def test_catch_warnings_percentage_change(self, warning_list=None):
"""
Test that two warnings are generated if the minimisation
does not result in a convergence. The first warning reports a that
the minimisation did not result in convergence, whilst the second
warning reports that the percentage change in the final iteration was
greater than the tolerated value.
The ensemble mean is the predictor.
"""
initial_guess = np.array([0, 1, 5000, 1], dtype=np.float64)
predictor = "mean"
distribution = "truncnorm"
plugin = Plugin(tolerance=self.tolerance, max_iterations=5)
plugin.process(
initial_guess,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
warning_msg_min = "Minimisation did not result in convergence after"
warning_msg_iter = "The final iteration resulted in a percentage "
self.assertTrue(
any(item.category == UserWarning for item in warning_list))
self.assertTrue(
any(warning_msg_min in str(item) for item in warning_list))
self.assertTrue(
any(warning_msg_iter in str(item) for item in warning_list))
def test_mean_predictor_point_by_point_sites(self):
"""
Test that the expected coefficients are generated when the ensemble
mean is the predictor for a normal distribution and coefficients are
calculated independently at each site location. The coefficients are
calculated by minimising the CRPS.
"""
forecast_spot_cube = self.historic_forecast_spot_cube.collapsed(
"realization", iris.analysis.MEAN)
forecast_var_spot_cube = forecast_spot_cube.copy()
forecast_var_spot_cube.data = forecast_var_spot_cube.data / 10.0
predictor = "mean"
distribution = "norm"
initial_guess = np.broadcast_to(
self.initial_guess_for_mean,
(
len(self.truth.coord(axis="y").points) *
len(self.truth.coord(axis="x").points),
len(self.initial_guess_for_mean),
),
)
plugin = Plugin(tolerance=self.tolerance, point_by_point=True)
result = plugin.process(
initial_guess,
forecast_spot_cube,
self.truth_spot_cube,
forecast_var_spot_cube,
predictor,
distribution,
)
self.assertEMOSCoefficientsAlmostEqual(
result, self.expected_mean_coefficients_point_by_point_sites)
def test_realizations_predictor_point_by_point(self):
"""
Test that the expected coefficients are generated when the ensemble
realizations are the predictor for a normal distribution and
coefficients are calculated independently at each grid point. The
coefficients are calculated by minimising the CRPS.
"""
predictor = "realizations"
distribution = "norm"
initial_guess = np.broadcast_to(
self.initial_guess_for_realization,
(
len(self.truth.coord(axis="y").points) *
len(self.truth.coord(axis="x").points),
len(self.initial_guess_for_realization),
),
)
# Use a larger value for the tolerance to terminate sooner to avoid
# minimising in computational noise.
plugin = Plugin(tolerance=0.01, point_by_point=True)
result = plugin.process(
initial_guess,
self.forecast_predictor_realizations,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertArrayAlmostEqual(
result,
self.expected_realizations_coefficients_point_by_point,
decimal=2)
def test_point_by_point_with_nans(self):
"""
Test that the expected coefficients are generated when the ensemble
mean is the predictor for a normal distribution and coefficients are
calculated independently at each grid point with one grid point
having NaN values for the truth.
"""
predictor = "mean"
distribution = "norm"
self.truth.data[:, 0, 0] = np.nan
plugin = Plugin(predictor,
tolerance=self.tolerance,
point_by_point=True)
result = plugin.process(
self.initial_guess_spot_mean,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
distribution,
)
self.expected_mean_coefficients_point_by_point[:, 0,
0] = self.initial_guess_for_mean
self.assertEMOSCoefficientsAlmostEqual(
result, self.expected_mean_coefficients_point_by_point)
def test_mean_predictor_point_by_point(self):
"""
Test that the expected coefficients are generated when the ensemble
mean is the predictor for a normal distribution and coefficients are
calculated independently at each grid point. The coefficients are
calculated by minimising the CRPS.
