def test_combinatory_matcher_n2():
n = 1000
x = np.arange(n)
y = np.arange(n) * 0.5
index = pd.date_range(start="2000-01-01", periods=n, freq="D")
df = pd.DataFrame({'x': x, 'y': y}, columns=['x', 'y'], index=index)
df2 = pd.DataFrame({'x': x, 'y': y}, columns=['x', 'y'], index=index)
df3 = pd.DataFrame({'x': x, 'y': y}, columns=['x', 'y'], index=index)
df_dict = {'data1': df, 'data2': df2, 'data3': df3}
temp_matcher = temporal_matchers.BasicTemporalMatching()
matched = temp_matcher.combinatory_matcher(df_dict, 'data1')
assert sorted(list(matched)) == sorted([('data1', 'data2'),
('data1', 'data3')])
assert sorted(list(matched[('data1',
'data2')].columns)) == sorted([('data1', 'x'),
('data1', 'y'),
('data2', 'x'),
('data2', 'y')])
assert sorted(list(matched[('data1',
'data3')].columns)) == sorted([('data1', 'x'),
('data1', 'y'),
('data3', 'x'),
('data3', 'y')])
def test_validation_n3_k2_temporal_matching_no_matches():
tst_results = {}
datasets = setup_two_without_overlap()
dm = DataManager(
datasets,
"DS1",
read_ts_names={d: "read" for d in ["DS1", "DS2", "DS3"]},
)
process = Validation(
dm,
"DS1",
temporal_matcher=temporal_matchers.BasicTemporalMatching(
window=1 / 24.0
).combinatory_matcher,
scaling="lin_cdf_match",
metrics_calculators={
(3, 2): metrics_calculators.BasicMetrics(
other_name="k1"
).calc_metrics
},
)
jobs = process.get_processing_jobs()
for job in jobs:
results = process.calc(*job)
assert sorted(list(results)) == sorted(list(tst_results))
def test_validation_error_n2_k2():
datasets = setup_TestDatasets()
dm = DataManager(
datasets,
"DS1",
read_ts_names={d: "read" for d in ["DS1", "DS2", "DS3"]},
)
# n less than number of datasets is no longer allowed
with pytest.raises(ValueError):
Validation(
dm,
"DS1",
temporal_matcher=temporal_matchers.BasicTemporalMatching(
window=1 / 24.0
).combinatory_matcher,
scaling="lin_cdf_match",
metrics_calculators={
(2, 2): metrics_calculators.BasicMetrics(
other_name="k1"
).calc_metrics
},
)
def test_validation_n3_k2_temporal_matching_no_matches2():
tst_results = {
(("DS1", "x"), ("DS3", "y")): {
"n_obs": np.array([1000], dtype=np.int32),
"tau": np.array([np.nan], dtype=np.float32),
"gpi": np.array([4], dtype=np.int32),
"RMSD": np.array([0.0], dtype=np.float32),
"lon": np.array([4.0]),
"p_tau": np.array([np.nan], dtype=np.float32),
"BIAS": np.array([0.0], dtype=np.float32),
"p_rho": np.array([0.0], dtype=np.float32),
"rho": np.array([1.0], dtype=np.float32),
"lat": np.array([4.0]),
"R": np.array([1.0], dtype=np.float32),
"p_R": np.array([0.0], dtype=np.float32),
},
(("DS1", "x"), ("DS3", "x")): {
"n_obs": np.array([1000], dtype=np.int32),
"tau": np.array([np.nan], dtype=np.float32),
"gpi": np.array([4], dtype=np.int32),
"RMSD": np.array([0.0], dtype=np.float32),
"lon": np.array([4.0]),
"p_tau": np.array([np.nan], dtype=np.float32),
"BIAS": np.array([0.0], dtype=np.float32),
"p_rho": np.array([0.0], dtype=np.float32),
"rho": np.array([1.0], dtype=np.float32),
"lat": np.array([4.0]),
"R": np.array([1.0], dtype=np.float32),
"p_R": np.array([0.0], dtype=np.float32),
},
}
datasets = setup_three_with_two_overlapping()
