Select Git revision
MinCQGraalpy.py
-
Baptiste Bauvin authoredBaptiste Bauvin authored
MinCQGraalpy.py 17.01 KiB
# -*- coding: utf-8 -*-
"""MinCq algorithm.
Related papers:
[1] From PAC-Bayes Bounds to Quadratic Programs for Majority Votes (Laviolette et al., 2011)
[2] Risk Bounds for the Majority Vote: From a PAC-Bayesian Analysis to a Learning Algorithm (Germain et al., 2015)
"""
from __future__ import print_function, division, absolute_import
import logging
from operator import xor
import numpy as np
from scipy.linalg import sqrtm
from scipy.spatial.distance import pdist, squareform
from sklearn.metrics.pairwise import rbf_kernel
from sklearn.utils.validation import check_X_y
from sklearn.ensemble import VotingClassifier
from sklearn.manifold import SpectralEmbedding
from sklearn.utils.graph import graph_laplacian
from sklearn.preprocessing import LabelEncoder
from ..Monoview.Additions.BoostUtils import ConvexProgram, StumpsClassifiersGenerator
from ..Monoview.MonoviewUtils import BaseMonoviewClassifier, CustomUniform
from ..Metrics import zero_one_loss
# logger = logging.getLogger('MinCq')
class MinCqClassifier(VotingClassifier):
"""
Base MinCq algorithm learner. See [1, 2].
This version is an attempt of creating a more general version of MinCq, that handles multiclass classfication.
For binary classification, use RegularizedMinCqClassifer.
Parameters
----------
mu : float
The fixed value of the first moment of the margin.
"""
def __init__(self, estimators_generator=None, estimators=None, mu=0.001, omega=0.5, use_binary=False, zeta=0, gamma=1, n_neighbors=5):
if estimators is None:
estimators = []
super().__init__(estimators=estimators, voting='soft', flatten_transform=False)
self.estimators_generator = estimators_generator
self.mu = mu
self.omega = omega
self.use_binary = use_binary
self.zeta = zeta
self.gamma = gamma
self.n_neighbors = n_neighbors
def fit(self, X, y):
"""Fit the estimators and learn the weights.
Parameters
----------
X : array-like, shape = [n_samples, n_features]
Training vectors, where n_samples is the number of samples and
n_features is the number of features.
y : array-like, shape = [n_samples]
Target values. If y is a masked-array (numpy.ma), the masked values are unlabeled examples.
Returns
-------
self : object
"""
# Validations
assert 0 < self.mu <= 1, "MinCqClassifier: mu parameter must be in (0, 1]"
assert xor(bool(self.estimators_generator), bool(self.estimators)), "MinCqClassifier: exactly one of estimator_generator or estimators must be used."
X, y = check_X_y(X, y)
# Fit the estimators using VotingClassifier's fit method. This will also fit a LabelEncoder that can be
# used to "normalize" labels (0, 1, 2, ...). In the case of binary classification, the two classes will be 0 and 1.
# First, ensure that the weights are reset to None (as cloning a VotingClassifier keeps the weights)
self.weights = None
# TODO: Ensure estimators can deal with masked arrays
# If we use an estimator generator, use the data-dependant estimator generator to generate them, and fit again.
if self.estimators:
super().fit(X, y)
else:
self.le_ = LabelEncoder()
self.le_.fit(y)
self.clean_me =True
if isinstance(y, np.ma.MaskedArray):
transformed_y = np.ma.MaskedArray(self.le_.transform(y), y.mask)
else:
# transformed_y = self.le_.transform(y)
transformed_y = y
self.estimators_generator.fit(X, transformed_y)
self.estimators = [('ds{}'.format(i), estimator) for i, estimator in enumerate(self.estimators_generator.estimators_)]
super().fit(X, y)