"""
predictor = "mean"
distribution = "norm"
initial_guess = np.broadcast_to(
self.initial_guess_for_mean,
(
len(self.truth.coord(axis="y").points) *
len(self.truth.coord(axis="x").points),
len(self.initial_guess_for_mean),
),
)
plugin = Plugin(tolerance=self.tolerance, point_by_point=True)
result = plugin.process(
initial_guess,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertEMOSCoefficientsAlmostEqual(
result, self.expected_mean_coefficients_point_by_point)
def test_realizations_predictor_max_iterations(self):
"""
Test that the plugin returns a list of coefficients
equal to specific values, when the ensemble realizations are the
predictor assuming a truncated normal distribution and the value
specified for the max_iterations is overridden. The coefficients are
calculated by minimising the CRPS and using a set default value for
the initial guess.
"""
predictor = "realizations"
max_iterations = 1000
distribution = "truncnorm"
plugin = Plugin(tolerance=self.tolerance,
max_iterations=max_iterations)
result = plugin.process(
self.initial_guess_for_realization,
self.forecast_predictor_realizations,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertEMOSCoefficientsAlmostEqual(
result, self.expected_realizations_coefficients)
def test_catch_warnings(self, warning_list=None):
"""
Test that a warning is generated if the minimisation
does not result in a convergence. The ensemble mean is the predictor.
"""
predictor = "mean"
distribution = "truncated_gaussian"
plugin = Plugin(tolerance=self.tolerance, max_iterations=10)
plugin.process(self.initial_guess_for_mean,
self.forecast_predictor_mean, self.truth,
self.forecast_variance, predictor, distribution)
warning_msg = "Minimisation did not result in convergence after"
self.assertTrue(
any(item.category == UserWarning for item in warning_list))
self.assertTrue(any(warning_msg in str(item) for item in warning_list))
def test_mean_predictor_point_by_point_sites_additional_predictor(self):
"""
Test that the plugin returns a numpy array with the expected
coefficients. Coefficients are calculated independently at each site.
The ensemble mean and altitude are the predictors.
"""
predictor = "mean"
distribution = "norm"
plugin = Plugin(predictor, tolerance=self.tolerance, point_by_point=True)
result = plugin.process(
self.ig_spot_mean_additional_predictor,
self.fp_additional_predictor_spot,
self.truth_spot_cube,
self.forecast_variance_spot,
distribution,
)
self.assertEMOSCoefficientsAlmostEqual(
result, self.expected_point_by_point_sites_additional_predictor
)
def test_mean_predictor_point_by_point_sites(self):
"""
Test that the expected coefficients are generated when the ensemble
mean is the predictor for a normal distribution and coefficients are
calculated independently at each site location. The coefficients are
calculated by minimising the CRPS.
"""
predictor = "mean"
distribution = "norm"
plugin = Plugin(predictor, tolerance=self.tolerance, point_by_point=True)
result = plugin.process(
self.initial_guess_spot_mean,
self.forecast_predictor_spot,
self.truth_spot_cube,
self.forecast_variance_spot,
distribution,
)
self.assertEMOSCoefficientsAlmostEqual(
result, self.expected_mean_coefficients_point_by_point_sites
)
def test_mean_predictor_max_iterations(self):
"""
Test that the plugin returns a list of coefficients
equal to specific values, when the ensemble mean is the predictor
assuming a normal distribution and the value specified for the
max_iterations is overridden. The coefficients are calculated by
minimising the CRPS.
"""
predictor = "mean"
max_iterations = 400
distribution = "norm"
plugin = Plugin(tolerance=self.tolerance,
max_iterations=max_iterations)
result = plugin.process(
self.initial_guess_for_mean,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertEMOSCoefficientsAlmostEqual(result,
self.expected_mean_coefficients)
class Test_process_truncated_normal_distribution(SetupTruncatedNormalInputs,
EnsembleCalibrationAssertions
):
"""
Test minimising the CRPS for a truncated normal distribution.
Either the ensemble mean or the individual ensemble realizations are used
as the predictors.