dm = DataManager(
datasets,
"DS1",
read_ts_names={d: "read" for d in ["DS1", "DS2", "DS3"]},
)
process = Validation(
dm,
"DS1",
temporal_matcher=temporal_matchers.BasicTemporalMatching(
window=1 / 24.0
).combinatory_matcher,
scaling="lin_cdf_match",
metrics_calculators={
(3, 2): metrics_calculators.BasicMetrics(
other_name="k1"
).calc_metrics
},
)
jobs = process.get_processing_jobs()
for job in jobs:
results = process.calc(*job)
assert sorted(list(results)) == sorted(list(tst_results))
def test_validation_n3_k2():
tst_results = {
(('DS1', 'x'), ('DS3', 'y')): {
'n_obs': np.array([1000], dtype=np.int32),
'tau': np.array([np.nan], dtype=np.float32),
'gpi': np.array([4], dtype=np.int32),
'RMSD': np.array([0.], dtype=np.float32),
'lon': np.array([4.]),
'p_tau': np.array([np.nan], dtype=np.float32),
'BIAS': np.array([0.], dtype=np.float32),
'p_rho': np.array([0.], dtype=np.float32),
'rho': np.array([1.], dtype=np.float32),
'lat': np.array([4.]),
'R': np.array([1.], dtype=np.float32),
'p_R': np.array([0.], dtype=np.float32)},
(('DS1', 'x'), ('DS2', 'y')): {
'n_obs': np.array([1000], dtype=np.int32),
'tau': np.array([np.nan], dtype=np.float32),
'gpi': np.array([4], dtype=np.int32),
'RMSD': np.array([0.], dtype=np.float32),
'lon': np.array([4.]),
'p_tau': np.array([np.nan], dtype=np.float32),
'BIAS': np.array([0.], dtype=np.float32),
'p_rho': np.array([0.], dtype=np.float32),
'rho': np.array([1.], dtype=np.float32),
'lat': np.array([4.]),
'R': np.array([1.], dtype=np.float32),
'p_R': np.array([0.], dtype=np.float32)},
(('DS1', 'x'), ('DS3', 'x')): {
'n_obs': np.array([1000], dtype=np.int32),
'tau': np.array([np.nan], dtype=np.float32),
'gpi': np.array([4], dtype=np.int32),
'RMSD': np.array([0.], dtype=np.float32),
'lon': np.array([4.]),
'p_tau': np.array([np.nan], dtype=np.float32),
'BIAS': np.array([0.], dtype=np.float32),
'p_rho': np.array([0.], dtype=np.float32),
'rho': np.array([1.], dtype=np.float32),
'lat': np.array([4.]),
'R': np.array([1.], dtype=np.float32),
'p_R': np.array([0.], dtype=np.float32)}}
datasets = setup_TestDatasets()
process = Validation(
datasets, 'DS1',
temporal_matcher=temporal_matchers.BasicTemporalMatching(
window=1 / 24.0).combinatory_matcher,
scaling='lin_cdf_match',
metrics_calculators={
(3, 2): metrics_calculators.BasicMetrics(other_name='k1').calc_metrics})
jobs = process.get_processing_jobs()
for job in jobs:
results = process.calc(*job)
assert sorted(list(results)) == sorted(list(tst_results))
def test_validation_n3_k2_masking_no_data_remains():
datasets = setup_TestDatasets()
# setup masking datasets
grid = grids.CellGrid(np.array([1, 2, 3, 4]), np.array([1, 2, 3, 4]),
np.array([4, 4, 2, 1]), gpis=np.array([1, 2, 3, 4]))
mds1 = GriddedTsBase("", grid, MaskingTestDataset)
mds2 = GriddedTsBase("", grid, MaskingTestDataset)
mds = {
'masking1': {
'class': mds1,
'columns': ['x'],
'args': [],
'kwargs': {'limit': 500},
'use_lut': False,
'grids_compatible': True},
'masking2': {
'class': mds2,
'columns': ['x'],
'args': [],
'kwargs': {'limit': 1000},
'use_lut': False,
'grids_compatible': True}
}
process = Validation(
datasets, 'DS1',
