To increase usability of minimizer it should be possible to set the t…

…olerance and the max number of iteration (#84)

Co-authored-by: Andrew Sazonov <[email protected]>

src/easyscience/fitting/fitter.py

@@ -26,11 +26,13 @@ class Fitter:
     def __init__(self, fit_object, fit_function: Callable):
         self._fit_object = fit_object
         self._fit_function = fit_function
-        self._dependent_dims = None
+        self._dependent_dims: int = None
+        self._tolerance: float = None
+        self._max_evaluations: int = None
 
-        self._enum_current_minimizer = DEFAULT_MINIMIZER
-        self._minimizer: MinimizerBase  # _minimizer is set in the create method
-        self._update_minimizer(self._enum_current_minimizer)
+        self._minimizer: MinimizerBase = None  # set in _update_minimizer
+        self._enum_current_minimizer: AvailableMinimizers = None  # set in _update_minimizer
+        self._update_minimizer(DEFAULT_MINIMIZER)
 
     def fit_constraints(self) -> list:
         return self._minimizer.fit_constraints()
@@ -110,6 +112,42 @@ def minimizer(self) -> MinimizerBase:
         """
         return self._minimizer
 
+    @property
+    def tolerance(self) -> float:
+        """
+        Get the tolerance for the minimizer.
+
+        :return: Tolerance for the minimizer
+        """
+        return self._tolerance
+
+    @tolerance.setter
+    def tolerance(self, tolerance: float) -> None:
+        """
+        Set the tolerance for the minimizer.
+
+        :param tolerance: Tolerance for the minimizer
+        """
+        self._tolerance = tolerance
+
+    @property
+    def max_evaluations(self) -> int:
+        """
+        Get the maximal number of evaluations for the minimizer.
+
+        :return: Maximal number of steps for the minimizer
+        """
+        return self._max_evaluations
+
+    @max_evaluations.setter
+    def max_evaluations(self, max_evaluations: int) -> None:
+        """
+        Set the maximal number of evaluations for the minimizer.
+
+        :param max_evaluations: Maximal number of steps for the minimizer
+        """
+        self._max_evaluations = max_evaluations
+
     @property
     def fit_function(self) -> Callable:
         """
@@ -175,7 +213,7 @@ def fit(self) -> Callable:
         re-constitute the independent variables and once the fit is completed, reshape the inputs to those expected.
         """
 
-        @functools.wraps(self.minimizer.fit)
+        @functools.wraps(self._minimizer.fit)
         def inner_fit_callable(
             x: np.ndarray,
             y: np.ndarray,
@@ -202,7 +240,14 @@ def inner_fit_callable(
             constraints = self._minimizer.fit_constraints()
             self.fit_function = fit_fun_wrap
             self._minimizer.set_fit_constraint(constraints)
-            f_res = self.minimizer.fit(x_fit, y_new, weights=weights, **kwargs)
+            f_res = self._minimizer.fit(
+                x_fit,
+                y_new,
+                weights=weights,
+                tolerance=self._tolerance,
+                max_evaluations=self._max_evaluations,
+                **kwargs,
+            )
 
             # Postcompute
             fit_result = self._post_compute_reshaping(f_res, x, y)

src/easyscience/fitting/minimizers/minimizer_base.py

@@ -40,7 +40,7 @@ def __init__(
         self,
         obj,  #: BaseObj,
         fit_function: Callable,
-        minimizer_enum: Optional[AvailableMinimizers] = None,
+        minimizer_enum: AvailableMinimizers,
     ):  # todo after constraint changes, add type hint: obj: BaseObj  # noqa: E501
         if minimizer_enum.method not in self.supported_methods():
             raise FitError(f'Method {minimizer_enum.method} not available in {self.__class__}')
@@ -58,6 +58,10 @@ def __init__(
     def all_constraints(self) -> List[ObjConstraint]:
         return [*self._constraints, *self._object._constraints]
 
+    @property
+    def enum(self) -> AvailableMinimizers:
+        return self._minimizer_enum
+
     @property
     def name(self) -> str:
         return self._minimizer_enum.name
@@ -83,6 +87,8 @@ def fit(
         model: Optional[Callable] = None,
         parameters: Optional[Parameter] = None,
         method: Optional[str] = None,
+        tolerance: Optional[float] = None,
+        max_evaluations: Optional[int] = None,
         **kwargs,
     ) -> FitResults:
         """
@@ -129,7 +135,7 @@ def evaluate(self, x: np.ndarray, minimizer_parameters: Optional[dict[str, float
 
         return self._fit_function(x, **minimizer_parameters, **kwargs)
 