"""
def setUp(self):
"""Set up expected output."""
super().setUp()
self.tolerance = 1e-4
self.plugin = Plugin(tolerance=self.tolerance)
self.expected_mean_coefficients = [0.3958, 0.9854, -0.0, 0.621]
self.expected_realizations_coefficients = [
0.1898,
-0.1558,
0.4452,
0.8877,
-0.1331,
-0.0002,
]
@ManageWarnings(
ignored_messages=[
"Collapsing a non-contiguous coordinate.",
"The final iteration resulted in",
"invalid value encountered in",
"divide by zero encountered in",
],
warning_types=[
UserWarning, UserWarning, RuntimeWarning, RuntimeWarning
],
)
def test_basic_mean_predictor(self):
"""
Test that the plugin returns a numpy array. The ensemble mean
is the predictor.
"""
predictor = "mean"
distribution = "truncnorm"
result = self.plugin.process(
self.initial_guess_for_mean,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertIsInstance(result, np.ndarray)
self.assertEMOSCoefficientsAlmostEqual(result,
self.expected_mean_coefficients)
@ManageWarnings(
ignored_messages=[
"Collapsing a non-contiguous coordinate.",
"Minimisation did not result in convergence",
"invalid value encountered in",
"divide by zero encountered in",
],
warning_types=[
UserWarning, UserWarning, RuntimeWarning, RuntimeWarning
],
)
def test_basic_realizations_predictor(self):
"""
Test that the plugin returns a numpy array with the expected
coefficients. The ensemble realizations are the predictor.
"""
predictor = "realizations"
distribution = "truncnorm"
result = self.plugin.process(
self.initial_guess_for_realization,
self.forecast_predictor_realizations,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertIsInstance(result, np.ndarray)
self.assertEMOSCoefficientsAlmostEqual(
result, self.expected_realizations_coefficients)
@ManageWarnings(
ignored_messages=["Collapsing a non-contiguous coordinate."])
def test_mean_predictor_keyerror(self):
"""
Test that an exception is raised when the distribution requested is
not an available option when the predictor is the
ensemble mean.
"""
predictor = "mean"
distribution = "foo"
msg = "Distribution requested"
with self.assertRaisesRegex(KeyError, msg):
self.plugin.process(
self.initial_guess_for_mean,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
@ManageWarnings(
ignored_messages=[
"Collapsing a non-contiguous coordinate.",
"Minimisation did not result in convergence",
"The final iteration resulted in",
"invalid value encountered in",
"divide by zero encountered in",
],
warning_types=[
UserWarning,
UserWarning,
UserWarning,
RuntimeWarning,
RuntimeWarning,
],
)
def test_mean_predictor_max_iterations(self):
"""
Test that the plugin returns a list of coefficients
equal to specific values, when the ensemble mean is the predictor
assuming a truncated normal distribution and the value specified
for the max_iterations is overridden. The coefficients are
calculated by minimising the CRPS.
"""
predictor = "mean"
max_iterations = 400
distribution = "truncnorm"
plugin = Plugin(tolerance=self.tolerance,
max_iterations=max_iterations)
result = plugin.process(
self.initial_guess_for_mean,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertEMOSCoefficientsAlmostEqual(result,
self.expected_mean_coefficients)
@ManageWarnings(
ignored_messages=[
"Collapsing a non-contiguous coordinate.",
"Minimisation did not result in convergence",
"invalid value encountered in",
"divide by zero encountered in",
],
warning_types=[
UserWarning, UserWarning, RuntimeWarning, RuntimeWarning
],
)
def test_realizations_predictor_max_iterations(self):
"""
Test that the plugin returns a list of coefficients
equal to specific values, when the ensemble realizations are the
predictor assuming a truncated normal distribution and the value
specified for the max_iterations is overridden. The coefficients are
calculated by minimising the CRPS.