temporal_matcher=temporal_matchers.BasicTemporalMatching(
window=1 / 24.0).combinatory_matcher,
scaling='lin_cdf_match',
metrics_calculators={
(3, 2): metrics_calculators.BasicMetrics(other_name='k1').calc_metrics},
masking_datasets=mds)
gpi_info = (1, 1, 1)
ref_df = datasets['DS1']['class'].read(1)
with warnings.catch_warnings():
warnings.filterwarnings('ignore', category=DeprecationWarning)
new_ref_df = process.mask_dataset(ref_df, gpi_info)
assert len(new_ref_df) == 0
nptest.assert_allclose(new_ref_df.x.values, np.arange(1000, 1000))
jobs = process.get_processing_jobs()
for job in jobs:
with warnings.catch_warnings():
warnings.filterwarnings('ignore', category=DeprecationWarning)
results = process.calc(*job)
tst = []
assert sorted(list(results)) == sorted(list(tst))
for key, tst_key in zip(sorted(results),
sorted(tst)):
nptest.assert_almost_equal(results[key]['n_obs'],
tst[tst_key]['n_obs'])
def __init__(self,
datasets,
spatial_ref,
metrics_calculators,
temporal_matcher=None,
temporal_window=1 / 24.0,
temporal_ref=None,
masking_datasets=None,
period=None,
scaling='lin_cdf_match',
scaling_ref=None):
if type(datasets) is DataManager:
self.data_manager = datasets
else:
self.data_manager = DataManager(datasets, spatial_ref, period)
self.temp_matching = temporal_matcher
if self.temp_matching is None:
self.temp_matching = temporal_matchers.BasicTemporalMatching(
window=temporal_window).combinatory_matcher
self.temporal_ref = temporal_ref
if self.temporal_ref is None:
self.temporal_ref = self.data_manager.reference_name
self.metrics_c = metrics_calculators
for n, k in self.metrics_c:
if n < len(self.data_manager.datasets.keys()):
raise ValueError('n must be equal to the number of datasets')
self.masking_dm = None
if masking_datasets is not None:
# add temporal reference dataset to the masking datasets since it
# is necessary for temporally matching the masking datasets to the
# common time stamps. Use _reference here to make a clash with the
# names of the masking datasets unlikely
masking_datasets.update(
{'_reference': datasets[self.temporal_ref]})
self.masking_dm = DataManager(masking_datasets,
'_reference',
period=period)
if type(scaling) == str:
self.scaling = DefaultScaler(scaling)
else:
self.scaling = scaling
self.scaling_ref = scaling_ref
if self.scaling_ref is None:
self.scaling_ref = self.data_manager.reference_name
self.luts = self.data_manager.get_luts()
def __init__(self, datasets, spatial_ref, metrics_calculators,
temporal_matcher=None, temporal_window=1 / 24.0,
temporal_ref=None,
masking_datasets=None,
period=None,
scaling='lin_cdf_match', scaling_ref=None):
if isinstance(datasets, DataManager):
self.data_manager = datasets
else:
self.data_manager = DataManager(datasets, spatial_ref, period)
self.temp_matching = temporal_matcher
if self.temp_matching is None:
warnings.warn(
"You are using the default temporal matcher. If you are using one of the"
" newer metric calculators (PairwiseIntercomparisonMetrics,"
" TripleCollocationMetrics) you should probably use `make_combined_temporal_matcher`"
" instead. Have a look at the documentation of the metric calculators for more info."
