cleanup

tests/benchmark-models/test_petab_benchmark.py

@@ -72,6 +72,7 @@ class GradientCheckSettings:
     ss_sensitivity_mode: amici.SteadyStateSensitivityMode = (
         amici.SteadyStateSensitivityMode.integrateIfNewtonFails
     )
+    noise_level: float = 0.05
 
 
 settings = defaultdict(GradientCheckSettings)
@@ -105,7 +106,7 @@ class GradientCheckSettings:
     atol_check=1e-5,
     rtol_check=1e-4,
 )
-# NOTE: Newton method fails badly for this model
+# NOTE: Newton method fails badly with ASA for this model
 settings["Blasi_CellSystems2016"] = GradientCheckSettings(
     atol_check=1e-12,
     rtol_check=1e-4,
@@ -193,10 +194,7 @@ def test_benchmark_gradient(model, scale, sensitivity_method, request):
     petab.flatten_timepoint_specific_output_overrides(petab_problem)
 
     # Only compute gradient for estimated parameters.
-    parameter_df_free = petab_problem.parameter_df.loc[
-        petab_problem.x_free_ids
-    ]
-    parameter_ids = list(parameter_df_free.index)
+    parameter_ids = petab_problem.x_free_ids
     cur_settings = settings[model]
 
     # Setup AMICI objects.
@@ -209,6 +207,7 @@ def test_benchmark_gradient(model, scale, sensitivity_method, request):
     amici_solver.setRelativeTolerance(cur_settings.rtol_sim)
     amici_solver.setMaxSteps(int(1e5))
     amici_solver.setSensitivityMethod(sensitivity_method)
+    # TODO: we should probably test all sensitivity modes
     amici_model.setSteadyStateSensitivityMode(cur_settings.ss_sensitivity_mode)
 
     amici_function, amici_derivative = simulate_petab_to_cached_functions(
@@ -224,23 +223,20 @@ def test_benchmark_gradient(model, scale, sensitivity_method, request):
         # cache=not debug,
         cache=False,
     )
-    noise_level = 0.05
     np.random.seed(cur_settings.rng_seed)
 
     # find a point where the derivative can be computed
     for _ in range(5):
         if scale:
-            point = np.asarray(
-                list(
-                    petab_problem.scale_parameters(
-                        dict(parameter_df_free.nominalValue)
-                    ).values()
-                )
+            point = petab_problem.x_nominal_free_scaled
+            point_noise = (
+                np.random.randn(len(point)) * cur_settings.noise_level
             )
-            point_noise = np.random.randn(len(point)) * noise_level
         else:
-            point = parameter_df_free.nominalValue.values
-            point_noise = np.random.randn(len(point)) * point * noise_level
+            point = petab_problem.x_nominal_free_unscaled
+            point_noise = (
+                np.random.randn(len(point)) * point * cur_settings.noise_level
+            )
         point += point_noise  # avoid small gradients at nominal value
 
         try: