-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathBayesianDataLogisticRegression.py
180 lines (154 loc) · 8.2 KB
/
BayesianDataLogisticRegression.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
import random
import numpy as np
from netcal.metrics import ECE
from scipy.stats import beta, uniform, norm
from sklearn.metrics import accuracy_score, log_loss
from sklearn.linear_model import LogisticRegression
from DatasetGenerator import SyntheticGaussianGenerator, SyntheticCauchyGenerator, SyntheticBetaGenerator
import warnings
warnings.filterwarnings('once')
random.seed(0)
def log_pdf_norm(mu, sigma, sample):
return norm.logpdf(sample, mu, sigma)
def sigmoid(m, b, X):
return 1 / (1 + np.exp(-m*X-b))
def optimal_sigmoid_preds(mu0, mu1, sig, X, base_generator_type = 'gaussian'):
if base_generator_type == 'gaussian':
W = (mu1 - mu0) / (sig * sig)
b = (mu0 * mu0 - mu1 * mu1) / (2 * sig * sig)
else:
print('A7A')
if base_generator_type == 'cauchy':
generator = SyntheticCauchyGenerator(p = 0.5)
generator.set_cauchy_parameters(mu0, mu1, sig, sig)
elif base_generator_type == 'beta':
generator = SyntheticBetaGenerator(p = 0.5)
generator.set_beta_parameters(mu0, mu1, sig, sig)
XX, y = generator.generate(int(1e2))
XX = XX.reshape(-1, 1)
tmp_clf = LogisticRegression(random_state=0, solver='lbfgs', penalty = 'none', C = 0)
tmp_clf.fit(XX, y)
W = tmp_clf.coef_[0]; b = tmp_clf.intercept_[0]
return sigmoid(W, b, X)
class uniform_prior():
def __init__(self, l0min, l0max, l1min, l1max, lsmin=1, lsmax=1):
self.l0min = l0min; self.l0max = l0max #negative class
self.l1min = l1min; self.l1max = l1max #positive class
self.lsmin = lsmin; self.lsmax = lsmax #standard deviation
def sample(self, num_samples):
np.random.seed(42)
mu0s = uniform.rvs(self.l0min, self.l0max - self.l0min, size = num_samples)
mu1s = uniform.rvs(self.l1min, self.l1max - self.l1min, size = num_samples)
sigmas = uniform.rvs(self.lsmin, self.lsmax - self.lsmin, size = num_samples)
return mu0s, mu1s, sigmas
def draw_priors(self, priorN = 1e5):
#prior of mean of negative distribution
mu0P = uniform.rvs(self.l0min, self.l0max - self.l0min, size = int(priorN))
#prior of mean of positive distribution
mu1P = uniform.rvs(self.l1min, self.l1max - self.l1min, size = int(priorN))
#prior of stdev distribution
sigmaP = uniform.rvs(self.lsmin, self.lsmax - self.lsmin, size = int(priorN))
return mu0P, mu1P, sigmaP
def log_pdf(self, a, b, samples):
return uniform.logpdf(samples, a, b - a)
def log_pdf_prior(self, mu0s, mu1s):
return np.sum(self.log_pdf(self.l0min, self.l0max, mu0s)) + np.sum(self.log_pdf(self.l1min, self.l1max, mu1s))
def log_likelihood(self, mu0P, mu1P, sigP):
log_pdf0 = np.sum(self.log_pdf(self.l0min, self.l0max - self.l0min, self.mu0P))
log_pdf1 = np.sum(self.log_pdf(self.l1min, self.l1max - self.l1min, self.mu1P))
if self.lsmax > self.lsmin:
log_pdfs = np.sum(self.log_pdf(self.lsmin, self.lsmax - self.lsmin, self.sigmaP))
else:
log_pdfs = 0
return log_pdf0 + log_pdf1 + log_pdfs
class beta_prior():
def __init__(self, a0, b0, a1, b1, shift=1, lsmin=1, lsmax=1):
self.a0 = a0; self.b0 = b0
self.a1 = a1; self.b1 = b1
self.shift = shift
self.lsmin = lsmin; self.lsmax = lsmax
def sample(self, num_samples):
np.random.seed(42)
mu0s = beta.rvs(self.a0, self.b0, size = num_samples)
mu1s = beta.rvs(self.a1, self.b1, size = num_samples) + self.shift
sigmas = uniform.rvs(self.lsmin, self.lsmax - self.lsmin, size = num_samples)
return mu0s, mu1s, sigmas
def draw_priors(self, priorN = 1e5):
#prior of mean of negative distribution
mu0P = beta.rvs(self.a0, self.b0, size = int(priorN))
#prior of mean of positive distribution
mu1P = beta.rvs(self.a1, self.b1, size = int(priorN)) + self.shift
#prior of distribution stddev is always 1
sigmaP = np.ones(int(priorN))
