def test_custom_metric_passed_to_bootstrap_compute(PM_da_initialized_1d,
PM_da_control_1d):
"""Test custom metric in bootstrap_perfect_model."""
comparison = 'e2c'
dim = 'init'
np.random.seed(42)
actual = bootstrap_perfect_model(
PM_da_initialized_1d,
PM_da_control_1d,
comparison=comparison,
metric=my_mse,
iterations=ITERATIONS,
dim=dim,
)
expected = bootstrap_perfect_model(
PM_da_initialized_1d,
PM_da_control_1d,
comparison=comparison,
metric='mse',
iterations=ITERATIONS,
dim=dim,
)
assert_allclose(actual, expected, rtol=0.1, atol=1)
def test_bootstrap_pm_assign_attrs():
"""Test assigning attrs for bootstrap_perfect_model."""
v = 'tos'
metric = 'pearson_r'
comparison = 'm2e'
ITERATIONS = 3
sig = 95
da = load_dataset('MPI-PM-DP-1D')[v].isel(area=1, period=-1)
control = load_dataset('MPI-control-1D')[v].isel(area=1, period=-1)
actual = bootstrap_perfect_model(
da,
control,
metric=metric,
comparison=comparison,
iterations=ITERATIONS,
sig=sig,
).attrs
assert actual['metric'] == metric
assert actual['comparison'] == comparison
assert actual['bootstrap_iterations'] == ITERATIONS
assert str(round((1 - sig / 100) / 2,
3)) in actual['confidence_interval_levels']
if metric == 'pearson_r':
assert actual['units'] == 'None'
assert 'bootstrap' in actual['skill_calculated_by_function']
def test_bootstrap_pm_assign_attrs():
"""Test assigning attrs for bootstrap_perfect_model."""
v = "tos"
metric = "pearson_r"
comparison = "m2e"
ITERATIONS = 3
sig = 95
da = load_dataset("MPI-PM-DP-1D")[v].isel(area=1, period=-1)
control = load_dataset("MPI-control-1D")[v].isel(area=1, period=-1)
actual = bootstrap_perfect_model(
da,
control,
metric=metric,
comparison=comparison,
iterations=ITERATIONS,
sig=sig,
).attrs
assert actual["metric"] == metric
assert actual["comparison"] == comparison
assert actual["bootstrap_iterations"] == ITERATIONS
assert str(round((1 - sig / 100) / 2,
3)) in actual["confidence_interval_levels"]
if metric == "pearson_r":
assert actual["units"] == "None"
assert "bootstrap" in actual["skill_calculated_by_function"]
def test_bootstrap_perfect_model_da_not_nan(PM_da_ds, PM_da_control, metric,
comparison):
actual = bootstrap_perfect_model(PM_da_ds,
PM_da_control,
metric=metric,
comparison=comparison,
sig=50,
bootstrap=2).isnull().any()
assert actual == False
def test_pvalue_from_bootstrapping(pm_da_ds1d, pm_da_control1d, metric):
"""Test that pvalue of initialized ensemble first lead is close to 0."""
sig = 95
actual = (bootstrap_perfect_model(
pm_da_ds1d,
pm_da_control1d,
metric=metric,
bootstrap=20,
comparison='e2c',
sig=sig,
).sel(kind='uninit', results='p').isel(lead=0))
assert actual < 2 * (1 - sig / 100)
def peakmem_bootstrap_perfect_model(self, metric, comparison):
"""Take memory peak for `bootstrap_perfect_model`."""