"""
predictor = "realizations"
max_iterations = 1000
distribution = "truncnorm"
plugin = Plugin(tolerance=self.tolerance,
max_iterations=max_iterations)
result = plugin.process(
self.initial_guess_for_realization,
self.forecast_predictor_realizations,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertEMOSCoefficientsAlmostEqual(
result, self.expected_realizations_coefficients)
@ManageWarnings(
record=True,
ignored_messages=["Collapsing a non-contiguous coordinate."])
def test_catch_warnings(self, warning_list=None):
"""
Test that a warning is generated if the minimisation
does not result in a convergence. The ensemble mean is the predictor.
"""
predictor = "mean"
distribution = "truncnorm"
plugin = Plugin(tolerance=self.tolerance, max_iterations=10)
plugin.process(
self.initial_guess_for_mean,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
warning_msg = "Minimisation did not result in convergence after"
self.assertTrue(
any(item.category == UserWarning for item in warning_list))
self.assertTrue(any(warning_msg in str(item) for item in warning_list))
@ManageWarnings(
record=True,
ignored_messages=["Collapsing a non-contiguous coordinate."])
def test_catch_warnings_percentage_change(self, warning_list=None):
"""
Test that two warnings are generated if the minimisation
does not result in a convergence. The first warning reports a that
the minimisation did not result in convergence, whilst the second
warning reports that the percentage change in the final iteration was
greater than the tolerated value.
The ensemble mean is the predictor.
"""
initial_guess = np.array([0, 1, 5000, 1], dtype=np.float64)
predictor = "mean"
distribution = "truncnorm"
plugin = Plugin(tolerance=self.tolerance, max_iterations=5)
plugin.process(
initial_guess,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
warning_msg_min = "Minimisation did not result in convergence after"
warning_msg_iter = "The final iteration resulted in a percentage "
self.assertTrue(
any(item.category == UserWarning for item in warning_list))
self.assertTrue(
any(warning_msg_min in str(item) for item in warning_list))
self.assertTrue(
any(warning_msg_iter in str(item) for item in warning_list))
class Test_process_normal_distribution(SetupNormalInputs,
EnsembleCalibrationAssertions):
"""
Test minimising the CRPS for a normal distribution.
Either the ensemble mean or the individual ensemble realizations are used
as the predictors.
"""
def setUp(self):
"""Set up expected output.
The coefficients are in the order [alpha, beta, gamma, delta].
"""
super().setUp()
self.tolerance = 1e-4
self.plugin = Plugin(tolerance=self.tolerance)
self.expected_mean_coefficients = [-0.0003, 1.0013, 0.0012, 0.5945]
self.expected_realizations_coefficients = [
0.0254,
0.4349,
0.39,
0.8122,
-0.0016,
0.2724,
]
self.expected_mean_coefficients_point_by_point = np.array(
[
[
[0.0015, 0.0037, -0.002],
[-0.0009, 0.0008, 0.0015],
[-0.0046, 0.0053, -0.0038],
],
[
[1.0039, 1.0035, 1.0009],
[1.0013, 1.0011, 1.001],
[1.002, 1.0015, 1.0008],
],
[
[0.0017, -0.0009, -0.0002],
[0.0054, 0.0003, -0.0002],
[-0.0001, -0.0018, 0.0002],
],
[
[-0.0, 0.0007, -0.0009],
[0.0003, -0.0001, -0.001],
[-0.0013, 0.0, 0.0006],
],
],
dtype=np.float32,
)
self.expected_mean_coefficients_point_by_point_sites = np.array(
[
[0.0017, 0.0017, 0.0017, 0.0017],
[1.0036, 1.0036, 1.0036, 1.0036],
[0.0017, 0.0017, 0.0017, 0.0017],
[-0.0, -0.0, -0.0, 0.0],
],
dtype=np.float32,
)
self.expected_realizations_coefficients_point_by_point = np.array(
[
[
[0.0001, 0.0001, 0.0001],
[0.0001, 0.0001, 0.0],
[0.0001, 0.0001, 0.0001],
],
[
[0.579, 0.5793, 0.5782],
[0.5782, 0.5778, 0.5781],
[0.5786, 0.5782, 0.5783],
],
[
[0.5795, 0.5786, 0.5782],
[0.5783, 0.578, 0.5767],
[0.5791, 0.578, 0.5763],
],
[
[0.5773, 0.5769, 0.5763],
[0.5769, 0.5771, 0.5782],
[0.5764, 0.5773, 0.5783],
],
[
[0.0001, 0.0001, 0.0001],
[0.0001, 0.0001, 0.0001],
[0.0001, 0.0001, 0.0],
],
[
[1.0194, 1.0143, 1.0199],
[1.0199, 1.02, 1.013],
[1.0144, 0.9885, 1.0246],
],
],
dtype=np.float32,
)
@ManageWarnings(
ignored_messages=[
"Collapsing a non-contiguous coordinate.",
"Minimisation did not result in convergence",
"divide by zero encountered in",
],
warning_types=[UserWarning, UserWarning, RuntimeWarning],
)
def test_basic_mean_predictor(self):
"""
Test that the plugin returns a numpy array with the expected
coefficients. The ensemble mean is the predictor.
