def test_predictions_with_additional_regressors(sample_data):
'''
TODO: update
'''
my_chronos = Chronos(n_changepoints=0, max_iter=100)
my_chronos.add_regressors("reg1", "mul")
my_chronos.add_regressors("reg2", "mul")
# Check not including the regressor throws an error
with pytest.raises(KeyError):
my_chronos.fit(sample_data)
# Now check with the regressor
sample_data['reg1'] = [1] * sample_data.shape[0]
sample_data['reg2'] = [1] * sample_data.shape[0]
my_chronos.fit(sample_data)
future_df = my_chronos.make_future_dataframe(include_history=False)
# Check not including the regressor throws an error
with pytest.raises(KeyError):
predictions = my_chronos.predict(future_df)
# Now check with the regressor
future_df['reg1'] = [1] * future_df.shape[0]
future_df['reg2'] = [1] * future_df.shape[0]
predictions = my_chronos.predict(future_df)
predictions.drop('y', axis=1, inplace=True)
assert (predictions.isna().sum().sum() == 0)
def test_plotting_no_history(capsys, monkeypatch, sample_data):
'''
Test the plotting of data when no history is provided.
This was an issue raised in issue #2.
'''
for distribution in chronos_utils.SUPPORTED_DISTRIBUTIONS:
my_chronos = Chronos(n_changepoints=0, max_iter=100, distribution=distribution)
my_chronos.fit(sample_data)
future_df = my_chronos.make_future_dataframe(include_history=False)
predictions = my_chronos.predict(future_df)
fig = plt.figure(figsize=(15,5))
gs = gridspec.GridSpec(1,1, figure=fig)
gs_section = gs[0,0]
ax = fig.add_subplot(gs_section)
#chronos_plotting.plot_predictions(predictions, my_chronos, fig=fig, gs_section=gs_section)
chronos_plotting.plot_predictions(predictions, my_chronos, axs=ax)
plt.savefig("test_prediction_no_history.png")
std_error = capsys.readouterr().err
assert(std_error == "")
os.remove("test_prediction_no_history.png")
def test_error_for_nan_values(sample_data):
'''
Tests that Chronos complains when given
nan values to fit on, but not to predict on
'''
def add_dummy_regressors(data):
z = data.index.values
y = 0.01 * z + np.sin(z / 30)
data['y'] = y
# should be easy since the target is just dummy1 + dummy2
dummy1 = 0.01 * z
dummy2 = np.sin(z / 30)
data['dummy1'] = dummy1
data['dummy2'] = dummy2
return data
sample_data = add_dummy_regressors(sample_data)
my_chronos = Chronos(max_iter=200)
# add dummies
my_chronos.add_regressors("dummy1")
my_chronos.add_regressors("dummy2")
# check nan values fail for all columns
for col in ['ds', 'y', 'dummy1', 'dummy2']:
sample_data_with_nans = sample_data.copy()
sample_data_with_nans.loc[0, col] = np.nan
with pytest.raises(ValueError):
my_chronos.fit(sample_data_with_nans)
# check original data works
my_chronos.fit(sample_data)
future_df = my_chronos.make_future_dataframe()
future_df = add_dummy_regressors(future_df)
# check nan values fail for all columns EXCEPT
# the target column when making predictions
for col in ['ds', 'y', 'dummy1', 'dummy2']:
future_df_with_nans = future_df.copy()
future_df_with_nans.loc[0, col] = np.nan
if (col == 'y'):
predictions = my_chronos.predict(future_df_with_nans)
else:
with pytest.raises(ValueError):
predictions = my_chronos.predict(future_df_with_nans)
def test_prediction_with_easy_extra_regressors(sample_data):
'''
Test that when the size of the data is too small, the
number of changepoints gets adjusted
'''
z = sample_data.index.values
y = 0.01 * z + np.sin(z / 30)
sample_data['y'] = y
# should be easy since the target is just dummy1 + dummy2
dummy1 = 0.01 * z
dummy2 = np.sin(z / 30)
sample_data['dummy1'] = dummy1
sample_data['dummy2'] = dummy2
for distribution in chronos_utils.SUPPORTED_DISTRIBUTIONS:
# Student t distribution has issues with the long tails
if (distribution != chronos_utils.StudentT_dist_code):
my_chronos = Chronos(max_iter=200, distribution=distribution)
# add dummies
my_chronos.add_regressors("dummy1")
my_chronos.add_regressors("dummy2")
my_chronos.fit(sample_data)
future_df = my_chronos.make_future_dataframe()
# Add dummies to future df
z = future_df.index.values
y = 0.01 * z + np.sin(z / 30)
future_df['y'] = y
dummy1 = 0.01 * z
dummy2 = np.sin(z / 30)
future_df['dummy1'] = dummy1
future_df['dummy2'] = dummy2
# Make predictions
predictions = my_chronos.predict(future_df)
# The predictions should be almost the same as the target
assert (np.mean(np.abs(predictions['y'] - predictions['yhat'])) <
0.1)
def test_predictions_not_nan(sample_data):
'''
Make sure no nans are returned during the prediction process
'''
for method in ["MAP", "MLE"]:
for distribution in chronos_utils.SUPPORTED_DISTRIBUTIONS:
my_chronos = Chronos(n_changepoints=0,
max_iter=100,
distribution=distribution,
method=method)
my_chronos.fit(sample_data)
future_df = my_chronos.make_future_dataframe(include_history=False)
predictions = my_chronos.predict(future_df)
predictions.drop('y', axis=1, inplace=True)
#print(predictions)
assert (predictions.isna().sum().sum() == 0)
def test_prediction_too_small_for_default_changepoints(sample_data):
'''
Test that when the size of the data is too small, the
number of changepoints gets adjusted
'''
for distribution in chronos_utils.SUPPORTED_DISTRIBUTIONS:
my_chronos = Chronos(max_iter=100,
n_changepoints=20,
distribution=distribution)
# This should raise a warning about the size of the data and changepoint number
with pytest.warns(RuntimeWarning):
my_chronos.fit(sample_data.iloc[:10])
future_df = my_chronos.make_future_dataframe()
predictions = my_chronos.predict(future_df)
assert (my_chronos._Chronos__number_of_changepoints <
sample_data.shape[0])
def test_plotting(capsys, monkeypatch, sample_data):
'''
Test the plotting of data when no history is provided.
This was an issue raised in issue #2.
'''
for distribution in chronos_utils.SUPPORTED_DISTRIBUTIONS:
my_chronos = Chronos(n_changepoints=0, max_iter=100, distribution=distribution)
my_chronos.fit(sample_data)
future_df = my_chronos.make_future_dataframe(include_history=True)
predictions = my_chronos.predict(future_df)
with pytest.warns(UserWarning):
chronos_plotting.plot_components(predictions, my_chronos, figure_name="test_prediction.png")
std_error = capsys.readouterr().err
assert(std_error == "")
os.remove("test_prediction.png")