Merge pull request #2454 from janezd/distances

[ENH] Proper calculation of distances

Author: astaric

Orange/distance/__init__.py

@@ -1,261 +1,6 @@
-import numpy as np
-from scipy import stats
-import sklearn.metrics as skl_metrics
+from .distance import (Distance, DistanceModel,
+                       Euclidean, Manhattan, Cosine, Jaccard,
+                       SpearmanR, SpearmanRAbsolute, PearsonR, PearsonRAbsolute,
+                       Mahalanobis, MahalanobisDistance)
 
-from Orange import data
-from Orange.misc import DistMatrix
-from Orange.preprocess import SklImpute
-
-__all__ = ['Euclidean', 'Manhattan', 'Cosine', 'Jaccard', 'SpearmanR',
-           'SpearmanRAbsolute', 'PearsonR', 'PearsonRAbsolute', 'Mahalanobis',
-           'MahalanobisDistance']
-
-
-def _preprocess(table):
-    """Remove categorical attributes and impute missing values."""
-    if not len(table):
-        return table
-    new_domain = data.Domain(
-        [a for a in table.domain.attributes if a.is_continuous],
-        table.domain.class_vars,
-        table.domain.metas)
-    new_data = table.transform(new_domain)
-    new_data = SklImpute()(new_data)
-    return new_data
-
-
-def _orange_to_numpy(x):
-    """Convert :class:`Orange.data.Table` and :class:`Orange.data.RowInstance`
-    to :class:`numpy.ndarray`.
-    """
-    if isinstance(x, data.Table):
-        return x.X
-    elif isinstance(x, data.Instance):
-        return np.atleast_2d(x.x)
-    elif isinstance(x, np.ndarray):
-        return np.atleast_2d(x)
-    else:
-        return x    # e.g. None
-
-
-class Distance:
-    def __call__(self, e1, e2=None, axis=1, impute=False):
-        """
-        :param e1: input data instances, we calculate distances between all
-            pairs
-        :type e1: :class:`Orange.data.Table` or
-            :class:`Orange.data.RowInstance` or :class:`numpy.ndarray`
-        :param e2: optional second argument for data instances if provided,
-            distances between each pair, where first item is from e1 and
-            second is from e2, are calculated
-        :type e2: :class:`Orange.data.Table` or
-            :class:`Orange.data.RowInstance` or :class:`numpy.ndarray`
-        :param axis: if axis=1 we calculate distances between rows, if axis=0
-            we calculate distances between columns
-        :type axis: int
-        :param impute: if impute=True all NaN values in matrix are replaced
-            with 0
-        :type impute: bool
-        :return: the matrix with distances between given examples
-        :rtype: :class:`Orange.misc.distmatrix.DistMatrix`
-        """
-        raise NotImplementedError(
-            'Distance is an abstract class and should not be used directly.')
-
-
-class SklDistance(Distance):
-    """Generic scikit-learn distance."""
-    def __init__(self, metric, name, supports_sparse):
-        """
-        Args:
-            metric: The metric to be used for distance calculation
-            name (str): Name of the distance
-            supports_sparse (boolean): Whether this metric works on sparse data
-                or not.
-        """
-        self.metric = metric
-        self.name = name
-        self.supports_sparse = supports_sparse
-
-    def __call__(self, e1, e2=None, axis=1, impute=False):
-        x1 = _orange_to_numpy(e1)
-        x2 = _orange_to_numpy(e2)
-        if axis == 0:
-            x1 = x1.T
-            if x2 is not None:
-                x2 = x2.T
-        dist = skl_metrics.pairwise.pairwise_distances(
-                x1, x2, metric=self.metric)
-        if isinstance(e1, data.Table) or isinstance(e1, data.RowInstance):
-            dist = DistMatrix(dist, e1, e2, axis)
-        else:
-            dist = DistMatrix(dist)
-        return dist
-
-Euclidean = SklDistance('euclidean', 'Euclidean', True)
-Manhattan = SklDistance('manhattan', 'Manhattan', True)
-Cosine = SklDistance('cosine', 'Cosine', True)
-Jaccard = SklDistance('jaccard', 'Jaccard', False)
-
-
-class SpearmanDistance(Distance):
-    """ Generic Spearman's rank correlation coefficient. """
-    def __init__(self, absolute, name):
-        """
-        Constructor for Spearman's and Absolute Spearman's distances.
-
-        Args:
-            absolute (boolean): Whether to use absolute values or not.
-            name (str): Name of the distance
-
-        Returns:
-            If absolute=True return Spearman's Absolute rank class else return
-                Spearman's rank class.
-        """
-        self.absolute = absolute
-        self.name = name
-        self.supports_sparse = False
-
-    def __call__(self, e1, e2=None, axis=1, impute=False):
-        x1 = _orange_to_numpy(e1)
-        x2 = _orange_to_numpy(e2)
-        if x2 is None:
-            x2 = x1
-        slc = len(x1) if axis == 1 else x1.shape[1]
-        rho, _ = stats.spearmanr(x1, x2, axis=axis)
