-
Notifications
You must be signed in to change notification settings - Fork 0
/
test_real_datasets.py
194 lines (163 loc) · 8.12 KB
/
test_real_datasets.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
181
182
183
184
185
186
187
188
189
190
191
192
193
194
from __future__ import division
import numpy as np
from cwc.data_wrappers import reject
from cwc.evaluation.rgp import RGP
from cwc.evaluation.confidence_intervals import ConfidenceIntervals
from sklearn import datasets
from sklearn.cross_validation import StratifiedKFold
from sklearn.mixture import GMM
from sklearn import svm
from sklearn import tree
import matplotlib.pyplot as plt
def load_dataset(name="iris"):
if name == "iris":
dataset = datasets.load_iris()
if name == "digits":
dataset = datasets.load_digits()
return dataset.data, dataset.target
def remove_reject_class(data, labels, reject_class=0):
true_data = data[labels != reject_class, :]
true_labels = labels[labels != reject_class]
remaining_labels = np.unique(true_labels)
counts = [np.sum(true_labels == l) for l in remaining_labels]
positive_label = remaining_labels[np.argmin(counts)]
true_labels = 1*(true_labels == positive_label)
return true_data, true_labels
def separate_reject_class(data, labels, reject_class=0):
data_reject = data[labels == reject_class, :]
true_data, true_labels = remove_reject_class(data, labels,
reject_class=reject_class)
return data_reject, true_data, true_labels
def separate_sets(x, y, test_fold, validation_fold, test_folds):
test_data = x[test_folds == test_fold, :]
test_labels = y[test_folds == test_fold]
validation_data = x[test_folds == validation_fold, :]
validation_labels = y[test_folds == validation_fold]
training_folds = np.logical_and(test_folds != test_fold,
test_folds != validation_fold)
training_data = x[training_folds, :]
training_labels = y[training_folds]
return [training_data, training_labels, validation_data, validation_labels,
test_data, test_labels]
def train_reject_model(x, r):
"""Train a classifier of training points
Returns a classifier that predicts high probability values for training
points and low probability values for reject points.
"""
# model_rej = svm.SVC(probability=True)
model_rej = tree.DecisionTreeClassifier(max_depth=4)
xr = np.vstack((x, r))
y = np.hstack((np.ones(np.alen(x)), np.zeros(np.alen(r)))).T
model_rej.fit(xr, y)
return model_rej
def train_classifier_model(x, y):
# model_clas = svm.SVC(probability=True)
model_clas = tree.DecisionTreeClassifier(max_depth=2)
model_clas = model_clas.fit(x, y)
return model_clas
def evaluate_test(test_data, test_labels, model_rej=None,
model_clas=None, step1_threshold=0.5,
step2_threshold=0.5, reject_class=0):
test_reject, true_test_data, true_test_labels = \
separate_reject_class(training_data, training_labels,
reject_class=reject_class)
if model_rej is not None:
step1_reject_scores = model_rej.predict_proba(test_reject)[:,1]
step1_test_scores = model_rej.predict_proba(true_test_data)[:,1]
else:
step1_reject_scores = np.ones(np.alen(test_reject))
step1_test_scores = np.ones(np.alen(true_test_data))
step2_test_scores = model_clas.predict_proba(true_test_data)[:,1]
n_accepted_rejects = np.sum(step1_reject_scores >= step1_threshold)
accepted_data = step1_test_scores >= step1_threshold
n_accepted_data = np.sum(accepted_data)
n_correct_positives = np.sum(np.logical_and(np.logical_and(
accepted_data, true_test_labels == 1),
step2_test_scores >= step2_threshold))
n_correct_negatives = np.sum(np.logical_and(np.logical_and(
accepted_data, true_test_labels == 0),
step2_test_scores < step2_threshold))
precision_1 = n_accepted_data / (n_accepted_data + n_accepted_rejects)
n = np.alen(true_test_data)
recall_1 = n_accepted_data / n
accuracy_2 = (n_correct_positives + n_correct_negatives) / n
return [recall_1, precision_1, accuracy_2]
# print("recall_1 = {}, precision_1 = {}, accuracy_2 = {},"
# " accuracy'_2 = {}".format(recall_1, precision_1, accuracy_2,
# accuracy_2 * precision_1))
if __name__ == "__main__":
np.random.seed(1)
x, y = load_dataset("iris")
n_folds = 5
skf = StratifiedKFold(y, n_folds=n_folds)
reject_class = 0
# accuracies of the second classifier when the first classifier is used
recalls_1 = np.zeros(n_folds)
precisions_1 = np.zeros(n_folds)
accuracies_2 = np.zeros(n_folds)
# accuracies of the second classifier without the first classifier
pr_1 = np.zeros(n_folds)
acc_2 = np.zeros(n_folds)
for test_fold in np.arange(n_folds):
validation_fold = test_fold + 1
if test_fold == (n_folds - 1):
validation_fold = 0
[training_data, training_labels, validation_data, validation_labels,
test_data, test_labels] = separate_sets(x, y, test_fold,
validation_fold, skf.test_folds)
'''
Training and validation sets preparation. The class that will be
treated as reject is removed from the data. The generated data
will be used to simulate knowledge about the reject class.
The reject class is used for testing.
'''
true_training_data, true_training_labels = \
remove_reject_class(training_data, training_labels,
reject_class=reject_class)
r = reject.create_reject_data(true_training_data,
proportion=1, method='uniform_hsphere',
pca=True, pca_variance=0.9,
pca_components=0, hshape_cov=0,
hshape_prop_in=0.99, hshape_multiplier=2)
true_validation_data, true_validation_labels = \
remove_reject_class(validation_data, validation_labels,
reject_class=reject_class)
v_r = reject.create_reject_data(true_validation_data,
proportion=1, method='uniform_hsphere',
pca=True, pca_variance=0.9,
pca_components=0, hshape_cov=0,
hshape_prop_in=0.99, hshape_multiplier=2)
# Classifier of reject data
model_rej = train_reject_model(true_training_data, r)
# Classifier of training data
model_clas = train_classifier_model(true_training_data,
true_training_labels)
step1_reject_scores = model_rej.predict_proba(v_r)[:,1]
step1_validation_scores = model_rej.predict_proba(true_validation_data)[:,1]
step2_validation_scores = model_clas.predict_proba(true_validation_data)[:,1]
'''
TODO: find optimal classification threshold for second classifier
'''
rgp = RGP(step1_reject_scores, step1_validation_scores,
step2_validation_scores, true_validation_labels,
step2_threshold=0.5)
print("Area = {}".format(rgp.calculate_area()))
fig = plt.figure('RGP')
fig.clf()
rgp.plot(fig)
step1_threshold = rgp.get_optimal_step1_threshold()
print("Optimal threshold for the first classifier = {}".format(
step1_threshold))
[recalls_1[test_fold], precisions_1[test_fold], accuracies_2[test_fold]] = \
evaluate_test(
test_data, test_labels, model_rej=model_rej,
model_clas=model_clas, step1_threshold=step1_threshold,
step2_threshold=0.5, reject_class=reject_class)
[re_1, pr_1[test_fold], acc_2[test_fold]] = evaluate_test(
test_data, test_labels, model_rej=None,
model_clas=model_clas, step1_threshold=0.5,
step2_threshold=0.5, reject_class=reject_class)
values = np.append((accuracies_2*precisions_1).reshape(-1, 1),
(acc_2*pr_1).reshape(-1, 1), 1)
intervals = ConfidenceIntervals(values, ['Protected', 'Original'], n_samples=10, alpha=0.05)
intervals.plot()