dim = "member" if metric in PROBABILISTIC_METRICS else None
ensure_loaded(
bootstrap_perfect_model(
self.ds,
self.control,
metric=metric,
comparison=comparison,
iterations=self.iterations,
dim=dim,
))
def Sef2018_Fig1_Different_PH_Definitions(ds, control, unit='PgC/yr', sig=95, bootstrap=1000):
# from esmtools.prediction import predictability_horizon
from PMMPIESM.plot import _set_integer_xaxis
rsig = (100 - sig)/100
_control = control
_ds = ds
ss = compute_perfect_model(
_ds, _control, metric='rmse', comparison='m2e')
ss['lead'] = np.arange(1, ss.lead.size + 1)
# ss.name = 'every'
ss_boot = bootstrap_perfect_model(_ds, _control, metric='rmse',
comparison='m2e', sig=sig, bootstrap=bootstrap)
ss_p = ss_boot.sel(kind='uninit', results='p')
ss_ci_high = ss_boot.sel(kind='uninit', results='low_ci')
ph_Spring_2019 = predictability_horizon(
ss.where(ss_p < rsig)).values
b_m2e, ph_Sef_2018, c_m2e = fit_ph_int(ss.to_series())
print('ph_Sef_2018', ph_Sef_2018)
print('ph_Spring_2019', int(ph_Spring_2019))
fig, ax = plt.subplots(figsize=(10, 4))
std = _control.std('time').values
every_color = 'mediumorchid'
ss.name = 'skill'
ss.to_dataframe().plot(ax=ax, label='skill', color='k', marker='o')
t_fit = np.arange(0, _ds.lead.size)
ax.plot(t_fit[1:], func(t_fit, b_m2e, ph_Sef_2018, c_m2e)[1:],
linewidth=3, color=every_color, label='Sef 2018 breakpoint fit')
ax.axvline(x=ph_Sef_2018, linestyle='-.',
color=every_color, label='PH Sef 2018')
ax.axhline(y=std, ls='--', c='k', alpha=.3, label='std control')
ax.axhline(y=ss_ci_high.mean('lead'), ls=':',
c='royalblue', label='Bootstrapped high CI')
ax.axvline(x=ph_Spring_2019, ls='-.', c='royalblue',
label='PH Spring 2019')
ax.set_xlabel('Lead Time [time]')
ax.set_ylabel('RMSE [' + unit + ']')
ax.set_ylim([0, ss.max() * 1.1])
ax.set_xlim([0, 10])
_set_integer_xaxis(ax)
ax.legend(frameon=False, ncol=2)
ax.set_xticks(range(1, 11))
ax.set_title(
' Global oceanic CO$_2$ flux: Differences in definitions of Predictability Horizon')
if savefig:
plt.tight_layout()
plt.savefig('FigureSI_Differences_PH_definition')
def test_mpi_he_plot_bootstrapped_skill_over_leadyear_ds(
PM_ds_initialized_1d, PM_ds_control_1d):
"""
Checks plots from bootstrap_perfect_model works for xr.Dataset with one variable.
"""
res = bootstrap_perfect_model(
PM_ds_initialized_1d,
PM_ds_control_1d,
metric="pearson_r",
iterations=ITERATIONS,
)
res_ax = plot_bootstrapped_skill_over_leadyear(res)
assert res_ax is not None
def test_mpi_pm_plot_bootstrapped_skill_over_leadyear_da(
PM_da_initialized_1d, PM_da_control_1d):
"""
Checks plots from bootstrap_perfect_model works for xr.DataArray.
"""
res = bootstrap_perfect_model(
PM_da_initialized_1d,
PM_da_control_1d,
metric='pearson_r',
iterations=ITERATIONS,
)
res_ax = plot_bootstrapped_skill_over_leadyear(res)
assert res_ax is not None
def test_pvalue_from_bootstrapping(PM_da_initialized_1d, PM_da_control_1d,
metric):
"""Test that pvalue of initialized ensemble first lead is close to 0."""
sig = 95
actual = (bootstrap_perfect_model(
PM_da_initialized_1d,
PM_da_control_1d,
metric=metric,
iterations=ITERATIONS,
comparison="e2c",
sig=sig,
dim="init",
).sel(skill="uninitialized", results="p").isel(lead=0))
assert actual.values < 2 * (1 - sig / 100)
def test_bootstrap_PM_keep_lead_attrs(
PM_da_initialized_3d, PM_da_control_3d,
):
"""Test bootstrap_perfect_model works lazily."""