# We clean the estimators attribute (as we do not want it to be cloned later)
# self.estimators_ = []
# logger.info("Training started...")
# logger.info("Training dataset shape: {}".format(str(np.shape(X))))
# logger.info("Number of voters: {}".format(len(self.estimators_)))
# Preparation and resolution of the quadratic program
# logger.info("Preparing and solving QP...")
self.weights = self._solve(X, y)
if self.clean_me:
self.estimators = []
return self
# def evaluate_metrics(self, X, y, metrics_list=None, functions_list=None):
# if metrics_list is None:
# metrics_list = [zero_one_loss]
#
# if functions_list is None:
# functions_list = []
# else:
# raise NotImplementedError
#
# # Predict, evaluate metrics.
# predictions = self.predict(X)
# metrics_results = {metric.__name__: metric(y, predictions) for metric in metrics_list}
#
# metrics_dataframe = ResultsDataFrame([metrics_results])
# return metrics_dataframe
def _binary_classification_matrix(self, X):
probas = self.transform(X)
predicted_labels = np.argmax(probas, axis=2)
predicted_labels[predicted_labels == 0] = -1
values = np.max(probas, axis=2)
return (predicted_labels * values).T
def _multiclass_classification_matrix(self, X, y):
probas = self.transform(X).swapaxes(0, 1)
matrix = probas[np.arange(probas.shape[0]), :, y]
return (matrix - self.omega)
def predict(self, X):
if not self.estimators:
self.estimators = [('ds{}'.format(i), estimator) for i, estimator in
enumerate(self.estimators_generator.estimators_)]
self.clean_me = True
pred = super().predict(X)
if self.clean_me:
self.estimators = []
return pred
def _solve(self, X, y):
y = self.le_.transform(y)
if self.use_binary:
assert len(self.le_.classes_) == 2
# TODO: Review the number of labeled examples when adding back the support for transductive learning.
classification_matrix = self._binary_classification_matrix(X)
# We use {-1, 1} labels.
binary_labels = np.copy(y)
binary_labels[y == 0] = -1
multi_matrix = binary_labels.reshape((len(binary_labels), 1)) * classification_matrix
else:
multi_matrix = self._multiclass_classification_matrix(X, y)
n_examples, n_voters = np.shape(multi_matrix)
ftf = 1.0 / n_examples * multi_matrix.T.dot(multi_matrix)
yf = np.mean(multi_matrix, axis=0)
# Objective function.
objective_matrix = 2 * ftf
objective_vector = None
# Equality constraints (first moment of the margin equal to mu, Q sums to one)
equality_matrix = np.vstack((yf.reshape((1, n_voters)), np.ones((1, n_voters))))
equality_vector = np.array([self.mu, 1.0])
# Lower and upper bounds, no quasi-uniformity.
lower_bound = 0.0
# TODO: In the case of binary classification, no upper bound will give
# bad results. Using 1/n works, as it brings back the l_infinity
# regularization normally given by the quasi-uniformity constraint.
# upper_bound = 2.0/n_voters
upper_bound = None
weights = self._solve_qp(objective_matrix, objective_vector, equality_matrix, equality_vector, lower_bound, upper_bound)
# Keep learning information for further use.
self.learner_info_ = {}
# We count the number of non-zero weights, including the implicit voters.
# TODO: Verify how we define non-zero weights here, could be if the weight is near 1/2n.
n_nonzero_weights = np.sum(np.asarray(weights) > 1e-12)
n_nonzero_weights += np.sum(np.asarray(weights) < 1.0 / len(self.estimators_) - 1e-12)
self.learner_info_.update(n_nonzero_weights=n_nonzero_weights)
return weights
def _solve_qp(self, objective_matrix, objective_vector, equality_matrix, equality_vector, lower_bound, upper_bound):
try:
qp = ConvexProgram()
qp.quadratic_func, qp.linear_func = objective_matrix, objective_vector
qp.add_equality_constraints(equality_matrix, equality_vector)
qp.add_lower_bound(lower_bound)
qp.add_upper_bound(upper_bound)
return qp.solve()
except Exception:
# logger.warning("Error while solving the quadratic program.")
raise
class RegularizedBinaryMinCqClassifier(MinCqClassifier):
"""MinCq, version published in [1] and [2], where the regularization comes from the enforced quasi-uniformity
of the posterior distributino on the symmetric hypothesis space. This version only works with {-1, 1} labels.
