Add DistfitFitter distribution fitter wrapping the distfit package

.all-contributorsrc

@@ -272,6 +272,16 @@
         "bug"
       ]
     },
+    {
+      "login": "pyarchana",
+      "name": "Archana K.",
+      "avatar_url": "https://avatars.githubusercontent.com/u/93970069?v=4",
+      "profile": "https://github.com/pyarchana",
+      "contributions": [
+        "code",
+        "test"
+      ]
+    },
     {
       "login": "Kushagra651",
       "name": "Kushagra",

docs/source/api_reference/distfitter.rst

@@ -21,6 +21,7 @@ Parametric fitters
     :toctree: auto_generated/
     :template: class.rst
 
+    DistfitFitter
     MOMFitter
     NormalFitter
 

skpro/distfitter/__init__.py

@@ -1,10 +1,12 @@
 """Distribution fitter estimators."""
 # copyright: skpro developers, BSD-3-Clause License (see LICENSE file)
 
+from skpro.distfitter._distfitfitter import DistfitFitter
 from skpro.distfitter._momfitter import MOMFitter
 from skpro.distfitter._normalfitter import NormalFitter
 
 __all__ = [
+    "DistfitFitter",
     "NormalFitter",
     "MOMFitter",
 ]

skpro/distfitter/_distfit_adapter.py

@@ -0,0 +1,52 @@
+"""Generic scalar distribution wrapping a named scipy distribution.
+
+Used to turn the output of ``distfit`` (a scipy distribution name plus
+shape/loc/scale parameters) into a fitted ``skpro`` ``BaseDistribution``,
+by reusing the existing ``_ScipyAdapter`` machinery.
+"""
+# copyright: skpro developers, BSD-3-Clause License (see LICENSE file)
+
+from skpro.distributions.adapters.scipy import _ScipyAdapter
+
+__all__ = ["_DistfitDistribution"]
+
+
+class _DistfitDistribution(_ScipyAdapter):
+    """Scalar distribution wrapping a scipy distribution selected by ``distfit``.
+
+    Parameters
+    ----------
+    dist_name : str
+        Name of a distribution in ``scipy.stats``, e.g. ``"norm"``, ``"gamma"``.
+    shape_args : tuple, optional (default=())
+        Positional shape parameters for the ``scipy.stats`` distribution.
+    dist_loc : float, optional (default=0.0)
+        Location parameter, passed as ``loc`` to the ``scipy.stats`` distribution.
+    dist_scale : float, optional (default=1.0)
+        Scale parameter, passed as ``scale`` to the ``scipy.stats`` distribution.
+    """
+
+    _tags = {
+        "authors": ["areychana"],
+        "capabilities:exact": ["mean", "var", "pdf", "log_pdf", "cdf", "ppf"],
+        "distr:measuretype": "continuous",
+        "distr:paramtype": "parametric",
+    }
+
+    def __init__(self, dist_name, shape_args=(), dist_loc=0.0, dist_scale=1.0):
+        self.dist_name = dist_name
+        self.shape_args = shape_args
+        self.dist_loc = dist_loc
+        self.dist_scale = dist_scale
+
+        super().__init__(index=None, columns=None)
+
+    def _get_scipy_object(self):
+        import scipy.stats
+
+        return getattr(scipy.stats, self.dist_name)
+
+    def _get_scipy_param(self):
+        args = list(self.shape_args)
+        kwds = {"loc": self.dist_loc, "scale": self.dist_scale}
+        return args, kwds