PM_da_initialized_3d.lead.attrs['units'] = 'years'
s = bootstrap_perfect_model(
PM_da_initialized_3d,
PM_da_control_3d,
iterations=ITERATIONS,
comparison='m2c',
metric='mse',
)
assert 'units' in s.lead.attrs
assert s.lead.attrs['units'] == PM_da_initialized_3d.lead.attrs['units']
def test_bootstrap_perfect_model_da3d_not_nan(PM_da_ds3d, PM_da_control3d,
metric, comparison):
"""
Checks that there are no NaNs on bootstrap init skill or uninit p of 3D da.
"""
actual = bootstrap_perfect_model(PM_da_ds3d,
PM_da_control3d,
metric=metric,
comparison=comparison,
sig=50,
bootstrap=2)
actual_init_skill = actual.sel(i='init', results='skill').isnull().any()
assert not actual_init_skill
actual_uninit_p = actual.sel(i='uninit', results='p').isnull().any()
assert not actual_uninit_p
def test_mpi_he_plot_bootstrapped_skill_over_leadyear_single_uninit_lead(
PM_da_initialized_1d, PM_da_control_1d):
"""
Checks plots from bootstrap_perfect_model works for xr.DataArray.
"""
res = bootstrap_perfect_model(
PM_da_initialized_1d,
PM_da_control_1d,
metric="pearson_r",
iterations=ITERATIONS,
)
# set all but first uninit lead to nan
res[:, 2, 1:] = [np.nan] * (res.lead.size - 1)
res_ax = plot_bootstrapped_skill_over_leadyear(res)
assert res_ax is not None
def test_bootstrap_perfect_model_da1d_not_nan(PM_da_initialized_1d, PM_da_control_1d):
"""
Checks that there are no NaNs on bootstrap perfect_model of 1D da.
"""
actual = bootstrap_perfect_model(
PM_da_initialized_1d,
PM_da_control_1d,
metric='rmse',
comparison='e2c',
sig=50,
iterations=ITERATIONS,
)
actual_init_skill = actual.sel(kind='init', results='skill').isnull().any()
assert not actual_init_skill
actual_uninit_p = actual.sel(kind='uninit', results='p').isnull().any()
assert not actual_uninit_p
def test_bootstrap_perfect_model_keeps_lead_units(PM_da_initialized_1d,
PM_da_control_1d):
"""Test that lead units is kept in compute."""
sig = 95
units = "years"
PM_da_initialized_1d.lead.attrs["units"] = "years"
actual = bootstrap_perfect_model(
PM_da_initialized_1d,
PM_da_control_1d,
metric="mse",
iterations=ITERATIONS,
comparison="e2c",
sig=sig,
dim="init",
)
assert actual.lead.attrs["units"] == units
def test_bootstrap_perfect_model_da1d_not_nan(pm_da_ds1d, pm_da_control1d):
"""
Checks that there are no NaNs on bootstrap perfect_model of 1D da.
"""
actual = bootstrap_perfect_model(
pm_da_ds1d,
pm_da_control1d,
metric='rmse',
comparison='e2c',
sig=50,
bootstrap=2,
)
actual_init_skill = actual.sel(kind='init', results='skill').isnull().any()
assert not actual_init_skill
actual_uninit_p = actual.sel(kind='uninit', results='p').isnull().any()
assert not actual_uninit_p
def test_bootstrap_perfect_model_keeps_lead_units(
PM_da_initialized_1d, PM_da_control_1d
):
"""Test that lead units is kept in compute."""
sig = 95
units = 'years'
PM_da_initialized_1d.lead.attrs['units'] = 'years'
actual = bootstrap_perfect_model(
PM_da_initialized_1d,
PM_da_control_1d,
metric='mse',
iterations=ITERATIONS,
comparison='e2c',
sig=sig,
dim='init',
)
assert actual.lead.attrs['units'] == units
def test_mpi_pm_plot_bootstrapped_skill_over_leadyear():
"""
Checks plots from bootstrap_perfect_model works.
