From 8f4b1568bca1602e1489e4e8c7602fb5b5f2b9b2 Mon Sep 17 00:00:00 2001
From: Alexandre Gramfort <alexandre.gramfort@inria.fr>
Date: Mon, 15 Nov 2010 18:44:31 +0100
Subject: [PATCH] ENH : adding support for predict_log_proba in Naive Bayes
---
scikits/learn/naive_bayes.py | 38 +++++++++++++++++--------
scikits/learn/tests/test_naive_bayes.py | 6 +++-
2 files changed, 31 insertions(+), 13 deletions(-)
diff --git a/scikits/learn/naive_bayes.py b/scikits/learn/naive_bayes.py
index a4635827f8..8f3a4a53e9 100644
--- a/scikits/learn/naive_bayes.py
+++ b/scikits/learn/naive_bayes.py
@@ -7,6 +7,7 @@ import numpy as np
from .base import BaseEstimator, ClassifierMixin
+
class GNB(BaseEstimator, ClassifierMixin):
"""
Gaussian Naive Bayes (GNB)
@@ -56,6 +57,7 @@ class GNB(BaseEstimator, ClassifierMixin):
--------
"""
+
def __init__(self):
pass
@@ -65,8 +67,8 @@ class GNB(BaseEstimator, ClassifierMixin):
proba_y = []
unique_y = np.unique(y)
for yi in unique_y:
- theta.append(np.mean(X[y==yi,:], 0))
- sigma.append(np.var(X[y==yi,:], 0))
+ theta.append(np.mean(X[y==yi, :], 0))
+ sigma.append(np.var(X[y==yi, :], 0))
proba_y.append(np.float(np.sum(y==yi)) / np.size(y))
self.theta = np.array(theta)
self.sigma = np.array(sigma)
@@ -74,23 +76,35 @@ class GNB(BaseEstimator, ClassifierMixin):
self.unique_y = unique_y
return self
-
def predict(self, X):
- y_pred = self.unique_y[np.argmax(self.predict_proba(X),1)]
+ y_pred = self.unique_y[np.argmax(self.predict_proba(X), 1)]
return y_pred
-
- def predict_proba(self, X):
+ def _joint_log_likelihood(self, X):
joint_log_likelihood = []
for i in range(np.size(self.unique_y)):
jointi = np.log(self.proba_y[i])
- n_ij = - 0.5 * np.sum(np.log(np.pi*self.sigma[i,:]))
- n_ij -= 0.5 * np.sum( ((X - self.theta[i,:])**2) /\
- (self.sigma[i,:]),1)
- joint_log_likelihood.append(jointi+n_ij)
+ n_ij = - 0.5 * np.sum(np.log(np.pi*self.sigma[i, :]))
+ n_ij -= 0.5 * np.sum(((X - self.theta[i, :])**2) / \
+ (self.sigma[i, :]), 1)
+ joint_log_likelihood.append(jointi + n_ij)
joint_log_likelihood = np.array(joint_log_likelihood).T
+ return joint_log_likelihood
+
+ def predict_proba(self, X):
+ joint_log_likelihood = self._joint_log_likelihood(X)
proba = np.exp(joint_log_likelihood)
- proba = proba / np.sum(proba,1)[:,np.newaxis]
+ proba = proba / np.sum(proba, 1)[:, np.newaxis]
return proba
-
+ def predict_log_proba(self, X):
+ log_proba = self._joint_log_likelihood(X)
+ # Compute a sum of logs without underflow. Equivalent to:
+ # log_proba -= np.log(np.sum(np.exp(log_proba), axis=1))[:, np.newaxis]
+ B = np.max(log_proba, axis=1)[:, np.newaxis]
+ logaB = log_proba - B
+ sup = logaB > -np.inf
+ aB = np.zeros_like(logaB)
+ aB[sup] = np.exp(logaB[sup])
+ log_proba -= np.log(np.sum(aB, axis=1))[:, np.newaxis] + B
+ return log_proba
diff --git a/scikits/learn/tests/test_naive_bayes.py b/scikits/learn/tests/test_naive_bayes.py
index 9bbf7a4519..0f4f74437e 100644
--- a/scikits/learn/tests/test_naive_bayes.py
+++ b/scikits/learn/tests/test_naive_bayes.py
@@ -1,5 +1,5 @@
import numpy as np
-from numpy.testing import assert_array_equal
+from numpy.testing import assert_array_equal, assert_array_almost_equal
from .. import naive_bayes
@@ -19,3 +19,7 @@ def test_gnb():
y_pred = clf.fit(X, y).predict(X)
assert_array_equal(y_pred, y)
+ y_pred_proba = clf.predict_proba(X)
+ y_pred_log_proba = clf.predict_log_proba(X)
+ assert_array_almost_equal(np.log(y_pred_proba), y_pred_log_proba, 8)
+
--
GitLab