skpro/distfitter/_distfitfitter.py

@@ -0,0 +1,147 @@
+"""Distribution fitter wrapping the distfit package."""
+# copyright: skpro developers, BSD-3-Clause License (see LICENSE file)
+
+from skpro.distfitter.base import BaseDistFitter
+
+__author__ = ["areychana"]
+
+
+class DistfitFitter(BaseDistFitter):
+    """Fit a parametric distribution to data using the ``distfit`` package.
+
+    Wraps `distfit <https://github.com/erdogant/distfit>`_'s parametric fitting
+    procedure: a set of candidate ``scipy.stats`` distributions is fitted to the
+    data, scored by a goodness-of-fit statistic, and the best-scoring
+    distribution is returned as a fitted ``skpro`` scalar distribution.
+
+    Only ``distfit``'s ``method="parametric"`` mode is currently supported.
+    ``distfit``'s ``"quantile"``, ``"percentile"``, and ``"discrete"`` modes
+    produce differently shaped model output that this fitter does not (yet)
+    convert into a ``skpro`` distribution.
+
+    Parameters
+    ----------
+    distr : str or list of str, optional (default="popular")
+        Candidate distribution(s) to fit and compare, passed to ``distfit``.
+        ``"popular"`` tests ``[norm, expon, pareto, dweibull, t, genextreme,
+        gamma, lognorm, beta, uniform, loggamma]``; ``"full"`` tests all
+        ``scipy.stats`` continuous distributions; a single name (e.g.
+        ``"norm"``) or a list of names restricts to those distributions.
+    stats : str, optional (default="RSS")
+        Goodness-of-fit statistic used by ``distfit`` to rank candidates.
+        One of ``"RSS"``, ``"wasserstein"``, ``"ks"``, ``"energy"``,
+        ``"goodness_of_fit"``.
+    bins : int or "auto", optional (default="auto")
+        Histogram bin size used internally by ``distfit``.
+    random_state : int, optional (default=None)
+        Random state passed to ``distfit``.
+
+    Attributes
+    ----------
+    dist_name_ : str
+        Name of the best-fitting ``scipy.stats`` distribution, as chosen by
+        ``distfit``.
+    shape_args_ : tuple
+        Fitted shape parameters for ``dist_name_``, as returned by ``distfit``.
+    dist_loc_ : float
+        Fitted location parameter for ``dist_name_``.
+    dist_scale_ : float
+        Fitted scale parameter for ``dist_name_``.
+    fit_summary_ : pandas.DataFrame
+        Full ``distfit`` summary table of all candidate distributions tried,
+        with their scores.
+
+    Examples
+    --------
+    >>> import pandas as pd
+    >>> from skpro.distfitter import DistfitFitter
+    >>> X = pd.DataFrame([1.0, 2.0, 3.0, 4.0, 5.0, 4.5, 3.5, 2.5, 1.5])
+
+    >>> fitter = DistfitFitter(distr="norm")
+    >>> fitter.fit(X)
+    DistfitFitter(distr='norm')
+    >>> dist = fitter.proba()
+    """
+
+    _tags = {
+        "authors": ["areychana"],
+        "python_dependencies": ["distfit"],
+    }
+
+    def __init__(self, distr="popular", stats="RSS", bins="auto", random_state=None):
+        self.distr = distr
+        self.stats = stats
+        self.bins = bins
+        self.random_state = random_state
+
+        super().__init__()
+
+    def _fit(self, X, C=None):
+        """Fit the best-scoring distfit distribution to the data.
+
+        Parameters
+        ----------
+        X : pandas DataFrame
+            Data to fit the distribution to.
+        C : ignored
+
+        Returns
+        -------
+        self : reference to self
+        """
+        from distfit import distfit as _distfit
+
+        vals = X.values.ravel()
+
+        dfit = _distfit(
+            method="parametric",
+            distr=self.distr,
+            stats=self.stats,
+            bins=self.bins,
+            random_state=self.random_state,
+            verbose="warning",
+        )
+        dfit.fit_transform(vals)
+
+        model = dfit.model
+        self.dist_name_ = model["name"]
+        self.shape_args_ = tuple(model["arg"])
+        self.dist_loc_ = float(model["loc"])
+        self.dist_scale_ = float(model["scale"])
+        self.fit_summary_ = dfit.summary
+
+        return self
+
+    def _proba(self):
+        """Return the best-fitting distribution found by distfit.
+
+        Returns
+        -------
+        dist : skpro BaseDistribution (scalar)
+        """
+        from skpro.distfitter._distfit_adapter import _DistfitDistribution
+
+        return _DistfitDistribution(
+            dist_name=self.dist_name_,
+            shape_args=self.shape_args_,
+            dist_loc=self.dist_loc_,
+            dist_scale=self.dist_scale_,
+        )
+
+    @classmethod
+    def get_test_params(cls, parameter_set="default"):
+        """Return testing parameter settings for the estimator.
+
+        Parameters
+        ----------
+        parameter_set : str, default="default"
+            Name of the set of test parameters to return.
+
+        Returns
+        -------
+        params : dict or list of dict
+            Parameters to create testing instances of the class.
+        """
+        params1 = {"distr": "norm"}
+        params2 = {"distr": ["norm", "expon"], "stats": "wasserstein"}
+        return [params1, params2]