)
self.temp_matching = temporal_matchers.BasicTemporalMatching(
window=temporal_window).combinatory_matcher
self.temporal_ref = temporal_ref
if self.temporal_ref is None:
self.temporal_ref = self.data_manager.reference_name
self.metrics_c = metrics_calculators
for n, k in self.metrics_c:
if n < len(self.data_manager.datasets.keys()):
raise ValueError('n must be equal to the number of datasets')
self.masking_dm = None
if masking_datasets is not None:
# add temporal reference dataset to the masking datasets since it
# is necessary for temporally matching the masking datasets to the
# common time stamps. Use _reference here to make a clash with the
# names of the masking datasets unlikely
masking_datasets.update(
{'_reference': datasets[self.temporal_ref]})
self.masking_dm = DataManager(masking_datasets, '_reference',
period=period)
if type(scaling) == str:
self.scaling = DefaultScaler(scaling)
else:
self.scaling = scaling
self.scaling_ref = scaling_ref
if self.scaling_ref is None:
self.scaling_ref = self.data_manager.reference_name
self.luts = self.data_manager.get_luts()
def test_validation_error_n2_k2():
datasets = setup_TestDatasets()
dm = DataManager(datasets, 'DS1', read_ts_names={d: 'read' for d in ['DS1', 'DS2', 'DS3']})
# n less than number of datasets is no longer allowed
with pytest.raises(ValueError):
process = Validation(
dm, 'DS1',
temporal_matcher=temporal_matchers.BasicTemporalMatching(
window=1 / 24.0).combinatory_matcher,
scaling='lin_cdf_match',
metrics_calculators={
(2, 2): metrics_calculators.BasicMetrics(other_name='k1').calc_metrics})
def test_validation_n2_k2_temporal_matching_no_matches():
tst_results = {}
datasets = setup_two_without_overlap()
process = Validation(
datasets, 'DS1',
temporal_matcher=temporal_matchers.BasicTemporalMatching(
window=1 / 24.0).combinatory_matcher,
scaling='lin_cdf_match',
metrics_calculators={
(2, 2): metrics_calculators.BasicMetrics(other_name='k1').calc_metrics})
jobs = process.get_processing_jobs()
for job in jobs:
results = process.calc(*job)
assert sorted(list(results)) == sorted(list(tst_results))
def test_validation_n3_k2_masking():
# test result for one gpi in a cell
tst_results_one = {
(('DS1', 'x'), ('DS3', 'y')): {
'n_obs': np.array([250], dtype=np.int32)},
(('DS1', 'x'), ('DS2', 'y')): {
'n_obs': np.array([250], dtype=np.int32)},
(('DS1', 'x'), ('DS3', 'x')): {
'n_obs': np.array([250], dtype=np.int32)}}
# test result for two gpis in a cell
tst_results_two = {
(('DS1', 'x'), ('DS3', 'y')): {
'n_obs': np.array([250, 250], dtype=np.int32)},
(('DS1', 'x'), ('DS2', 'y')): {
'n_obs': np.array([250, 250], dtype=np.int32)},
(('DS1', 'x'), ('DS3', 'x')): {
'n_obs': np.array([250, 250], dtype=np.int32)}}
# cell 4 in this example has two gpis so it returns different results.
tst_results = {1: tst_results_one,
1: tst_results_one,
2: tst_results_two}
datasets = setup_TestDatasets()
# setup masking datasets
grid = grids.CellGrid(np.array([1, 2, 3, 4]), np.array([1, 2, 3, 4]),
np.array([4, 4, 2, 1]), gpis=np.array([1, 2, 3, 4]))
mds1 = GriddedTsBase("", grid, MaskingTestDataset)
mds2 = GriddedTsBase("", grid, MaskingTestDataset)
mds = {
'masking1': {
'class': mds1,
'columns': ['x'],
'args': [],
'kwargs': {'limit': 500},
'use_lut': False,
'grids_compatible': True},
'masking2': {
'class': mds2,
'columns': ['x'],
'args': [],
'kwargs': {'limit': 750},
'use_lut': False,
'grids_compatible': True}
}
process = Validation(
datasets, 'DS1',
temporal_matcher=temporal_matchers.BasicTemporalMatching(
window=1 / 24.0).combinatory_matcher,
scaling='lin_cdf_match',
