot as plt
x, y = data_utils.get_data_in_shape('EA', (['CPI'], ['CPI']), 1)
x_train, x_val, x_test, = data_utils.train_val_test_split(x, 12, 12)
y_train, y_val, y_test = data_utils.train_val_test_split(y, 12, 12)
input = Input(batch_shape=(1, 1,))
layer = layers.Reshape((1, 1))(input)
layer = layers.LSTM(10, stateful=True)(layer)
# layer = layers.LSTM(10, return_sequences=False, stateful=True)(layer)
layer = layers.Dense(1, activation='linear')(layer)
model = RecurrentModel(inputs=input, outputs=layer)
model.compile(optimizer='adam', loss='mse')
model.fit(x_train, y_train, epochs=200, batch_size=1)
pred = model.predict(x_val, batch_size=1)
fcast = model.forecast(x_val, y_val, batch_size=1)
from keras.utils.vis_utils import plot_model
plot_model(model)
model.summary()
plt.plot(y_val.values)
plt.plot(pred)
plt.show()
plt.plot(y_val.values)
plt.plot(fcast)
plt.show()
def persistence_mse(series):
columns = ['train pred', 'val pred', 'test pred']
lossess = OrderedDict([(i, 0) for i in columns])
y_pred, y_true = data_utils.get_xy_data(series, 1)
true_data = data_utils.train_val_test_split(y_true, 12, 12)
pred_data = data_utils.train_val_test_split(y_pred, 12, 12)
for pred, true, col in zip(pred_data, true_data, columns):
lossess[col] = mse(true, pred)
return pd.DataFrame(lossess, index=[0])
def set_data_params(self, **data_params):
self.x, self.y = data_utils.get_data_in_shape(**data_params)
if self.difference:
self.x, self.y = self.transformer.fit(self.x, self.y)
self.x_train, self.x_val, self.x_test, = data_utils.train_val_test_split(self.x, 12, 12)
self.y_train, self.y_val, self.y_test = data_utils.train_val_test_split(self.y, 12, 12)
self.data_params = data_params
self.is_multioutput = len(data_params['vars'][1]) > 1
self.variables = data_params['vars'][1]
self.is_data_set = True
self.is_fitted = False
def one_one():
from src.utils import data_utils
from keras import layers
x, y = data_utils.get_data_in_shape('EA', (['CPI'], ['CPI']), 1)
x_train, x_val, x_test, = data_utils.train_val_test_split(x, 12, 12)
y_train, y_val, y_test = data_utils.train_val_test_split(y, 12, 12)
input = Input(shape=(1,))
layer = Dense(1)(input)
model = ForecastModel(inputs=input, outputs=layer)
model.compile(optimizer='adam', loss='mse')
print model.forecast(x_val, y_val)
def persistence_prediction(series):
y_pred, y_true = data_utils.get_xy_data(series, 1)
y_train, y_val, y_test = data_utils.train_val_test_split(y_pred, 12, 12)
predictions = pd.concat([y_true, y_train, y_val, y_test], axis=1)
predictions.columns = [
'true values', 'train prediction', 'val prediction', 'test prediction'
]
return predictions
def many_one():
from src.utils import data_utils
from keras import layers
x, y = data_utils.get_data_in_shape('EA', (['CPI', 'GDP'], ['CPI']), 2)
x_train, x_val, x_test, = data_utils.train_val_test_split(x, 12, 12)
y_train, y_val, y_test = data_utils.train_val_test_split(y, 12, 12)
input1 = Input(shape=(2,))
input2 = Input(shape=(2,))
layer = layers.concatenate([input1, input2])
layer = Dense(1)(layer)
model = ForecastModel(inputs=[input1, input2], outputs=layer)
model.compile(optimizer='adam', loss='mse')
print model.forecast(x_val, y_val)
def main():
data_params = OrderedDict()
data_params['country'] = 'EA'
data_params['vars'] = (['CPI', 'GDP'], ['CPI', 'GDP'])
data_params['lags'] = (1, 1)
tr = DifferenceTransformer()
x, y = data_utils.get_data_in_shape(**data_params)
x, y = tr.fit(x, y)
x_train, x_val, x_test = data_utils.train_val_test_split(x, 12, 12)
y_train, y_val, y_test = data_utils.train_val_test_split(y, 12, 12)
curr_x = x_val
cuur_y = y_val
x_trans, y_trans = tr.transform(curr_x, cuur_y)
y_orig = tr.inverse_transform(y_trans.values, cuur_y)
print np.concatenate([cuur_y, y_orig], axis=1)
print y_orig