-        if np.isnan(rho).any() and impute:
-            rho = np.nan_to_num(rho)
-        if self.absolute:
-            dist = (1. - np.abs(rho)) / 2.
-        else:
-            dist = (1. - rho) / 2.
-        if isinstance(dist, np.float):
-            dist = np.array([[dist]])
-        elif isinstance(dist, np.ndarray):
-            dist = dist[:slc, slc:]
-        if isinstance(e1, data.Table) or isinstance(e1, data.RowInstance):
-            dist = DistMatrix(dist, e1, e2, axis)
-        else:
-            dist = DistMatrix(dist)
-        return dist
-
-SpearmanR = SpearmanDistance(absolute=False, name='Spearman')
-SpearmanRAbsolute = SpearmanDistance(absolute=True, name='Spearman absolute')
-
-
-class PearsonDistance(Distance):
-    """ Generic Pearson's rank correlation coefficient. """
-    def __init__(self, absolute, name):
-        """
-        Constructor for Pearson's and Absolute Pearson's distances.
-
-        Args:
-            absolute (boolean): Whether to use absolute values or not.
-            name (str): Name of the distance
-
-        Returns:
-            If absolute=True return Pearson's Absolute rank class else return
-                Pearson's rank class.
-        """
-        self.absolute = absolute
-        self.name = name
-        self.supports_sparse = False
-
-    def __call__(self, e1, e2=None, axis=1, impute=False):
-        x1 = _orange_to_numpy(e1)
-        x2 = _orange_to_numpy(e2)
-        if x2 is None:
-            x2 = x1
-        if axis == 0:
-            x1 = x1.T
-            x2 = x2.T
-        rho = np.array([[stats.pearsonr(i, j)[0] for j in x2] for i in x1])
-        if np.isnan(rho).any() and impute:
-            rho = np.nan_to_num(rho)
-        if self.absolute:
-            dist = (1. - np.abs(rho)) / 2.
-        else:
-            dist = (1. - rho) / 2.
-        if isinstance(e1, data.Table) or isinstance(e1, data.RowInstance):
-            dist = DistMatrix(dist, e1, e2, axis)
-        else:
-            dist = DistMatrix(dist)
-        return dist
-
-PearsonR = PearsonDistance(absolute=False, name='Pearson')
-PearsonRAbsolute = PearsonDistance(absolute=True, name='Pearson absolute')
-
-
-class MahalanobisDistance(Distance):
-    """Mahalanobis distance."""
-    def __init__(self, data=None, axis=1, name='Mahalanobis'):
-        self.name = name
-        self.supports_sparse = False
-        self.axis = None
-        self.VI = None
-        if data is not None:
-            self.fit(data, axis)
-
-    def fit(self, data, axis=1):
-        """
-        Compute the covariance matrix needed for calculating distances.
-
-        Args:
-            data: The dataset used for calculating covariances.
-            axis: If axis=1 we calculate distances between rows, if axis=0 we
-                calculate distances between columns.
-        """
-        x = _orange_to_numpy(data)
-        if axis == 0:
-            x = x.T
-        self.axis = axis
-        try:
-            c = np.cov(x.T)
-        except:
-            raise MemoryError("Covariance matrix is too large.")
-        try:
-            self.VI = np.linalg.inv(c)
-        except:
-            raise ValueError("Computation of inverse covariance matrix failed.")
-
-    def __call__(self, e1, e2=None, axis=None, impute=False):
-        assert self.VI is not None, \
-            "Mahalanobis distance must be initialized with the fit() method."
-
-        x1 = _orange_to_numpy(e1)
-        x2 = _orange_to_numpy(e2)
-
-        if axis is not None:
-            assert axis == self.axis, \
-                "Axis must match its value at initialization."
-        if self.axis == 0:
-            x1 = x1.T
-            if x2 is not None:
-                x2 = x2.T
-        if not x1.shape[1] == self.VI.shape[0] or \
-                x2 is not None and not x2.shape[1] == self.VI.shape[0]:
-            raise ValueError('Incorrect number of features.')
-
-        dist = skl_metrics.pairwise.pairwise_distances(
-                x1, x2, metric='mahalanobis', VI=self.VI)
-        if np.isnan(dist).any() and impute:
-            dist = np.nan_to_num(dist)
-        if isinstance(e1, data.Table) or isinstance(e1, data.RowInstance):
-            dist = DistMatrix(dist, e1, e2, self.axis)
-        else:
-            dist = DistMatrix(dist)
-        return dist
-
-
-# Only retain this to raise errors on use. Remove in some future version.
-class __MahalanobisDistanceError(MahalanobisDistance):
-    def _raise_error(self, *args, **kwargs):
-        raise RuntimeError(
-            "Invalid use of MahalanobisDistance.\n"
-            "Create a new MahalanobisDistance instance first, e.g.\n"
-            ">>> metric = MahalanobisDistance(data)\n"
-            ">>> dist = metric(data)"
-        )
-    fit = _raise_error
-    __call__ = _raise_error
-Mahalanobis = __MahalanobisDistanceError()
+from .base import _preprocess, remove_discrete_features, impute