return mu0P, mu1P, sigmaP
def log_pdf(self, a, b, samples):
return beta.logpdf(samples, a, b)
def log_pdf_prior(self, mu0s, mu1s):
return np.sum(self.log_pdf(self.a0, self.b0, mu0s)) + np.sum(self.log_pdf(self.a1, self.b1, (mu1s - self.shift)))
def log_likelihood(self):
log_pdf0 = np.sum(self.log_pdf(self.a0, self.b0, self.mu0P))
log_pdf1 = np.sum(self.log_pdf(self.a1, self.b1, (self.mu1P-self.shift)))
if self.lsmax > self.lsmin:
log_pdfs = np.sum(self.log_pdf(self.lsmin, self.lsmax - self.lsmin, self.sigmaP))
else:
log_pdfs = 0
return log_pdf0 + log_pdf1 + log_pdfs
class BayesianDataLogisticRegression(object):
def __init__(self, prior_obj, prior_type='uniform', prior_samples=int(1e5), base_generator_type = 'gaussian'):
self.prior_type = prior_type
self.prior = prior_obj
self.prior_samples = prior_samples
self.base_generator_type = base_generator_type
def calc_posts(self, S0, S1):
log_likelihood = np.sum(log_pdf_norm([[m] for m in self.mu0P], [[s] for s in self.sigmaP], S0), axis = 1)
log_likelihood += np.sum(log_pdf_norm([[m] for m in self.mu1P], [[s] for s in self.sigmaP], S1), axis = 1)
self.posts = log_likelihood + self.prior.log_pdf_prior(self.mu0P, self.mu1P)
self.posts = np.exp(self.posts - max(self.posts))
self.posts = np.divide(self.posts, self.posts.sum())
def fit(self, X_train, y_train):
self.mu0P, self.mu1P, self.sigmaP = self.prior.draw_priors()
self.calc_posts(S0 = X_train[np.where(y_train == 0)], S1 = X_train[np.where(y_train == 1)])
self.mean_mu0_coeff = np.sum(np.multiply(self.posts, self.mu0P));
self.mean_mu1_coeff = np.sum(np.multiply(self.posts, self.mu1P));
self.mean_sigma_coeff = np.sum(np.multiply(self.posts, self.sigmaP))
_, self.mu0_top, self.mu1_top, self.sigma_top = zip(*sorted(zip(self.posts, self.mu0P, self.mu1P, self.sigmaP)))
self.mu0_top = self.mu0_top[-1]; self.mu1_top = self.mu1_top[-1]; self.sigma_top = self.sigma_top[-1]
def predict(self, X, threshold = 0.5):
return self.predict_proba(X) >= threshold
def predict_logloss(self, X, y):
preds_probs = self.predict_proba(X)
#print(y[:10], preds_probs[:10], log_loss(y[:10], preds_probs[:10], labels = [0, 1]))
return log_loss(y, preds_probs, labels = [0, 1])
def predict_accuracy(self, X, y, threshold = 0.5):
return accuracy_score(y, self.predict(X, threshold = threshold))
def predict_ece(self, X, y, bins = 10):
ece = ECE(bins)
calibrated_score = ece.measure(self.predict_proba(X), y)
return calibrated_score
def predict_ece_logloss(self, X, y, bins = 10):
preds_probs = self.predict_proba(X)
ece = ECE(bins)
calibrated_score = ece.measure(preds_probs, y)
return calibrated_score, log_loss(y, preds_probs, labels = [0, 1])
class BayesianDataLogisticRegressionMeanPreds(BayesianDataLogisticRegression):
def predict_proba(self, X):
if X.ndim == 1:
X = X.reshape(-1, 1)
preds_probs = np.zeros((len(X), 1), dtype=np.float64)
for mu0, mu1, sig, post in zip(self.mu0P, self.mu1P, self.sigmaP, self.posts):
prior_output = optimal_sigmoid_preds(mu0, mu1, sig, X, base_generator_type = self.base_generator_type)
preds_probs = np.add(preds_probs, np.multiply(post, prior_output))
preds_probs = np.squeeze(preds_probs)
return preds_probs
class BayesianDataLogisticRegressionMeanCoeffs(BayesianDataLogisticRegression):
def predict_proba(self, X):
if X.ndim == 1:
X = X.reshape(-1, 1)
preds_probs = optimal_sigmoid_preds(self.mean_mu0_coeff, self.mean_mu1_coeff, self.mean_sigma_coeff, X, base_generator_type = self.base_generator_type)
return np.squeeze(preds_probs)
class BayesianDataLogisticRegressionMAP(BayesianDataLogisticRegression):
def predict_proba(self, X):
if X.ndim == 1:
X = X.reshape(-1, 1)
preds_probs = optimal_sigmoid_preds(self.mu0_top, self.mu1_top, self.sigma_top, X, base_generator_type = self.base_generator_type)
return np.squeeze(preds_probs)