"""
predictor = "mean"
distribution = "norm"
result = self.plugin.process(
self.initial_guess_for_mean,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertIsInstance(result, np.ndarray)
self.assertEqual(result.dtype, np.float32)
self.assertEMOSCoefficientsAlmostEqual(result,
self.expected_mean_coefficients)
@ManageWarnings(
ignored_messages=[
"Collapsing a non-contiguous coordinate.",
"Minimisation did not result in convergence",
"divide by zero encountered in",
"invalid value encountered in",
],
warning_types=[
UserWarning, UserWarning, RuntimeWarning, RuntimeWarning
],
)
def test_basic_realizations_predictor(self):
"""
Test that the plugin returns a numpy array with the expected
coefficients. The ensemble realizations are the predictor.
"""
predictor = "realizations"
distribution = "norm"
result = self.plugin.process(
self.initial_guess_for_realization,
self.forecast_predictor_realizations,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertIsInstance(result, np.ndarray)
self.assertEqual(result.dtype, np.float32)
self.assertEMOSCoefficientsAlmostEqual(
result, self.expected_realizations_coefficients)
@ManageWarnings(
ignored_messages=["Collapsing a non-contiguous coordinate."])
def test_mean_predictor_keyerror(self):
"""
Test that the minimisation has resulted in a KeyError, if the
distribution that has been requested was not within the dictionary
containing the minimisation functions.
"""
predictor = "mean"
distribution = "foo"
msg = "Distribution requested"
with self.assertRaisesRegex(KeyError, msg):
self.plugin.process(
self.initial_guess_for_mean,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
@ManageWarnings(
ignored_messages=[
"Collapsing a non-contiguous coordinate.",
"Minimisation did not result in convergence",
"divide by zero encountered in",
],
warning_types=[UserWarning, UserWarning, RuntimeWarning],
)
def test_mean_predictor_max_iterations(self):
"""
Test that the plugin returns a list of coefficients
equal to specific values, when the ensemble mean is the predictor
assuming a normal distribution and the value specified for the
max_iterations is overridden. The coefficients are calculated by
minimising the CRPS.
"""
predictor = "mean"
max_iterations = 400
distribution = "norm"
plugin = Plugin(tolerance=self.tolerance,
max_iterations=max_iterations)
result = plugin.process(
self.initial_guess_for_mean,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertEMOSCoefficientsAlmostEqual(result,
self.expected_mean_coefficients)
@ManageWarnings(
ignored_messages=[
"Collapsing a non-contiguous coordinate.",
"Minimisation did not result in convergence",
"divide by zero encountered in",
"invalid value encountered in",
],
warning_types=[
UserWarning, UserWarning, RuntimeWarning, RuntimeWarning
],
)
def test_realizations_predictor_max_iterations(self):
"""
Test that the plugin returns a list of coefficients
equal to specific values, when the ensemble realizations are the
predictor assuming a truncated normal distribution and the value
specified for the MAX_ITERATIONS is overridden. The coefficients are
calculated by minimising the CRPS.