metrics_calculators={
(3, 2): metrics_calculators.BasicMetrics(other_name='k1').calc_metrics},
masking_datasets=mds)
gpi_info = (1, 1, 1)
ref_df = datasets['DS1']['class'].read_ts(1)
new_ref_df = process.mask_dataset(ref_df, gpi_info)
assert len(new_ref_df) == 250
nptest.assert_allclose(new_ref_df.x.values, np.arange(750, 1000))
jobs = process.get_processing_jobs()
for job in jobs:
results = process.calc(*job)
tst = tst_results[len(job[0])]
assert sorted(list(results)) == sorted(list(tst))
for key, tst_key in zip(sorted(results),
sorted(tst)):
nptest.assert_almost_equal(results[key]['n_obs'],
tst[tst_key]['n_obs'])
def test_validation_n3_k2_masking():
# test result for one gpi in a cell
tst_results_one = {
(("DS1", "x"), ("DS3", "y")): {
"n_obs": np.array([250], dtype=np.int32)
},
(("DS1", "x"), ("DS2", "y")): {
"n_obs": np.array([250], dtype=np.int32)
},
(("DS1", "x"), ("DS3", "x")): {
"n_obs": np.array([250], dtype=np.int32)
},
(("DS2", "y"), ("DS3", "x")): {
"n_obs": np.array([250], dtype=np.int32)
},
(("DS2", "y"), ("DS3", "y")): {
"n_obs": np.array([250], dtype=np.int32)
},
}
# test result for two gpis in a cell
tst_results_two = {
(("DS1", "x"), ("DS3", "y")): {
"n_obs": np.array([250, 250], dtype=np.int32)
},
(("DS1", "x"), ("DS2", "y")): {
"n_obs": np.array([250, 250], dtype=np.int32)
},
(("DS1", "x"), ("DS3", "x")): {
"n_obs": np.array([250, 250], dtype=np.int32)
},
(("DS2", "y"), ("DS3", "x")): {
"n_obs": np.array([250, 250], dtype=np.int32)
},
(("DS2", "y"), ("DS3", "y")): {
"n_obs": np.array([250, 250], dtype=np.int32)
},
}
# cell 4 in this example has two gpis so it returns different results.
tst_results = {1: tst_results_one, 1: tst_results_one, 2: tst_results_two}
datasets = setup_TestDatasets()
# setup masking datasets
grid = grids.CellGrid(
np.array([1, 2, 3, 4]),
np.array([1, 2, 3, 4]),
np.array([4, 4, 2, 1]),
gpis=np.array([1, 2, 3, 4]),
)
mds1 = GriddedTsBase("", grid, MaskingTestDataset)
mds2 = GriddedTsBase("", grid, MaskingTestDataset)
mds = {
"masking1": {
"class": mds1,
"columns": ["x"],
"args": [],
"kwargs": {"limit": 500},
"use_lut": False,
"grids_compatible": True,
},
"masking2": {
"class": mds2,
"columns": ["x"],
"args": [],
"kwargs": {"limit": 750},
"use_lut": False,
"grids_compatible": True,
},
}
process = Validation(
datasets,
"DS1",
temporal_matcher=temporal_matchers.BasicTemporalMatching(
window=1 / 24.0
).combinatory_matcher,
scaling="lin_cdf_match",
metrics_calculators={
(3, 2): metrics_calculators.BasicMetrics(
other_name="k1"
).calc_metrics
},
masking_datasets=mds,
)
gpi_info = (1, 1, 1)
ref_df = datasets["DS1"]["class"].read(1)
with warnings.catch_warnings():
warnings.simplefilter(
"ignore", category=DeprecationWarning
) # read_ts is hard coded when using mask_data
new_ref_df = process.mask_dataset(ref_df, gpi_info)
assert len(new_ref_df) == 250
nptest.assert_allclose(new_ref_df.x.values, np.arange(750, 1000))
jobs = process.get_processing_jobs()
for job in jobs:
with warnings.catch_warnings():
# most warnings here are caused by the read_ts function that cannot
# be changed when using a masking data set
warnings.simplefilter("ignore", category=DeprecationWarning)
results = process.calc(*job)
tst = tst_results[len(job[0])]
assert sorted(list(results)) == sorted(list(tst))
for key, tst_key in zip(sorted(results), sorted(tst)):
nptest.assert_almost_equal(
results[key]["n_obs"], tst[tst_key]["n_obs"]
)
def test_validation_n3_k2_masking_no_data_remains():
datasets = setup_TestDatasets()
# setup masking datasets
grid = grids.CellGrid(
np.array([1, 2, 3, 4]),
np.array([1, 2, 3, 4]),
np.array([4, 4, 2, 1]),
gpis=np.array([1, 2, 3, 4]),
)
mds1 = GriddedTsBase("", grid, MaskingTestDataset)
mds2 = GriddedTsBase("", grid, MaskingTestDataset)
mds = {
"masking1": {
"class": mds1,
"columns": ["x"],
"args": [],
"kwargs": {"limit": 500},
"use_lut": False,
"grids_compatible": True,
},
"masking2": {
"class": mds2,
"columns": ["x"],
"args": [],
"kwargs": {"limit": 1000},
"use_lut": False,
"grids_compatible": True,
},
}
process = Validation(
datasets,
"DS1",
temporal_matcher=temporal_matchers.BasicTemporalMatching(
window=1 / 24.0
).combinatory_matcher,
scaling="lin_cdf_match",
metrics_calculators={
(3, 2): metrics_calculators.BasicMetrics(
other_name="k1"
).calc_metrics
},
masking_datasets=mds,
)
gpi_info = (1, 1, 1)
ref_df = datasets["DS1"]["class"].read(1)
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=DeprecationWarning)
new_ref_df = process.mask_dataset(ref_df, gpi_info)
assert len(new_ref_df) == 0
nptest.assert_allclose(new_ref_df.x.values, np.arange(1000, 1000))
jobs = process.get_processing_jobs()
for job in jobs:
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=DeprecationWarning)
results = process.calc(*job)
tst = []
assert sorted(list(results)) == sorted(list(tst))
for key, tst_key in zip(sorted(results), sorted(tst)):
nptest.assert_almost_equal(
results[key]["n_obs"], tst[tst_key]["n_obs"]
)
def test_dfdict_combined_temporal_collocation():
ref_dr = pd.date_range("2000", "2020", freq="YS")
dr1 = pd.date_range("2000", "2015", freq="YS")
dr2 = pd.date_range("2005", "2020", freq="YS")
ref_df = pd.DataFrame({"ref": np.arange(len(ref_dr))}, index=ref_dr)
df1 = pd.DataFrame({
"k1": np.arange(len(dr1)),
"k2": np.arange(len(dr1))
},
index=dr1)
df2 = pd.DataFrame({
"k1": np.arange(len(dr2)),
"k2": np.arange(len(dr2))
},
index=dr2)
dfs = {"refkey": ref_df, "df1key": df1, "df2key": df2}
window = pd.Timedelta(days=300)
matched = temporal_matchers.dfdict_combined_temporal_collocation(
dfs, "refkey", 2, window=window, n=3, combined_dropna=True)
# keys are the same, only refkey is missing
key = ("refkey", "df1key", "df2key")
assert list(matched.keys()) == [key]
# overlap is only 11 timestamps
assert matched[key].shape == (11, 5)
overlap_dr = pd.date_range("2005", "2015", freq="YS")
assert np.all(matched[key].index == overlap_dr)
# test with ASCAT and ISMN data
here = Path(__file__).resolve().parent
ascat = pd.read_csv(here / "ASCAT.csv", index_col=0, parse_dates=True)
ismn = pd.read_csv(here / "ISMN.csv", index_col=0, parse_dates=True)
dfs = {"ASCAT": ascat[["sm"]], "ISMN": ismn[["soil_moisture"]]}
refname = "ISMN"
window = pd.Timedelta(12, "H")
old_matcher = temporal_matchers.BasicTemporalMatching().combinatory_matcher
new_matcher = temporal_matchers.make_combined_temporal_matcher(window)
expected = old_matcher(dfs, refname, k=2, n=2)
new = new_matcher(dfs, refname, k=2, n=2)
key = ("ISMN", "ASCAT")
assert list(expected.keys()) == [key]
assert list(new.keys()) == [key]
# We have to do an extra dropna for the old matcher, because the old
# matcher doesn't do this by itself.
# This is normally done within validation.py, `get_data_for_result_tuple`,
# but since the combined matcher should exclude all data where even a
# single entry misses (so that all only have common data) this is done
# before in the new matcher (the combined matcher, whereas the old one is
# the combinatory matcher)
exp = expected[key].dropna()
assert exp.shape == new[key].shape
for col in new[key]:
np.testing.assert_equal(exp[col].values, new[key][col].values)