[1] From PAC-Bayes Bounds to Quadratic Programs for Majority Votes (Laviolette et al., 2011)
[2] Risk Bounds for the Majority Vote: From a PAC-Bayesian Analysis to a Learning Algorithm (Germain et al., 2015)
"""
def fit(self, X, y):
# We first fit and learn the weights.
super().fit(X, y)
# Validations
if isinstance(y, np.ma.MaskedArray):
assert len(self.classes_[np.where(np.logical_not(self.classes_.mask))]) == 2, "RegularizedBinaryMinCqClassifier: only supports binary classification."
else:
assert len(self.classes_), "RegularizedBinaryMinCqClassifier: only supports binary classification."
# Then we "reverse" the negative weights and their associated voter's output.
for i, weight in enumerate(self.weights):
if weight < 0:
# logger.debug("Reversing decision of a binary voter")
self.weights[i] *= -1
self.estimators_[i].reverse_decision()
return self
def _solve(self, X, y):
if isinstance(y, np.ma.MaskedArray):
y = np.ma.MaskedArray(self.le_.transform(y), y.mask)
else:
y = self.le_.transform(y)
classification_matrix = self._binary_classification_matrix(X)
n_examples, n_voters = np.shape(classification_matrix)
if self.zeta == 0:
np.transpose(classification_matrix)
ftf = np.dot(np.transpose(classification_matrix),classification_matrix)
else:
I = np.eye(n_examples)
L = build_laplacian(X, n_neighbors=self.n_neighbors)
ftf = classification_matrix.T.dot(I + (self.zeta / n_examples) * L).dot(classification_matrix)
# We use {-1, 1} labels.
binary_labels = np.ma.copy(y)
binary_labels[np.ma.where(y == 0)] = -1
# Objective function.
ftf_mean = np.mean(ftf, axis=1)
objective_matrix = 2.0 / n_examples * ftf
objective_vector = -1.0 / n_examples * ftf_mean.T
# Equality constraint: first moment of the margin fixed to mu, only using labeled examples.
if isinstance(y, np.ma.MaskedArray):
labeled = np.where(np.logical_not(y.mask))[0]
binary_labels = binary_labels[labeled]
else:
labeled = range(len(y))
yf = binary_labels.T.dot(classification_matrix[labeled])
yf_mean = np.mean(yf)
equality_matrix = 2.0 / len(labeled) * yf
equality_vector = self.mu + 1.0 / len(labeled) * yf_mean
# Lower and upper bounds (quasi-uniformity constraints)
lower_bound = 0.0
upper_bound = 1.0 / n_voters
try:
weights = self._solve_qp(objective_matrix, objective_vector, equality_matrix, equality_vector, lower_bound, upper_bound)
except ValueError as e:
if "domain error" in e.args:
weights = np.ones(len(self.estimators_))
# Keep learning information for further use.
self.learner_info_ = {}
# We count the number of non-zero weights, including the implicit voters.
# TODO: Verify how we define non-zero weights here, could be if the weight is near 1/2n.
n_nonzero_weights = np.sum(np.asarray(weights) > 1e-12)
n_nonzero_weights += np.sum(np.asarray(weights) < 1.0 / len(self.estimators_) - 1e-12)
self.learner_info_.update(n_nonzero_weights=n_nonzero_weights)
# Conversion of the weights of the n first voters to weights on the implicit 2n voters.
# See Section 7.1 of [2] for an explanation.
return np.array([2 * q - 1.0 / len(self.estimators_) for q in weights])