"""
da = load_dataset('MPI-PM-DP-1D').isel(area=1, period=-1)
PM_da_ds1d = da['tos']
da = load_dataset('MPI-control-1D').isel(area=1, period=-1)
PM_da_control1d = da['tos']
# sig = 95
bootstrap = 5
res = bootstrap_perfect_model(
PM_da_ds1d, PM_da_control1d, metric='pearson_r', bootstrap=bootstrap
)
res_ax = plot_bootstrapped_skill_over_leadyear(res, 95)
assert res_ax is not None
def test_bootstrap_perfect_model_da1d_not_nan(PM_da_initialized_1d, PM_da_control_1d):
"""
Checks that there are no NaNs on bootstrap perfect_model of 1D da.
"""
actual = bootstrap_perfect_model(
PM_da_initialized_1d,
PM_da_control_1d,
metric="rmse",
comparison="e2c",
dim="init",
sig=50,
iterations=ITERATIONS,
)
actual_init_skill = actual.sel(kind="initialized", results="skill").isnull().any()
assert not actual_init_skill
actual_uninit_p = actual.sel(kind="uninitialized", results="p").isnull().any()
assert not actual_uninit_p
def test_bootstrap_PM_lazy_results(
PM_da_initialized_3d, PM_da_control_3d, chunk, comparison
):
"""Test bootstrap_perfect_model works lazily."""
if chunk:
PM_da_initialized_3d = PM_da_initialized_3d.chunk({'lead': 2}).persist()
PM_da_control_3d = PM_da_control_3d.chunk({'time': -1}).persist()
else:
PM_da_initialized_3d = PM_da_initialized_3d.compute()
PM_da_control_3d = PM_da_control_3d.compute()
s = bootstrap_perfect_model(
PM_da_initialized_3d,
PM_da_control_3d,
iterations=ITERATIONS,
comparison=comparison,
metric='mse',
)
assert dask.is_dask_collection(s) == chunk
def test_bootstrap_pm_dim(pm_da_ds1d, pm_da_control1d):
"""Test whether bootstrap_hindcast calcs skill over member dim and
returns init dim."""
actual = bootstrap_perfect_model(
pm_da_ds1d,
pm_da_control1d,
metric='rmse',
dim='member',
comparison='m2c',
bootstrap=3,
)
assert 'init' in actual.dims
for kind in ['init', 'uninit']:
actualk = actual.sel(kind=kind, results='skill')
if 'init' in actualk.coords:
actualk = actualk.mean('init')
actualk = actualk.isnull().any()
assert not actualk
def test_bootstrap_pm_dim(PM_da_initialized_1d, PM_da_control_1d):
"""Test whether bootstrap_hindcast calcs skill over member dim and
returns init dim."""
actual = bootstrap_perfect_model(
PM_da_initialized_1d,
PM_da_control_1d,
metric='rmse',
dim='member',
comparison='m2c',
iterations=ITERATIONS,
resample_dim='member',
)
assert 'init' in actual.dims
for kind in ['init', 'uninit']:
actualk = actual.sel(kind=kind, results='skill')
if 'init' in actualk.coords:
actualk = actualk.mean('init')
actualk = actualk.isnull().any()
assert not actualk
def test_bootstrap_perfect_model_ds1d_not_nan(PM_ds_ds1d, PM_ds_control1d,
metric, comparison):
"""
Checks that there are no NaNs on bootstrap init skill or uninit p of 1D ds time series.
"""
actual = bootstrap_perfect_model(PM_ds_ds1d,
PM_ds_control1d,
metric=metric,
comparison=comparison,
sig=50,
bootstrap=2)
for var in actual.data_vars:
actual_init_skill = actual[var].sel(i='init',
results='skill').isnull().any()
assert not actual_init_skill
for var in actual.data_vars:
actual_uninit_p = actual[var].sel(i='uninit',
results='p').isnull().any()
assert not actual_uninit_p
def test_bootstrap_perfect_model_da1d_not_nan_probabilistic(
PM_da_initialized_1d, PM_da_control_1d, metric, comparison):
"""
Checks that there are no NaNs on perfect model probabilistic metrics of 1D
time series.