"""
predictor = "realizations"
max_iterations = 1000
distribution = "norm"
plugin = Plugin(tolerance=self.tolerance,
max_iterations=max_iterations)
result = plugin.process(
self.initial_guess_for_realization,
self.forecast_predictor_realizations,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertEMOSCoefficientsAlmostEqual(
result, self.expected_realizations_coefficients)
@ManageWarnings(
ignored_messages=[
"Collapsing a non-contiguous coordinate.",
"Minimisation did not result in convergence",
"divide by zero encountered in",
],
warning_types=[UserWarning, UserWarning, RuntimeWarning],
)
def test_mean_predictor_point_by_point(self):
"""
Test that the expected coefficients are generated when the ensemble
mean is the predictor for a normal distribution and coefficients are
calculated independently at each grid point. The coefficients are
calculated by minimising the CRPS.
"""
predictor = "mean"
distribution = "norm"
initial_guess = np.broadcast_to(
self.initial_guess_for_mean,
(
len(self.truth.coord(axis="y").points) *
len(self.truth.coord(axis="x").points),
len(self.initial_guess_for_mean),
),
)
plugin = Plugin(tolerance=self.tolerance, point_by_point=True)
result = plugin.process(
initial_guess,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertEMOSCoefficientsAlmostEqual(
result, self.expected_mean_coefficients_point_by_point)
@ManageWarnings(
ignored_messages=[
"Collapsing a non-contiguous coordinate.",
"Minimisation did not result in convergence",
"divide by zero encountered in",
],
warning_types=[UserWarning, UserWarning, RuntimeWarning],
)
def test_mean_predictor_point_by_point_sites(self):
"""
Test that the expected coefficients are generated when the ensemble
mean is the predictor for a normal distribution and coefficients are
calculated independently at each site location. The coefficients are
calculated by minimising the CRPS.
"""
forecast_spot_cube = self.historic_forecast_spot_cube.collapsed(
"realization", iris.analysis.MEAN)
forecast_var_spot_cube = forecast_spot_cube.copy()
forecast_var_spot_cube.data = forecast_var_spot_cube.data / 10.0
predictor = "mean"
distribution = "norm"
initial_guess = np.broadcast_to(
self.initial_guess_for_mean,
(
len(self.truth.coord(axis="y").points) *
len(self.truth.coord(axis="x").points),
len(self.initial_guess_for_mean),
),
)
plugin = Plugin(tolerance=self.tolerance, point_by_point=True)
result = plugin.process(
initial_guess,
forecast_spot_cube,
self.truth_spot_cube,
forecast_var_spot_cube,
predictor,
distribution,
)
self.assertEMOSCoefficientsAlmostEqual(
result, self.expected_mean_coefficients_point_by_point_sites)
@ManageWarnings(
ignored_messages=[
"Collapsing a non-contiguous coordinate.",
"Minimisation did not result in convergence",
"divide by zero encountered in",
"invalid value encountered in",
],
warning_types=[
UserWarning, UserWarning, RuntimeWarning, RuntimeWarning
],
)
def test_realizations_predictor_point_by_point(self):
"""
Test that the expected coefficients are generated when the ensemble
realizations are the predictor for a normal distribution and
coefficients are calculated independently at each grid point. The
coefficients are calculated by minimising the CRPS.
"""
predictor = "realizations"
distribution = "norm"
initial_guess = np.broadcast_to(
self.initial_guess_for_realization,
(
len(self.truth.coord(axis="y").points) *
len(self.truth.coord(axis="x").points),
len(self.initial_guess_for_realization),
),
)
# Use a larger value for the tolerance to terminate sooner to avoid
# minimising in computational noise.
plugin = Plugin(tolerance=0.01, point_by_point=True)
result = plugin.process(
initial_guess,
self.forecast_predictor_realizations,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertArrayAlmostEqual(
result,
self.expected_realizations_coefficients_point_by_point,
decimal=2)
@ManageWarnings(
ignored_messages=[
"Collapsing a non-contiguous coordinate.",
"Minimisation did not result in convergence",
"divide by zero encountered in",
],
warning_types=[UserWarning, UserWarning, RuntimeWarning],
)
@ManageWarnings(
record=True,
ignored_messages=["Collapsing a non-contiguous coordinate."])
def test_catch_warnings(self, warning_list=None):
"""
Test that a warning is generated if the minimisation
does not result in a convergence. The ensemble mean is the predictor.
