375 lines
14 KiB
Python
375 lines
14 KiB
Python
# coding: utf-8
|
|
|
|
# Author: Johannes Schönberger
|
|
#
|
|
# License: BSD 3 clause
|
|
|
|
import numpy as np
|
|
|
|
from ..base import BaseEstimator, MetaEstimatorMixin, RegressorMixin, clone
|
|
from ..utils import check_random_state, check_array, check_consistent_length
|
|
from ..utils.random import sample_without_replacement
|
|
from .base import LinearRegression
|
|
|
|
|
|
_EPSILON = np.spacing(1)
|
|
|
|
|
|
def _dynamic_max_trials(n_inliers, n_samples, min_samples, probability):
|
|
"""Determine number trials such that at least one outlier-free subset is
|
|
sampled for the given inlier/outlier ratio.
|
|
|
|
Parameters
|
|
----------
|
|
n_inliers : int
|
|
Number of inliers in the data.
|
|
|
|
n_samples : int
|
|
Total number of samples in the data.
|
|
|
|
min_samples : int
|
|
Minimum number of samples chosen randomly from original data.
|
|
|
|
probability : float
|
|
Probability (confidence) that one outlier-free sample is generated.
|
|
|
|
Returns
|
|
-------
|
|
trials : int
|
|
Number of trials.
|
|
|
|
"""
|
|
inlier_ratio = n_inliers / float(n_samples)
|
|
nom = max(_EPSILON, 1 - probability)
|
|
denom = max(_EPSILON, 1 - inlier_ratio ** min_samples)
|
|
if nom == 1:
|
|
return 0
|
|
if denom == 1:
|
|
return float('inf')
|
|
return abs(float(np.ceil(np.log(nom) / np.log(denom))))
|
|
|
|
|
|
class RANSACRegressor(BaseEstimator, MetaEstimatorMixin, RegressorMixin):
|
|
"""RANSAC (RANdom SAmple Consensus) algorithm.
|
|
|
|
RANSAC is an iterative algorithm for the robust estimation of parameters
|
|
from a subset of inliers from the complete data set. More information can
|
|
be found in the general documentation of linear models.
|
|
|
|
A detailed description of the algorithm can be found in the documentation
|
|
of the ``linear_model`` sub-package.
|
|
|
|
Parameters
|
|
----------
|
|
base_estimator : object, optional
|
|
Base estimator object which implements the following methods:
|
|
|
|
* `fit(X, y)`: Fit model to given training data and target values.
|
|
* `score(X, y)`: Returns the mean accuracy on the given test data,
|
|
which is used for the stop criterion defined by `stop_score`.
|
|
Additionally, the score is used to decide which of two equally
|
|
large consensus sets is chosen as the better one.
|
|
|
|
If `base_estimator` is None, then
|
|
``base_estimator=sklearn.linear_model.LinearRegression()`` is used for
|
|
target values of dtype float.
|
|
|
|
Note that the current implementation only supports regression
|
|
estimators.
|
|
|
|
min_samples : int (>= 1) or float ([0, 1]), optional
|
|
Minimum number of samples chosen randomly from original data. Treated
|
|
as an absolute number of samples for `min_samples >= 1`, treated as a
|
|
relative number `ceil(min_samples * X.shape[0]`) for
|
|
`min_samples < 1`. This is typically chosen as the minimal number of
|
|
samples necessary to estimate the given `base_estimator`. By default a
|
|
``sklearn.linear_model.LinearRegression()`` estimator is assumed and
|
|
`min_samples` is chosen as ``X.shape[1] + 1``.
|
|
|
|
residual_threshold : float, optional
|
|
Maximum residual for a data sample to be classified as an inlier.
|
|
By default the threshold is chosen as the MAD (median absolute
|
|
deviation) of the target values `y`.
|
|
|
|
is_data_valid : callable, optional
|
|
This function is called with the randomly selected data before the
|
|
model is fitted to it: `is_data_valid(X, y)`. If its return value is
|
|
False the current randomly chosen sub-sample is skipped.
|
|
|
|
is_model_valid : callable, optional
|
|
This function is called with the estimated model and the randomly
|
|
selected data: `is_model_valid(model, X, y)`. If its return value is
|
|
False the current randomly chosen sub-sample is skipped.
|
|
Rejecting samples with this function is computationally costlier than
|
|
with `is_data_valid`. `is_model_valid` should therefore only be used if
|
|
the estimated model is needed for making the rejection decision.
|
|
|
|
max_trials : int, optional
|
|
Maximum number of iterations for random sample selection.
|
|
|
|
stop_n_inliers : int, optional
|
|
Stop iteration if at least this number of inliers are found.
|
|
|
|
stop_score : float, optional
|
|
Stop iteration if score is greater equal than this threshold.