-    def _get_method_dict(self, passed_method: Optional[str] = None) -> dict[str, str]:
+    def _get_method_kwargs(self, passed_method: Optional[str] = None) -> dict[str, str]:
         if passed_method is not None:
             if passed_method not in self.supported_methods():
                 raise FitError(f'Method {passed_method} not available in {self.__class__}')

src/easyscience/fitting/minimizers/minimizer_bumps.py

@@ -74,6 +74,8 @@ def fit(
         model: Optional[Callable] = None,
         parameters: Optional[Parameter] = None,
         method: Optional[str] = None,
+        tolerance: Optional[float] = None,
+        max_evaluations: Optional[int] = None,
         minimizer_kwargs: Optional[dict] = None,
         engine_kwargs: Optional[dict] = None,
         **kwargs,
@@ -97,7 +99,7 @@ def fit(
         :return: Fit results
         :rtype: ModelResult
         """
-        method_dict = self._get_method_dict(method)
+        method_dict = self._get_method_kwargs(method)
 
         if weights is None:
             weights = np.sqrt(np.abs(y))
@@ -107,10 +109,14 @@ def fit(
 
         if minimizer_kwargs is None:
             minimizer_kwargs = {}
-        # else:
-        #     minimizer_kwargs = {"fit_kws": minimizer_kwargs}
         minimizer_kwargs.update(engine_kwargs)
 
+        if tolerance is not None:
+            minimizer_kwargs['ftol'] = tolerance  # tolerance for change in function value
+            minimizer_kwargs['xtol'] = tolerance  # tolerance for change in parameter value, could be an independent value
+        if max_evaluations is not None:
+            minimizer_kwargs['steps'] = max_evaluations
+
         if model is None:
             model_function = self._make_model(parameters=parameters)
             model = model_function(x, y, weights)

src/easyscience/fitting/minimizers/minimizer_dfo.py

@@ -53,9 +53,7 @@ def supported_methods() -> List[str]:
 
     @staticmethod
     def all_methods() -> List[str]:
-        return [
-            'leastsq',
-        ]
+        return ['leastsq']
 
     def fit(
         self,
@@ -65,8 +63,8 @@ def fit(
         model: Optional[Callable] = None,
         parameters: Optional[List[Parameter]] = None,
         method: str = None,
-        xtol: float = 1e-6,
-        ftol: float = 1e-8,
+        tolerance: Optional[float] = None,
+        max_evaluations: Optional[int] = None,
         **kwargs,
     ) -> FitResults:
         """
@@ -110,6 +108,8 @@ def fit(
         stack_status = global_object.stack.enabled
         global_object.stack.enabled = False
 
+        kwargs = self._prepare_kwargs(tolerance, max_evaluations, **kwargs)
+
         try:
             model_results = self._dfo_fit(self._cached_pars, model, **kwargs)
             self._set_parameter_fit_result(model_results, stack_status)
@@ -239,7 +239,11 @@ def _gen_fit_results(self, fit_results, weights, **kwargs) -> FitResults:
         return results
 
     @staticmethod
-    def _dfo_fit(pars: Dict[str, Parameter], model: Callable, **kwargs):
+    def _dfo_fit(
+        pars: Dict[str, Parameter],
+        model: Callable,
+        **kwargs,
+    ):
         """
         Method to convert EasyScience styling to DFO-LS styling (yes, again)
 
@@ -261,13 +265,23 @@ def _dfo_fit(pars: Dict[str, Parameter], model: Callable, **kwargs):
             np.array([par.max for par in pars.values()]),
         )
         # https://numericalalgorithmsgroup.github.io/dfols/build/html/userguide.html
-        if np.isinf(bounds).any():
-            results = dfols.solve(model, pars_values, bounds=bounds, **kwargs)
-        else:
+        if not np.isinf(bounds).any():
             # It is only possible to scale (normalize) variables if they are bound (different from inf)
-            results = dfols.solve(model, pars_values, bounds=bounds, scaling_within_bounds=True, **kwargs)
+            kwargs['scaling_within_bounds'] = True
+
+        results = dfols.solve(model, pars_values, bounds=bounds, **kwargs)
 
         if 'Success' not in results.msg:
             raise FitError(f'Fit failed with message: {results.msg}')
 
         return results
+
+    @staticmethod
+    def _prepare_kwargs(tolerance: Optional[float] = None, max_evaluations: Optional[int] = None, **kwargs) -> dict[str:str]:
+        if max_evaluations is not None:
+            kwargs['maxfun'] = max_evaluations  # max number of function evaluations
+        if tolerance is not None:
+            if 0.1 < tolerance:  # dfo module throws errer if larger value
+                raise ValueError('Tolerance must be equal or smaller than 0.1')
+            kwargs['rhoend'] = tolerance  # size of the trust region
+        return kwargs

src/easyscience/fitting/minimizers/minimizer_lmfit.py

@@ -85,6 +85,8 @@ def fit(
         model: Optional[LMModel] = None,
         parameters: Optional[LMParameters] = None,
         method: Optional[str] = None,
+        tolerance: Optional[float] = None,
+        max_evaluations: Optional[int] = None,
         minimizer_kwargs: Optional[dict] = None,
         engine_kwargs: Optional[dict] = None,
         **kwargs,
@@ -110,19 +112,14 @@ def fit(
         :return: Fit results
         :rtype: ModelResult
         """
-        method_dict = self._get_method_dict(method)
-
         if weights is None:
             weights = 1 / np.sqrt(np.abs(y))
 
         if engine_kwargs is None:
             engine_kwargs = {}
 
-        if minimizer_kwargs is None:
-            minimizer_kwargs = {}
-        else:
-            minimizer_kwargs = {'fit_kws': minimizer_kwargs}
-        minimizer_kwargs.update(engine_kwargs)
+        method_kwargs = self._get_method_kwargs(method)
+        fit_kws_dict = self._get_fit_kws(method, tolerance, minimizer_kwargs)
 
