Refactor, reduce tolerance

ThirdParty/SuiteSparse/SuiteSparse_config/SuiteSparse_config.h

@@ -374,7 +374,7 @@ int SuiteSparse_divcomplex
     // be done via the SUITESPARSE_TIME macro, defined below:
     #define SUITESPARSE_TIMER_ENABLED
     #define SUITESPARSE_HAVE_CLOCK_GETTIME
-    #define SUITESPARSE_CONFIG_TIMER clock_gettime
+    #define SUITESPARSE_CONFIG_TIMER omp_get_wtime
     #if defined ( SUITESPARSE_TIMER_ENABLED )
         #if defined ( _OPENMP )
             // Avoid indirection through the library if the compilation unit

tests/benchmark-models/test_petab_benchmark.py

@@ -9,11 +9,8 @@
 import pytest
 from amici.petab.petab_import import import_petab_problem
 import benchmark_models_petab
-
-
-# Absolute and relative tolerances for finite difference gradient checks.
-ATOL: float = 1e-3
-RTOL: float = 1e-2
+from collections import defaultdict
+from dataclasses import dataclass
 
 repo_root = Path(__file__).parent.parent.parent
 
@@ -42,9 +39,45 @@
     debug_path.mkdir(exist_ok=True, parents=True)
 
 
+@dataclass
+class GradientCheckSettings:
+    # Absolute and relative tolerances for simulation
+    atol_sim: float = 1e-12
+    rtol_sim: float = 1e-12
+    # Absolute and relative tolerances for finite difference gradient checks.
+    atol_check: float = 1e-4
+    rtol_check: float = 1e-2
+    # Step sizes for finite difference gradient checks.
+    step_sizes = [
+        1e-1,
+        5e-2,
+        1e-2,
+        1e-3,
+        1e-4,
+        1e-5,
+    ]
+
+
+settings = defaultdict(GradientCheckSettings)
+settings["Smith_BMCSystBiol2013"] = GradientCheckSettings(
+    atol_sim=1e-10,
+    rtol_sim=1e-10,
+)
+settings["Oliveira_NatCommun2021"] = GradientCheckSettings(
+    atol_sim=1e-10,
+    rtol_sim=1e-10,
+)
+settings["Okuonghae_ChaosSolitonsFractals2020"] = GradientCheckSettings(
+    atol_sim=1e-14,
+    rtol_sim=1e-14,
+)
+settings["Okuonghae_ChaosSolitonsFractals2020"].step_sizes.insert(0, 0.2)
+
+
 # until fiddy is updated
 @pytest.mark.filterwarnings(
-    "ignore:Importing amici.petab_objective is deprecated.:DeprecationWarning"
+    "ignore:Importing amici.petab_objective is deprecated.:DeprecationWarning",
+    r"ignore:.*petab\.v1.*instead.*:DeprecationWarning",
 )
 @pytest.mark.filterwarnings("ignore:divide by zero encountered in log10")
 @pytest.mark.parametrize("scale", (True, False))
@@ -78,24 +111,16 @@ def test_benchmark_gradient(model, scale):
         petab_problem.x_free_ids
     ]
     parameter_ids = list(parameter_df_free.index)
+    cur_settings = settings[model]
 
     # Setup AMICI objects.
     amici_model = import_petab_problem(
         petab_problem,
         model_output_dir=benchmark_outdir / model,
     )
     amici_solver = amici_model.getSolver()
-    amici_solver.setAbsoluteTolerance(1e-12)
-    amici_solver.setRelativeTolerance(1e-12)
-    if model in (
-        "Smith_BMCSystBiol2013",
-        "Oliveira_NatCommun2021",
-    ):
-        amici_solver.setAbsoluteTolerance(1e-10)
-        amici_solver.setRelativeTolerance(1e-10)
-    elif model in ("Okuonghae_ChaosSolitonsFractals2020",):
-        amici_solver.setAbsoluteTolerance(1e-14)
-        amici_solver.setRelativeTolerance(1e-14)
+    amici_solver.setAbsoluteTolerance(cur_settings.atol_sim)
+    amici_solver.setRelativeTolerance(cur_settings.rtol_sim)
     amici_solver.setMaxSteps(int(1e5))
 
     if model in ("Brannmark_JBC2010",):
@@ -132,21 +157,10 @@ def test_benchmark_gradient(model, scale):
 
     expected_derivative = amici_derivative(point)
 
-    sizes = [
-        1e-1,
-        5e-2,
-        1e-2,
-        1e-3,
-        1e-4,
-        1e-5,
-    ]
-    if model in ("Okuonghae_ChaosSolitonsFractals2020",):
-        sizes.insert(0, 0.2)
-
     derivative = get_derivative(
         function=amici_function,
         point=point,
-        sizes=sizes,
+        sizes=cur_settings.step_sizes,
         direction_ids=parameter_ids,
         method_ids=[MethodId.CENTRAL, MethodId.FORWARD, MethodId.BACKWARD],
         success_checker=Consistency(atol=1e-5, rtol=1e-1),
@@ -160,7 +174,9 @@ def test_benchmark_gradient(model, scale):
             expectation=expected_derivative,
             point=point,
         )
-        success = check(rtol=RTOL, atol=ATOL)
+        success = check(
+            rtol=cur_settings.rtol_check, atol=cur_settings.atol_check
+        )
 
     if debug:
         df = pd.DataFrame(