|
|
|
|
stop_probability : float in range [0, 1], optional
|
|
RANSAC iteration stops if at least one outlier-free set of the training
|
|
data is sampled in RANSAC. This requires to generate at least N
|
|
samples (iterations)::
|
|
|
|
N >= log(1 - probability) / log(1 - e**m)
|
|
|
|
where the probability (confidence) is typically set to high value such
|
|
as 0.99 (the default) and e is the current fraction of inliers w.r.t.
|
|
the total number of samples.
|
|
|
|
residual_metric : callable, optional
|
|
Metric to reduce the dimensionality of the residuals to 1 for
|
|
multi-dimensional target values ``y.shape[1] > 1``. By default the sum
|
|
of absolute differences is used::
|
|
|
|
lambda dy: np.sum(np.abs(dy), axis=1)
|
|
|
|
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
|
|
number generator.
|
|
|
|
Attributes
|
|
----------
|
|
estimator_ : object
|
|
Best fitted model (copy of the `base_estimator` object).
|
|
|
|
n_trials_ : int
|
|
Number of random selection trials until one of the stop criteria is
|
|
met. It is always ``<= max_trials``.
|
|
|
|
inlier_mask_ : bool array of shape [n_samples]
|
|
Boolean mask of inliers classified as ``True``.
|
|
|
|
References
|
|
----------
|
|
.. [1] http://en.wikipedia.org/wiki/RANSAC
|
|
.. [2] http://www.cs.columbia.edu/~belhumeur/courses/compPhoto/ransac.pdf
|
|
.. [3] http://www.bmva.org/bmvc/2009/Papers/Paper355/Paper355.pdf
|
|
"""
|
|
|
|
def __init__(self, base_estimator=None, min_samples=None,
|
|
residual_threshold=None, is_data_valid=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):
|
|
|
|
self.base_estimator = base_estimator
|
|
self.min_samples = min_samples
|
|
self.residual_threshold = residual_threshold
|
|
self.is_data_valid = is_data_valid
|
|
self.is_model_valid = is_model_valid
|
|
self.max_trials = max_trials
|
|
self.stop_n_inliers = stop_n_inliers
|
|
self.stop_score = stop_score
|
|
self.stop_probability = stop_probability
|
|
self.residual_metric = residual_metric
|
|
self.random_state = random_state
|
|
|
|
def fit(self, X, y):
|
|
"""Fit estimator using RANSAC algorithm.
|
|
|
|
Parameters
|
|
----------
|
|
X : array-like or sparse matrix, shape [n_samples, n_features]
|
|
Training data.
|
|
|
|
y : array-like, shape = [n_samples] or [n_samples, n_targets]
|
|
Target values.
|
|
|
|
Raises
|
|
------
|
|
ValueError
|
|
If no valid consensus set could be found. This occurs if
|
|
`is_data_valid` and `is_model_valid` return False for all
|
|
`max_trials` randomly chosen sub-samples.
|
|
|
|
"""
|
|
X = check_array(X, accept_sparse='csr')
|
|
y = check_array(y, ensure_2d=False)
|
|
if y.ndim == 1:
|
|
y = y.reshape(-1, 1)
|
|
check_consistent_length(X, y)
|
|
|
|
if self.base_estimator is not None:
|
|
base_estimator = clone(self.base_estimator)
|
|
else:
|
|
base_estimator = LinearRegression()
|
|
|
|
if self.min_samples is None:
|
|
# assume linear model by default
|
|
min_samples = X.shape[1] + 1
|
|
elif 0 < self.min_samples < 1:
|
|
min_samples = np.ceil(self.min_samples * X.shape[0])
|
|
elif self.min_samples >= 1:
|
|
if self.min_samples % 1 != 0:
|
|
raise ValueError("Absolute number of samples must be an "
|
|
"integer value.")
|
|
min_samples = self.min_samples
|
|
else:
|
|
raise ValueError("Value for `min_samples` must be scalar and "
|
|
"positive.")
|
|
if min_samples > X.shape[0]:
|
|
raise ValueError("`min_samples` may not be larger than number "
|
|
"of samples ``X.shape[0]``.")
|
|
|
|
if self.stop_probability < 0 or self.stop_probability > 1:
|
|
raise ValueError("`stop_probability` must be in range [0, 1].")