"""
predictor = "mean"
distribution = "norm"
plugin = Plugin(tolerance=self.tolerance, max_iterations=10)
plugin.process(
self.initial_guess_for_mean,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
warning_msg = "Minimisation did not result in convergence after"
self.assertTrue(
any(item.category == UserWarning for item in warning_list))
self.assertTrue(any(warning_msg in str(item) for item in warning_list))
@ManageWarnings(
record=True,
ignored_messages=["Collapsing a non-contiguous coordinate."])
def test_catch_warnings_percentage_change(self, warning_list=None):
"""
Test that two warnings are generated if the minimisation
does not result in a convergence. The first warning reports a that
the minimisation did not result in convergence, whilst the second
warning reports that the percentage change in the final iteration was
greater than the tolerated value. The ensemble mean is the predictor.
"""
initial_guess = np.array([5000, 1, 0, 1], dtype=np.float64)
predictor = "mean"
distribution = "norm"
plugin = Plugin(tolerance=self.tolerance, max_iterations=5)
plugin.process(
initial_guess,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
warning_msg_min = "Minimisation did not result in convergence after"
warning_msg_iter = "The final iteration resulted in a percentage "
self.assertTrue(
any(item.category == UserWarning for item in warning_list))
self.assertTrue(
any(warning_msg_min in str(item) for item in warning_list))
self.assertTrue(
any(warning_msg_iter in str(item) for item in warning_list))
class Test_process_normal_distribution(SetupNormalInputs,
EnsembleCalibrationAssertions):
"""
Test minimising the CRPS for a normal distribution.
Either the ensemble mean or the individual ensemble realizations are used
as the predictors.
"""
def setUp(self):
"""Set up expected output.
The coefficients are in the order [alpha, beta, gamma, delta].
"""
super().setUp()
self.tolerance = 1e-4
self.plugin = Plugin(tolerance=self.tolerance)
self.expected_mean_coefficients = [-0.0008, 1.0009, 0.0023, 0.8070]
self.expected_realizations_coefficients = [
0.0427,
0.4117,
0.1946,
0.8907,
-0.1435,
0.037,
]
@ManageWarnings(
ignored_messages=[
"Collapsing a non-contiguous coordinate.",
"Minimisation did not result in convergence",
"divide by zero encountered in",
],
warning_types=[UserWarning, UserWarning, RuntimeWarning],
)
def test_basic_mean_predictor(self):
"""
Test that the plugin returns a numpy array with the expected
coefficients. The ensemble mean is the predictor.
"""
predictor = "mean"
distribution = "norm"
result = self.plugin.process(
self.initial_guess_for_mean,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertIsInstance(result, np.ndarray)
self.assertEqual(result.dtype, np.float32)
self.assertEMOSCoefficientsAlmostEqual(result,
self.expected_mean_coefficients)
@ManageWarnings(
ignored_messages=[
"Collapsing a non-contiguous coordinate.",
"Minimisation did not result in convergence",
"divide by zero encountered in",
"invalid value encountered in",
],
warning_types=[
UserWarning, UserWarning, RuntimeWarning, RuntimeWarning
],
)
def test_basic_realizations_predictor(self):
"""
Test that the plugin returns a numpy array with the expected
coefficients. The ensemble realizations are the predictor.
"""
predictor = "realizations"
distribution = "norm"
result = self.plugin.process(
self.initial_guess_for_realization,
self.forecast_predictor_realizations,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertIsInstance(result, np.ndarray)
self.assertEqual(result.dtype, np.float32)
self.assertEMOSCoefficientsAlmostEqual(
result, self.expected_realizations_coefficients)
@ManageWarnings(
ignored_messages=["Collapsing a non-contiguous coordinate."])
def test_mean_predictor_keyerror(self):
"""
Test that the minimisation has resulted in a KeyError, if the
distribution that has been requested was not within the dictionary
containing the minimisation functions.