         # Why do we do this? Because a fitting template has to have global_object instantiated outside pre-runtime
         from easyscience import global_object
@@ -134,7 +131,16 @@ def fit(
             if model is None:
                 model = self._make_model()
 
-            model_results = model.fit(y, x=x, weights=weights, **method_dict, **minimizer_kwargs, **kwargs)
+            model_results = model.fit(
+                y,
+                x=x,
+                weights=weights,
+                max_nfev=max_evaluations,
+                fit_kws=fit_kws_dict,
+                **method_kwargs,
+                **engine_kwargs,
+                **kwargs,
+            )
             self._set_parameter_fit_result(model_results, stack_status)
             results = self._gen_fit_results(model_results)
         except Exception as e:
@@ -143,6 +149,16 @@ def fit(
             raise FitError(e)
         return results
 
+    def _get_fit_kws(self, method: str, tolerance: float, minimizer_kwargs: dict[str:str]) -> dict[str:str]:
+        if minimizer_kwargs is None:
+            minimizer_kwargs = {}
+        if tolerance is not None:
+            if method in [None, 'least_squares', 'leastsq']:
+                minimizer_kwargs['ftol'] = tolerance
+            if method in ['differential_evolution', 'powell', 'cobyla']:
+                minimizer_kwargs['tol'] = tolerance
+        return minimizer_kwargs
+
     def convert_to_pars_obj(self, parameters: Optional[List[Parameter]] = None) -> LMParameters:
         """
         Create an lmfit compatible container with the `Parameters` converted from the base object.

tests/integration_tests/Fitting/test_fitter.py

@@ -140,6 +140,62 @@ def test_fit_result(fit_engine):
     check_fit_results(result, sp_sin, ref_sin, x, sp_ref1=sp_ref1, sp_ref2=sp_ref2)
 
 
+@pytest.mark.parametrize("fit_engine", [None, AvailableMinimizers.LMFit, AvailableMinimizers.Bumps, AvailableMinimizers.DFO])
+def test_basic_max_evaluations(fit_engine):
+    ref_sin = AbsSin(0.2, np.pi)
+    sp_sin = AbsSin(0.354, 3.05)
+
+    x = np.linspace(0, 5, 200)
+    y = ref_sin(x)
+
+    sp_sin.offset.fixed = False
+    sp_sin.phase.fixed = False
+
+    f = Fitter(sp_sin, sp_sin)
+    if fit_engine is not None:
+        try:
+            f.switch_minimizer(fit_engine)
+        except AttributeError:
+            pytest.skip(msg=f"{fit_engine} is not installed")
+    args = [x, y]
+    kwargs = {}
+    f.max_evaluations = 3
+    try:
+        result = f.fit(*args, **kwargs)
+        # Result should not be the same as the reference
+        assert sp_sin.phase.value != pytest.approx(ref_sin.phase.value, rel=1e-3)
+        assert sp_sin.offset.value != pytest.approx(ref_sin.offset.value, rel=1e-3)
+    except FitError as e:
+        # DFO throws a different error
+        assert "Objective has been called MAXFUN times" in str(e)
+
+
+@pytest.mark.parametrize("fit_engine,tolerance", [(None, 10), (AvailableMinimizers.LMFit, 10), (AvailableMinimizers.Bumps, 10), (AvailableMinimizers.DFO, 0.1)])
+def test_basic_tolerance(fit_engine, tolerance):
+    ref_sin = AbsSin(0.2, np.pi)
+    sp_sin = AbsSin(0.354, 3.05)
+
+    x = np.linspace(0, 5, 200)
+    y = ref_sin(x)
+
+    sp_sin.offset.fixed = False
+    sp_sin.phase.fixed = False
+
+    f = Fitter(sp_sin, sp_sin)
+    if fit_engine is not None:
+        try:
+            f.switch_minimizer(fit_engine)
+        except AttributeError:
+            pytest.skip(msg=f"{fit_engine} is not installed")
+    args = [x, y]
+    kwargs = {}
+    f.tolerance = tolerance
+    result = f.fit(*args, **kwargs)
+    # Result should not be the same as the reference
+    assert sp_sin.phase.value != pytest.approx(ref_sin.phase.value, rel=1e-3)
+    assert sp_sin.offset.value != pytest.approx(ref_sin.offset.value, rel=1e-3)
+
+
 @pytest.mark.parametrize("fit_method", ["leastsq", "powell", "cobyla"])
 def test_lmfit_methods(fit_method):
     ref_sin = AbsSin(0.2, np.pi)