"""
kwargs = {
"comparison": comparison,
"metric": metric,
}
category_edges = np.array([0, 0.5, 1])
if metric in probabilistic_metrics_requiring_logical:
def f(x):
return x > 0.5
kwargs["logical"] = f
elif metric == "threshold_brier_score":
kwargs["threshold"] = 0.5
elif metric == "contingency":
kwargs["forecast_category_edges"] = category_edges
kwargs["observation_category_edges"] = category_edges
kwargs["score"] = "accuracy"
elif metric == "rps":
kwargs["category_edges"] = category_edges
dim = (["member", "init"]
if metric in probabilistic_metrics_requiring_more_than_member_dim
else "member")
kwargs["dim"] = dim
assert (not compute_perfect_model(PM_da_initialized_1d, PM_da_control_1d,
**kwargs).isnull().all())
kwargs["iterations"] = ITERATIONS
kwargs["resample_dim"] = "member"
actual = bootstrap_perfect_model(PM_da_initialized_1d, PM_da_control_1d,
**kwargs)
for kind in ["initialized", "uninitialized"]:
actualk = actual.sel(kind=kind, results="skill")
actualk = actualk.isnull().all()
assert not actualk
def test_bootstrap_perfect_model_ds1d_not_nan(PM_ds_initialized_1d,
PM_ds_control_1d):
"""
Checks that there are no NaNs on bootstrap perfect_model of 1D ds.
"""
actual = bootstrap_perfect_model(
PM_ds_initialized_1d,
PM_ds_control_1d,
metric="rmse",
comparison="e2c",
dim="init",
sig=50,
iterations=ITERATIONS,
)
for var in actual.data_vars:
actual_init_skill = (actual[var].sel(
skill="initialized", results="verify skill").isnull().any())
assert not actual_init_skill
for var in actual.data_vars:
actual_uninit_p = (actual[var].sel(skill="uninitialized",
results="p").isnull().any())
assert not actual_uninit_p
def test_bootstrap_perfect_model_da1d_not_nan_probabilistic(
PM_da_initialized_1d, PM_da_control_1d, metric, comparison
):
"""
Checks that there are no NaNs on perfect model probabilistic metrics of 1D
time series.
"""
if 'threshold' in metric:
threshold = 10.5
else:
threshold = None
if metric == 'brier_score':
def func(x):
return x > 0
else:
func = None
actual = bootstrap_perfect_model(
PM_da_initialized_1d,
PM_da_control_1d,
comparison=comparison,
metric=metric,
threshold=threshold,
gaussian=True,
func=func,
iterations=ITERATIONS,
dim='member',
resample_dim='member',
)
for kind in ['init', 'uninit']:
actualk = actual.sel(kind=kind, results='skill')
if 'init' in actualk.coords:
actualk = actualk.mean('init')
actualk = actualk.isnull().any()
assert not actualk
s3d = compute_perfect_model(ds3d, control3d, metric=metric)
# s3d.plot(col='lead')
s3dapn = compute_perfect_model(
ds3d.sel(init=anom_pos_nino34_init.values), control3d, metric=metric)
s3daneun = compute_perfect_model(
ds3d.sel(init=anom_neu_nino34_init), control3d, metric=metric)
s3dann = compute_perfect_model(
ds3d.sel(init=anom_neg_nino34_init.values), control3d, metric=metric)
sa3d = xr.concat([s3d, s3dapn, s3daneun, s3dann], 'IC')
sa3d['IC'] = ['all', 'pos', 'neutral', 'neg']
warnings.simplefilter("ignore")
b = 100