skpro/distfitter/tests/test_distfitfitter.py

@@ -0,0 +1,86 @@
+"""Tests for the DistfitFitter distribution fitter."""
+# copyright: skpro developers, BSD-3-Clause License (see LICENSE file)
+
+import numpy as np
+import pandas as pd
+import pytest
+from skbase.utils.dependencies import _check_soft_dependencies
+
+from skpro.distributions.base import BaseDistribution
+
+DISTFIT_AVAILABLE = _check_soft_dependencies("distfit", severity="none")
+
+
+@pytest.mark.skipif(
+    not DISTFIT_AVAILABLE, reason="skip test if required soft dependency not present"
+)
+def test_distfitfitter_fits_known_normal():
+    """DistfitFitter restricted to 'norm' recovers close to true mean/scale."""
+    from skpro.distfitter import DistfitFitter
+
+    rng = np.random.RandomState(42)
+    X = pd.DataFrame(rng.normal(loc=5.0, scale=2.0, size=1000))
+
+    fitter = DistfitFitter(distr="norm")
+    fitter.fit(X)
+
+    assert fitter.dist_name_ == "norm"
+
+    dist = fitter.proba()
+
+    assert isinstance(dist, BaseDistribution)
+    assert dist.ndim == 0
+
+    mean = float(dist.mean())
+    var = float(dist.var())
+
+    assert np.isclose(mean, 5.0, atol=0.3)
+    assert np.isclose(var, 4.0, atol=1.0)
+
+
+@pytest.mark.skipif(
+    not DISTFIT_AVAILABLE, reason="skip test if required soft dependency not present"
+)
+def test_distfitfitter_stores_fit_summary():
+    """DistfitFitter stores a summary table of all candidate distributions."""
+    from skpro.distfitter import DistfitFitter
+
+    rng = np.random.RandomState(0)
+    X = pd.DataFrame(rng.normal(size=200))
+
+    fitter = DistfitFitter(distr=["norm", "expon"])
+    fitter.fit(X)
+
+    assert fitter.fit_summary_ is not None
+    assert len(fitter.fit_summary_) == 2
+    assert set(fitter.fit_summary_["name"]) == {"norm", "expon"}
+
+
+@pytest.mark.skipif(
+    not DISTFIT_AVAILABLE, reason="skip test if required soft dependency not present"
+)
+def test_distfitfitter_proba_before_fit_raises():
+    """Calling proba() before fit() must raise, per BaseDistFitter contract."""
+    from skpro.distfitter import DistfitFitter
+
+    fitter = DistfitFitter(distr="norm")
+
+    with pytest.raises(ValueError):
+        fitter.proba()
+
+
+@pytest.mark.skipif(
+    not DISTFIT_AVAILABLE, reason="skip test if required soft dependency not present"
+)
+def test_distfitfitter_get_params_roundtrip():
+    """get_params/set_params/clone round-trip as expected for skpro estimators."""
+    from skpro.distfitter import DistfitFitter
+
+    fitter = DistfitFitter(distr="norm", stats="wasserstein")
+    params = fitter.get_params()
+
+    assert params["distr"] == "norm"
+    assert params["stats"] == "wasserstein"
+
+    clone = fitter.clone()
+    assert clone.get_params() == fitter.get_params()