|
|
|
|
if self.residual_threshold is None:
|
|
# MAD (median absolute deviation)
|
|
residual_threshold = np.median(np.abs(y - np.median(y)))
|
|
else:
|
|
residual_threshold = self.residual_threshold
|
|
|
|
if self.residual_metric is None:
|
|
residual_metric = lambda dy: np.sum(np.abs(dy), axis=1)
|
|
else:
|
|
residual_metric = self.residual_metric
|
|
|
|
random_state = check_random_state(self.random_state)
|
|
|
|
try: # Not all estimator accept a random_state
|
|
base_estimator.set_params(random_state=random_state)
|
|
except ValueError:
|
|
pass
|
|
|
|
n_inliers_best = 0
|
|
score_best = np.inf
|
|
inlier_mask_best = None
|
|
X_inlier_best = None
|
|
y_inlier_best = None
|
|
|
|
# number of data samples
|
|
n_samples = X.shape[0]
|
|
sample_idxs = np.arange(n_samples)
|
|
|
|
n_samples, _ = X.shape
|
|
|
|
for self.n_trials_ in range(1, self.max_trials + 1):
|
|
|
|
# choose random sample set
|
|
subset_idxs = sample_without_replacement(n_samples, min_samples,
|
|
random_state=random_state)
|
|
X_subset = X[subset_idxs]
|
|
y_subset = y[subset_idxs]
|
|
|
|
# check if random sample set is valid
|
|
if (self.is_data_valid is not None
|
|
and not self.is_data_valid(X_subset, y_subset)):
|
|
continue
|
|
|
|
# fit model for current random sample set
|
|
base_estimator.fit(X_subset, y_subset)
|
|
|
|
# check if estimated model is valid
|
|
if (self.is_model_valid is not None and not
|
|
self.is_model_valid(base_estimator, X_subset, y_subset)):
|
|
continue
|
|
|
|
# residuals of all data for current random sample model
|
|
y_pred = base_estimator.predict(X)
|
|
if y_pred.ndim == 1:
|
|
y_pred = y_pred[:, None]
|
|
|
|
residuals_subset = residual_metric(y_pred - y)
|
|
|
|
# classify data into inliers and outliers
|
|
inlier_mask_subset = residuals_subset < residual_threshold
|
|
n_inliers_subset = np.sum(inlier_mask_subset)
|
|
|
|
# less inliers -> skip current random sample
|
|
if n_inliers_subset < n_inliers_best:
|
|
continue
|
|
|
|
# extract inlier data set
|
|
inlier_idxs_subset = sample_idxs[inlier_mask_subset]
|
|
X_inlier_subset = X[inlier_idxs_subset]
|
|
y_inlier_subset = y[inlier_idxs_subset]
|
|
|
|
# score of inlier data set
|
|
score_subset = base_estimator.score(X_inlier_subset,
|
|
y_inlier_subset)
|
|
|
|
# same number of inliers but worse score -> skip current random
|
|
# sample
|
|
if (n_inliers_subset == n_inliers_best
|
|
and score_subset < score_best):
|
|
continue
|
|
|
|
# save current random sample as best sample
|
|
n_inliers_best = n_inliers_subset
|
|
score_best = score_subset
|
|
inlier_mask_best = inlier_mask_subset
|
|
X_inlier_best = X_inlier_subset
|
|
y_inlier_best = y_inlier_subset
|
|
|
|
# break if sufficient number of inliers or score is reached
|
|
if (n_inliers_best >= self.stop_n_inliers
|
|
or score_best >= self.stop_score
|
|
or self.n_trials_
|
|
>= _dynamic_max_trials(n_inliers_best, n_samples,
|
|
min_samples,
|
|
self.stop_probability)):
|
|
break
|
|
|
|
# if none of the iterations met the required criteria
|
|
if inlier_mask_best is None:
|
|
raise ValueError(
|
|
"RANSAC could not find valid consensus set, because"
|
|
" either the `residual_threshold` rejected all the samples or"
|
|
" `is_data_valid` and `is_model_valid` returned False for all"
|
|
" `max_trials` randomly ""chosen sub-samples. Consider "
|
|
"relaxing the ""constraints.")
|
|
|
|
# estimate final model using all inliers
|
|
base_estimator.fit(X_inlier_best, y_inlier_best)
|
|
|
|
self.estimator_ = base_estimator
|
|
self.inlier_mask_ = inlier_mask_best
|
|
return self
|
|
|
|
def predict(self, X):
|
|
"""Predict using the estimated model.
|
|
|
|
This is a wrapper for `estimator_.predict(X)`.
|
|
|
|
Parameters
|
|
----------
|
|
X : numpy array of shape [n_samples, n_features]
|
|
|
|
Returns
|
|
-------
|
|
y : array, shape = [n_samples] or [n_samples, n_targets]
|
|
Returns predicted values.
|
|
"""
|
|
return self.estimator_.predict(X)
|
|
|
|
def score(self, X, y):
|
|
"""Returns the score of the prediction.
|
|
|
|
This is a wrapper for `estimator_.score(X, y)`.
|
|
|
|
Parameters
|
|
----------
|
|
X : numpy array or sparse matrix of shape [n_samples, n_features]
|
|
Training data.
|
|
|
|
y : array, shape = [n_samples] or [n_samples, n_targets]
|
|
Target values.
|
|
|
|
Returns
|
|
-------
|
|
z : float
|
|
Score of the prediction.
|
|
"""
|
|
return self.estimator_.score(X, y)
|