Merge pull request scikit-learn#5497 from MechCoder/ransac_residual

[MRG] Deprecate residual_metric and add support for loss in RANSAC

Author: MechCoder

doc/whats_new.rst

@@ -79,6 +79,9 @@ Enhancements
  - :class:`linear_model.RANSACRegressor` now supports ``sample_weights``.
  By `Imaculate`_.
 
+ - Add parameter ``loss`` to :class:`linear_model.RANSACRegressor` to measure the
+ error on the samples for every trial. By `Manoj Kumar`_.
+
 Bug fixes
 .........
 
@@ -114,6 +117,8 @@ API changes summary
  the :mod:`model_selection` module.
  (`#4294 <https://github.com/scikit-learn/scikit-learn/pull/4294>`_) by `Raghav R V`_.
 
+ - ``residual_metric`` has been deprecated in :class:`linear_model.RANSACRegressor`.
+ Use ``loss`` instead. By `Manoj Kumar`_.
 
 .. _changes_0_17:
 

sklearn/linear_model/ransac.py

@@ -5,6 +5,7 @@
 # License: BSD 3 clause
 
 import numpy as np
+import warnings
 
 from ..base import BaseEstimator, MetaEstimatorMixin, RegressorMixin, clone
 from ..utils import check_random_state, check_array, check_consistent_length
@@ -134,6 +135,22 @@ class RANSACRegressor(BaseEstimator, MetaEstimatorMixin, RegressorMixin):
 
  lambda dy: np.sum(np.abs(dy), axis=1)
 
+ NOTE: residual_metric is deprecated from 0.18 and will be removed in 0.20
+ Use ``loss`` instead.
+
+ loss: string, callable, optional, default "absolute_loss"
+ String inputs, "absolute_loss" and "squared_loss" are supported which
+ find the absolute loss and squared loss per sample
+ respectively.
+
+ If ``loss`` is a callable, then it should be a function that takes
+ two arrays as inputs, the true and predicted value and returns a 1-D
+ array with the ``i``th value of the array corresponding to the loss
+ on `X[i]`.
+
+ If the loss on a sample is greater than the ``residual_threshold``, then
+ this sample is classified as an outlier.
+
  random_state : integer or numpy.RandomState, optional
  The generator used to initialize the centers. If an integer is
  given, it fixes the seed. Defaults to the global numpy random
@@ -163,7 +180,7 @@ def __init__(self, base_estimator=None, min_samples=None,
  is_model_valid=None, max_trials=100,
  stop_n_inliers=np.inf, stop_score=np.inf,
  stop_probability=0.99, residual_metric=None,
- random_state=None):
+ loss='absolute_loss', random_state=None):
 
  self.base_estimator = base_estimator
  self.min_samples = min_samples
@@ -176,6 +193,7 @@ def __init__(self, base_estimator=None, min_samples=None,
  self.stop_probability = stop_probability
  self.residual_metric = residual_metric
  self.random_state = random_state
+ self.loss = loss
 
  def fit(self, X, y, sample_weight=None):
  """Fit estimator using RANSAC algorithm.
@@ -236,10 +254,33 @@ def fit(self, X, y, sample_weight=None):
  else:
  residual_threshold = self.residual_threshold
 
- if self.residual_metric is None:
- residual_metric = lambda dy: np.sum(np.abs(dy), axis=1)
+ if self.residual_metric is not None:
+ warnings.warn(
+ "'residual_metric' will be removed in version 0.20. Use "
+ "'loss' instead.", DeprecationWarning)
+
+ if self.loss == "absolute_loss":
+ if y.ndim == 1:
+ loss_function = lambda y_true, y_pred: np.abs(y_true - y_pred)
+ else:
+ loss_function = lambda \
+ y_true, y_pred: np.sum(np.abs(y_true - y_pred), axis=1)
+
+ elif self.loss == "squared_loss":
+ if y.ndim == 1:
+ loss_function = lambda y_true, y_pred: (y_true - y_pred) ** 2
+ else:
+ loss_function = lambda \
+ y_true, y_pred: np.sum((y_true - y_pred) ** 2, axis=1)
+
+ elif callable(self.loss):
+ loss_function = self.loss
+
  else:
- residual_metric = self.residual_metric
+ raise ValueError(
+ "loss should be 'absolute_loss', 'squared_loss' or a callable."
+ "Got %s. " % self.loss)
+
 
  random_state = check_random_state(self.random_state)
 
@@ -298,10 +339,15 @@ def fit(self, X, y, sample_weight=None):
 