"""
predictor = "mean"
distribution = "foo"
msg = "Distribution requested"
with self.assertRaisesRegex(KeyError, msg):
self.plugin.process(
self.initial_guess_for_mean,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
@ManageWarnings(
ignored_messages=[
"Collapsing a non-contiguous coordinate.",
"Minimisation did not result in convergence",
"divide by zero encountered in",
],
warning_types=[UserWarning, UserWarning, RuntimeWarning],
)
def test_mean_predictor_max_iterations(self):
"""
Test that the plugin returns a list of coefficients
equal to specific values, when the ensemble mean is the predictor
assuming a normal distribution and the value specified for the
max_iterations is overridden. The coefficients are calculated by
minimising the CRPS and using a set default value for the
initial guess.
"""
predictor = "mean"
max_iterations = 400
distribution = "norm"
plugin = Plugin(tolerance=self.tolerance,
max_iterations=max_iterations)
result = plugin.process(
self.initial_guess_for_mean,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertEMOSCoefficientsAlmostEqual(result,
self.expected_mean_coefficients)
@ManageWarnings(
ignored_messages=[
"Collapsing a non-contiguous coordinate.",
"Minimisation did not result in convergence",
"divide by zero encountered in",
"invalid value encountered in",
],
warning_types=[
UserWarning, UserWarning, RuntimeWarning, RuntimeWarning
],
)
def test_realizations_predictor_max_iterations(self):
"""
Test that the plugin returns a list of coefficients
equal to specific values, when the ensemble realizations are the
predictor assuming a truncated normal distribution and the value
specified for the MAX_ITERATIONS is overridden. The coefficients are
calculated by minimising the CRPS and using a set default value for
the initial guess.
"""
predictor = "realizations"
max_iterations = 1000
distribution = "norm"
plugin = Plugin(tolerance=self.tolerance,
max_iterations=max_iterations)
result = plugin.process(
self.initial_guess_for_realization,
self.forecast_predictor_realizations,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
self.assertEMOSCoefficientsAlmostEqual(
result, self.expected_realizations_coefficients)
@ManageWarnings(
record=True,
ignored_messages=["Collapsing a non-contiguous coordinate."])
def test_catch_warnings(self, warning_list=None):
"""
Test that a warning is generated if the minimisation
does not result in a convergence. The ensemble mean is the predictor.
"""
predictor = "mean"
distribution = "norm"
plugin = Plugin(tolerance=self.tolerance, max_iterations=10)
plugin.process(
self.initial_guess_for_mean,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
warning_msg = "Minimisation did not result in convergence after"
self.assertTrue(
any(item.category == UserWarning for item in warning_list))
self.assertTrue(any(warning_msg in str(item) for item in warning_list))
@ManageWarnings(
record=True,
ignored_messages=["Collapsing a non-contiguous coordinate."])
def test_catch_warnings_percentage_change(self, warning_list=None):
"""
Test that two warnings are generated if the minimisation
does not result in a convergence. The first warning reports a that
the minimisation did not result in convergence, whilst the second
warning reports that the percentage change in the final iteration was
greater than the tolerated value. The ensemble mean is the predictor.
"""
initial_guess = np.array([5000, 1, 0, 1], dtype=np.float64)
predictor = "mean"
distribution = "norm"
plugin = Plugin(tolerance=self.tolerance, max_iterations=5)
plugin.process(
initial_guess,
self.forecast_predictor_mean,
self.truth,
self.forecast_variance,
predictor,
distribution,
)
warning_msg_min = "Minimisation did not result in convergence after"
warning_msg_iter = "The final iteration resulted in a percentage "
self.assertTrue(
any(item.category == UserWarning for item in warning_list))
self.assertTrue(
any(warning_msg_min in str(item) for item in warning_list))
self.assertTrue(
any(warning_msg_iter in str(item) for item in warning_list))