s3d = bootstrap_perfect_model(ds3d, control3d, metric=metric, bootstrap=b)
s3dapn = bootstrap_perfect_model(
ds3d.sel(init=anom_pos_nino34_init.values), control3d, metric=metric, bootstrap=b)
s3daneun = bootstrap_perfect_model(
ds3d.sel(init=anom_neu_nino34_init), control3d, metric=metric, bootstrap=b)
s3dann = bootstrap_perfect_model(
ds3d.sel(init=anom_neg_nino34_init.values), control3d, metric=metric, bootstrap=b)
sa3d.sel(kind='uninit', results='p').plot(col='lead', row='IC', vmax=.2)
sa3d = xr.concat([s3d, s3dapn, s3daneun, s3dann], 'IC')
sa3d['IC'] = ['all', 'pos', 'neutral', 'neg']
skill = sa3d.sel(kind='init', results='skill').where(
sa3d.sel(results='p', kind='uninit') <= .05)
#sa3d.plot(col='lead', row='IC', robust=True)
#sa3d.name = 'CO$_2$ ACC'
metric = 'pearson_r'
v = 'co2_flux'
v='CO2'
metric = 'rmse'
for metric in ['pearson_r', 'rmse']:
for v in ['co2_flux', 'CO2']:
varstring = copy(v)
print(v, varstring)
ds = xr.open_dataset(data_path + 'ds_' + v + '_ym.nc').load()
control = xr.open_dataset(data_path + 'control_' + v + '_ym.nc').load()
ds, control = comply_climpred(ds, control)
# faster
ds = ds.isel(lead=slice(None, 6)).isel(member=slice(1,None))
bs = bootstrap_perfect_model(
ds, control, metric=metric, comparison=comparison, bootstrap=bootstrap, sig=sig)
if save_nc:
bs.to_netcdf('_'.join(['results', varstring, 'ym', 'metric', metric, 'comparison',
comparison, 'sig', str(sig), 'bootstrap', str(bootstrap)]) + '.nc')
skill = bs.sel(results='skill', kind='init').where(
bs.sel(results='p', kind='uninit') <= psig)
skill[v].plot(col='lead',col_wrap=4,yincrease=False)
# plots
bootstrap = 1000
comparison = 'm2e'
psig = .05
# prototyping
metric = 'rmse'
comparison = 'm2e'
sig = 95
psig = (100 - sig) / 100
bootstrap = 5000
control = xr.open_dataset(post_global + 'control_diagnosed_co2.nc')
ds = xr.open_dataset(post_global + 'ds_diagnosed_co2.nc')
# rename dims to climpred requirements
if 'ensemble' in ds.dims:
ds = ds.rename({'ensemble': 'init', 'time': 'lead'})
p = 'data/results/'
compute = False
if compute:
bs_acc = bootstrap_perfect_model(
ds, control, metric='pearson_r', comparison=comparison, bootstrap=bootstrap, sig=sig)
if save_nc:
bs_acc.to_netcdf(p+'_'.join(['results', 'global', 'ym', 'metric', 'pearson_r',
'comparison', comparison, 'sig', str(sig), 'bootstrap', str(bootstrap)]) + '.nc')
bs_rmse = bootstrap_perfect_model(
ds, control, metric='rmse', comparison=comparison, bootstrap=bootstrap, sig=sig)
if save_nc:
bs_rmse.to_netcdf(p+'_'.join(['results', 'global', 'ym', 'metric', 'rmse',
'comparison', comparison, 'sig', str(sig), 'bootstrap', str(bootstrap)]) + '.nc')
else:
bs_acc = xr.open_dataset(p+'_'.join(['results', 'global', 'ym', 'metric', 'pearson_r',
'comparison', comparison, 'sig', str(sig), 'bootstrap', str(bootstrap)]) + '.nc')
bs_rmse = xr.open_dataset(p+'_'.join(['results', 'global', 'ym', 'metric', 'rmse',
'comparison', comparison, 'sig', str(sig), 'bootstrap', str(bootstrap)]) + '.nc')