  # residuals of all data for current random sample model
  y_pred = base_estimator.predict(X)
- diff = y_pred - y
- if diff.ndim == 1:
- diff = diff.reshape(-1, 1)
- residuals_subset = residual_metric(diff)
+
+ # XXX: Deprecation: Remove this if block in 0.20
+ if self.residual_metric is not None:
+ diff = y_pred - y
+ if diff.ndim == 1:
+ diff = diff.reshape(-1, 1)
+ residuals_subset = self.residual_metric(diff)
+ else:
+ residuals_subset = loss_function(y, y_pred)
 
  # classify data into inliers and outliers
  inlier_mask_subset = residuals_subset < residual_threshold

sklearn/linear_model/tests/test_ransac.py

@@ -1,9 +1,18 @@
+from scipy import sparse
+
 import numpy as np
-from numpy.testing import assert_equal, assert_raises, assert_array_equal,assert_array_almost_equal
-from sklearn.utils.testing import assert_raises_regexp, assert_almost_equal, assert_less
 from scipy import sparse
+
+from numpy.testing import assert_equal, assert_raises
+from numpy.testing import assert_array_almost_equal
+from numpy.testing import assert_array_equal
+
 from sklearn.utils import check_random_state
-from sklearn.linear_model import LinearRegression, RANSACRegressor,Lasso
+from sklearn.utils.testing import assert_raises_regexp
+from sklearn.utils.testing import assert_less
+from sklearn.utils.testing import assert_warns
+from sklearn.utils.testing import assert_almost_equal
+from sklearn.linear_model import LinearRegression, RANSACRegressor, Lasso
 from sklearn.linear_model.ransac import _dynamic_max_trials
 
 
@@ -265,6 +274,7 @@ def test_ransac_multi_dimensional_targets():
  assert_equal(ransac_estimator.inlier_mask_, ref_inlier_mask)
 
 
+# XXX: Remove in 0.20
 def test_ransac_residual_metric():
  residual_metric1 = lambda dy: np.sum(np.abs(dy), axis=1)
  residual_metric2 = lambda dy: np.sum(dy ** 2, axis=1)
@@ -281,6 +291,38 @@ def test_ransac_residual_metric():
  residual_threshold=5, random_state=0,
  residual_metric=residual_metric2)
 
+ # multi-dimensional
+ ransac_estimator0.fit(X, yyy)
+ assert_warns(DeprecationWarning, ransac_estimator1.fit, X, yyy)
+ assert_warns(DeprecationWarning, ransac_estimator2.fit, X, yyy)
+ assert_array_almost_equal(ransac_estimator0.predict(X),
+ ransac_estimator1.predict(X))
+ assert_array_almost_equal(ransac_estimator0.predict(X),
+ ransac_estimator2.predict(X))
+
+ # one-dimensional
+ ransac_estimator0.fit(X, y)
+ assert_warns(DeprecationWarning, ransac_estimator2.fit, X, y)
+ assert_array_almost_equal(ransac_estimator0.predict(X),
+ ransac_estimator2.predict(X))
+
+def test_ransac_residual_loss():
+ loss_multi1 = lambda y_true, y_pred: np.sum(np.abs(y_true - y_pred), axis=1)
+ loss_multi2 = lambda y_true, y_pred: np.sum((y_true - y_pred) ** 2, axis=1)
+
+ loss_mono = lambda y_true, y_pred : np.abs(y_true - y_pred)
+ yyy = np.column_stack([y, y, y])
+
+ base_estimator = LinearRegression()
+ ransac_estimator0 = RANSACRegressor(base_estimator, min_samples=2,
+ residual_threshold=5, random_state=0)
+ ransac_estimator1 = RANSACRegressor(base_estimator, min_samples=2,
+ residual_threshold=5, random_state=0,
+ loss=loss_multi1)
+ ransac_estimator2 = RANSACRegressor(base_estimator, min_samples=2,
+ residual_threshold=5, random_state=0,
+ loss=loss_multi2)
+
  # multi-dimensional
  ransac_estimator0.fit(X, yyy)
  ransac_estimator1.fit(X, yyy)
@@ -292,9 +334,16 @@ def test_ransac_residual_metric():
 
  # one-dimensional
  ransac_estimator0.fit(X, y)
+ ransac_estimator2.loss = loss_mono
  ransac_estimator2.fit(X, y)
  assert_array_almost_equal(ransac_estimator0.predict(X),
  ransac_estimator2.predict(X))
+ ransac_estimator3 = RANSACRegressor(base_estimator, min_samples=2,
+ residual_threshold=5, random_state=0,
+ loss="squared_loss")
+ ransac_estimator3.fit(X, y)
+ assert_array_almost_equal(ransac_estimator0.predict(X),
+ ransac_estimator2.predict(X))
 
 
 def test_ransac_default_residual_threshold():