feat: restore optimizer settings after fits

src/cabinetry/fit/__init__.py

@@ -81,6 +81,7 @@ def _fit_model_pyhf(
     Returns:
         FitResults: object storing relevant fit results
     """
+    _, initial_optimizer = pyhf.get_backend()  # store initial optimizer settings
     pyhf.set_backend(pyhf.tensorlib, pyhf.optimize.minuit_optimizer(verbose=1))
 
     # strategy=None is currently not supported in pyhf
@@ -124,7 +125,7 @@ def _fit_model_pyhf(
         best_twice_nll,
         minos_uncertainty=minos_results,
     )
-
+    pyhf.set_backend(pyhf.tensorlib, initial_optimizer)  # restore optimizer settings
     return fit_results
 
 
@@ -174,6 +175,7 @@ def _fit_model_custom(
     Returns:
         FitResults: object storing relevant fit results
     """
+    _, initial_optimizer = pyhf.get_backend()  # store initial optimizer settings
     pyhf.set_backend(pyhf.tensorlib, pyhf.optimize.minuit_optimizer(verbose=1))
 
     # use parameter settings provided in function arguments if they exist, else defaults
@@ -236,7 +238,7 @@ def twice_nll_func(pars: np.ndarray) -> Any:
         best_twice_nll,
         minos_uncertainty=minos_results,
     )
-
+    pyhf.set_backend(pyhf.tensorlib, initial_optimizer)  # restore optimizer settings
     return fit_results
 
 
@@ -774,6 +776,7 @@ def limit(
     Returns:
         LimitResults: observed and expected limits, CLs values, and scanned points
     """
+    _, initial_optimizer = pyhf.get_backend()  # store initial optimizer settings
     pyhf.set_backend(pyhf.tensorlib, pyhf.optimize.minuit_optimizer(verbose=1))
 
     # use POI given by kwarg, fall back to POI specified in model
@@ -984,6 +987,7 @@ def _cls_minus_threshold(
         poi_arr,
         confidence_level,
     )
+    pyhf.set_backend(pyhf.tensorlib, initial_optimizer)  # restore optimizer settings
     return limit_results
 
 
@@ -1012,6 +1016,7 @@ def significance(
     Returns:
         SignificanceResults: observed and expected p-values and significances
     """
+    _, initial_optimizer = pyhf.get_backend()  # store initial optimizer settings
     pyhf.set_backend(pyhf.tensorlib, pyhf.optimize.minuit_optimizer(verbose=1))
 
     log.info("calculating discovery significance")
@@ -1044,4 +1049,5 @@ def significance(
     significance_results = SignificanceResults(
         obs_p_val, obs_significance, exp_p_val, exp_significance
     )
+    pyhf.set_backend(pyhf.tensorlib, initial_optimizer)  # restore optimizer settings
     return significance_results

tests/fit/test_fit.py

@@ -4,6 +4,7 @@
 
 import iminuit
 import numpy as np
+import pyhf
 import pytest
 
 from cabinetry import fit
@@ -29,13 +30,15 @@ def test_print_results(caplog):
 @pytest.mark.filterwarnings("ignore::RuntimeWarning")
 @mock.patch("cabinetry.fit._run_minos", return_value={"Signal strength": (-0.1, 0.2)})
 def test__fit_model_pyhf(mock_minos, example_spec, example_spec_multibin):
+    pyhf.set_backend("numpy", "scipy")
     model, data = model_utils.model_and_data(example_spec)
     fit_results = fit._fit_model_pyhf(model, data)
     assert np.allclose(fit_results.bestfit, [8.33624084, 1.1])
     assert np.allclose(fit_results.uncertainty, [0.38182003, 0.0])
     assert fit_results.labels == ["Signal strength", "staterror_Signal-Region[0]"]
     assert np.allclose(fit_results.best_twice_nll, 7.82495235)
     assert np.allclose(fit_results.corr_mat, [[1.0, 0.0], [0.0, 0.0]])
+    assert pyhf.get_backend()[1].name == "scipy"  # optimizer was reset
 
     # Asimov fit, with fixed gamma (fixed not to Asimov MLE)
     model, data = model_utils.model_and_data(example_spec, asimov=True)
@@ -104,13 +107,15 @@ def test__fit_model_pyhf(mock_minos, example_spec, example_spec_multibin):
 @pytest.mark.filterwarnings("ignore::RuntimeWarning")
 @mock.patch("cabinetry.fit._run_minos", return_value={"Signal strength": (-0.1, 0.2)})
 def test__fit_model_custom(mock_minos, example_spec, example_spec_multibin):
+    pyhf.set_backend("numpy", "scipy")
     model, data = model_utils.model_and_data(example_spec)
     fit_results = fit._fit_model_custom(model, data)
     assert np.allclose(fit_results.bestfit, [8.33625071, 1.1])
     assert np.allclose(fit_results.uncertainty, [0.38182151, 0.0])
     assert fit_results.labels == ["Signal strength", "staterror_Signal-Region[0]"]
     assert np.allclose(fit_results.best_twice_nll, 7.82495235)
     assert np.allclose(fit_results.corr_mat, [[1.0, 0.0], [0.0, 0.0]])
+    assert pyhf.get_backend()[1].name == "scipy"  # optimizer was reset
 
     # Asimov fit, with fixed gamma (fixed not to Asimov MLE)
     model, data = model_utils.model_and_data(example_spec, asimov=True)
@@ -659,6 +664,7 @@ def test_scan(mock_fit, example_spec):
 
 def test_limit(example_spec_with_background, caplog):
     caplog.set_level(logging.DEBUG)
+    pyhf.set_backend("numpy", "scipy")
 
     # expected values for results
     observed_limit = 0.749
@@ -686,6 +692,7 @@ def test_limit(example_spec_with_background, caplog):
     assert limit_results.confidence_level == 0.95
     # verify that POI values are sorted
     assert np.allclose(limit_results.poi_values, sorted(limit_results.poi_values))
+    assert pyhf.get_backend()[1].name == "scipy"  # optimizer was reset
     caplog.clear()
 
     # access negative POI values with lower bracket below zero
@@ -780,6 +787,7 @@ def test_limit(example_spec_with_background, caplog):
 
 
 def test_significance(example_spec_with_background):
+    pyhf.set_backend("numpy", "scipy")
     # increase observed data for smaller observed p-value
     example_spec_with_background["observations"][0]["data"] = [196]
 
@@ -789,6 +797,7 @@ def test_significance(example_spec_with_background):
     assert np.allclose(significance_results.observed_significance, 3.15402672)
     assert np.allclose(significance_results.expected_p_value, 0.00033333)
     assert np.allclose(significance_results.expected_significance, 3.40293444)
+    assert pyhf.get_backend()[1].name == "scipy"  # optimizer was reset
 
     # reduce signal for larger expected p-value
     example_spec_with_background["channels"][0]["samples"][0]["data"] = [30]

tests/test_integration.py

@@ -193,7 +193,7 @@ def test_integration(tmp_path, ntuple_creator, caplog):
     assert np.allclose(
         prediction_postfit.total_stdev_model_channels,
         [[41.043814, 45.814417, 20.439575]],
-        atol=2e-3,
+        atol=5e-3,
     )
     _ = cabinetry.visualize.data_mc(prediction_postfit, data, close_figure=True)