# def evaluate_metrics(self, X, y, metrics_list=None, functions_list=None):
# if metrics_list is None:
# metrics_list = [zero_one_loss]
#
# if functions_list is None:
# functions_list = []
#
# # Transductive setting: we only predict the X for labeled y
# if isinstance(y, np.ma.MaskedArray):
# labeled = np.where(np.logical_not(y.mask))[0]
# X = np.array(X[labeled])
# y = np.array(y[labeled])
#
# # Predict, evaluate metrics.
# predictions = self.predict(X)
# metrics_results = {metric.__name__: metric(y, predictions) for metric in metrics_list}
#
# # TODO: Repair in the case of non-{-1, 1} labels.
# assert set(y) == {-1, 1}
# classification_matrix = self._binary_classification_matrix(X)
#
# for function in functions_list:
# metrics_results[function.__name__] = function(classification_matrix, y, self.weights)
#
# metrics_dataframe = ResultsDataFrame([metrics_results])
# return metrics_dataframe
def build_laplacian(X, n_neighbors=None):
clf = SpectralEmbedding(n_neighbors=n_neighbors)
clf.fit(X)
w = clf.affinity_matrix_
laplacian = graph_laplacian(w, normed=True)
return laplacian
class MinCQGraalpy(RegularizedBinaryMinCqClassifier, BaseMonoviewClassifier):
def __init__(self, random_state=None, mu=0.01, self_complemented=True, n_stumps_per_attribute=1, **kwargs):
super(MinCQGraalpy, self).__init__(mu=mu,
estimators_generator=StumpsClassifiersGenerator(n_stumps_per_attribute=n_stumps_per_attribute, self_complemented=self_complemented),
)
self.param_names = ["mu", "n_stumps_per_attribute", "random_state"]
self.distribs = [CustomUniform(loc=0.05, state=2.0, multiplier="e-"),
[n_stumps_per_attribute], [random_state]]
self.n_stumps_per_attribute = n_stumps_per_attribute
self.classed_params = []
self.weird_strings = {}
self.random_state = random_state
if "nbCores" not in kwargs:
self.nbCores = 1
else:
self.nbCores = kwargs["nbCores"]
def canProbas(self):
"""Used to know if the classifier can return label probabilities"""
return True
def set_params(self, **params):
self.mu = params["mu"]
self.random_state = params["random_state"]
self.n_stumps_per_attribute = params["n_stumps_per_attribute"]
return self
def get_params(self, deep=True):
return {"random_state":self.random_state, "mu":self.mu, "n_stumps_per_attribute":self.n_stumps_per_attribute}
def getInterpret(self, directory, y_test):
interpret_string = ""
# interpret_string += "Train C_bound value : "+str(self.cbound_train)
# y_rework = np.copy(y_test)
# y_rework[np.where(y_rework==0)] = -1
# interpret_string += "\n Test c_bound value : "+str(self.majority_vote.cbound_value(self.x_test, y_rework))
return interpret_string
def get_name_for_fusion(self):
return "MCG"
def formatCmdArgs(args):
"""Used to format kwargs for the parsed args"""
kwargsDict = {"mu":args.MCG_mu,
"n_stumps_per_attribute":args.MCG_stumps}
return kwargsDict
def paramsToSet(nIter, randomState):
"""Used for weighted linear early fusion to generate random search sets"""
paramsSet = []
for _ in range(nIter):
paramsSet.append({})
return paramsSet
# if __name__ == '__main__':
# # Example usage.
# from sklearn.datasets import load_iris
# from sklearn.cross_validation import train_test_split
# from graalpy.utils.majority_vote import StumpsClassifiersGenerator
#
# # Load data, change {0, 1, 2} labels to {-1, 1}
# iris = load_iris()
# iris.target[np.where(iris.target == 0)] = -1
# iris.target[np.where(iris.target == 2)] = 1
# x_train, x_test, y_train, y_test = train_test_split(iris.data, iris.target, random_state=42)
#
# # Fit MinCq
# clf = RegularizedBinaryMinCqClassifier(estimators_generator=StumpsClassifiersGenerator())
# clf.fit(x_train, y_train)
#
# # Compare the best score of individual classifiers versus the score of the learned majority vote.
# print("Best training risk of individual voters: {:.4f}".format(1 - max([e.score(x_train, y_train) for e in clf.estimators_])))
# print("Training risk of the majority vote outputted by MinCq: {:.4f}".format(1 - clf.score(x_train, y_train)))
# print()
# print("Best testing risk of individual voters: {:.4f}".format(1 - max([e.score(x_test, y_test) for e in clf.estimators_])))
# print("Testing risk of the majority vote outputted by MinCq: {:.4f}".format(1 - clf.score(x_test, y_test)))