forked from marcotcr/lime-experiments
-
Notifications
You must be signed in to change notification settings - Fork 0
/
generate_data_for_compare_classifiers.py
180 lines (167 loc) · 8.9 KB
/
generate_data_for_compare_classifiers.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 sys
import copy
sys.path.append('..')
import time
import numpy as np
import scipy as sp
import sklearn
import xgboost
import xgboost.sklearn
import explainers
from load_datasets import *
from sklearn.metrics import accuracy_score
from sklearn import ensemble, cross_validation
import pickle
import parzen_windows
import argparse
def get_random_indices(labels, class_, probability):
nonzero = (labels == class_).nonzero()[0]
if nonzero.shape[0] == 0 or probability == 0:
return []
return np.random.choice(nonzero, int(probability * len(nonzero)) , replace=False)
def add_corrupt_feature(feature_name, clean_train, clean_test, dirty_train,
train_labels, test_labels, class_probs_dirty, class_probs_clean, fake_prefix='FAKE'):
"""clean_train, clean_test, dirty_train will be corrupted"""
for class_ in set(train_labels):
indices = get_random_indices(train_labels, class_, class_probs_clean[class_])
for i in indices:
clean_train[i] += ' %s%s%s' % (fake_prefix, feature_name, fake_prefix)
indices = get_random_indices(train_labels, class_, class_probs_dirty[class_])
for i in indices:
dirty_train[i] += ' %s%s%s' % (fake_prefix, feature_name, fake_prefix)
indices = get_random_indices(test_labels, class_, class_probs_clean[class_])
for i in indices:
clean_test[i] += ' %s%s%s' % (fake_prefix, feature_name, fake_prefix)
def corrupt_dataset(independent_features, train_data, train_labels, test_data, test_labels):
# independent_features: list [([.3, .8],[.5,.5], 3), ([.1, .1],[0, 0], 1)
# ...]. Each element in list is a tuple (l,l2, n) where l a list
# representing the probability of seeing the feature in each class in the
# dirty train data, l2 is a list representing the probability of seeing the
# feature in each class the clean test data and n is the number of features
# with this distribution to add.
# returns (clean_train, dirty_train, clean_test)
dirty_train = copy.deepcopy(train_data)
clean_train = copy.deepcopy(train_data)
clean_test = copy.deepcopy(test_data)
idx = 0
for probs, probs2, n in independent_features:
for i in range(n):
add_corrupt_feature('%d' % idx, clean_train, clean_test, dirty_train, train_labels, test_labels, probs, probs2)
idx += 1
return clean_train, dirty_train, clean_test
def main():
parser = argparse.ArgumentParser(description='Evaluate some explanations')
parser.add_argument('--dataset', '-d', type=str, required=True,help='dataset name')
parser.add_argument('--output_folder', '-o', type=str, required=True, help='output folder')
parser.add_argument('--num_features', '-k', type=int, required=True, help='num features')
parser.add_argument('--num_rounds', '-r', type=int, required=True, help='num rounds')
parser.add_argument('--start_id', '-i', type=int, default=0,required=False, help='output start id')
args = parser.parse_args()
dataset = args.dataset
train_data, train_labels, test_data, test_labels, class_names = LoadDataset(dataset)
rho = 25
kernel = lambda d: np.sqrt(np.exp(-(d**2) / rho ** 2))
local = explainers.GeneralizedLocalExplainer(kernel, explainers.data_labels_distances_mapping_text, num_samples=15000, return_mean=True, verbose=False, return_mapped=True)
# Found through cross validation
sigmas = {'multi_polarity_electronics': {'neighbors': 0.75, 'svm': 10.0, 'tree': 0.5,
'logreg': 0.5, 'random_forest': 0.5, 'embforest': 0.75},
'multi_polarity_kitchen': {'neighbors': 1.0, 'svm': 6.0, 'tree': 0.75,
'logreg': 0.25, 'random_forest': 6.0, 'embforest': 1.0},
'multi_polarity_dvd': {'neighbors': 0.5, 'svm': 0.75, 'tree': 8.0, 'logreg':
0.75, 'random_forest': 0.5, 'embforest': 5.0}, 'multi_polarity_books':
{'neighbors': 0.5, 'svm': 7.0, 'tree': 2.0, 'logreg': 1.0, 'random_forest':
1.0, 'embforest': 3.0}}
parzen1 = parzen_windows.ParzenWindowClassifier()
parzen1.sigma = sigmas[dataset]['random_forest']
parzen2 = parzen_windows.ParzenWindowClassifier()
parzen2.sigma = sigmas[dataset]['random_forest']
random = explainers.RandomExplainer()
for Z in range(args.num_rounds):
exps1 = {}
exps2 = {}
explainer_names = ['lime', 'parzen', 'random', 'greedy', 'mutual']
for expl in explainer_names:
exps1[expl] = []
exps2[expl] = []
print 'Round', Z
sys.stdout.flush()
fake_features_z = [([.1, .2], [.1,.1], 10)]#, ([.2, .1], [.1,.1], 10)]
clean_train, dirty_train, clean_test = corrupt_dataset(fake_features_z, train_data, train_labels, test_data, test_labels)
vectorizer = CountVectorizer(lowercase=False, binary=True)
dirty_train_vectors = vectorizer.fit_transform(dirty_train)
clean_train_vectors = vectorizer.transform(clean_train)
test_vectors = vectorizer.transform(clean_test)
terms = np.array(list(vectorizer.vocabulary_.keys()))
indices = np.array(list(vectorizer.vocabulary_.values()))
inverse_vocabulary = terms[np.argsort(indices)]
tokenizer = vectorizer.build_tokenizer()
c1 = ensemble.RandomForestClassifier(n_estimators=30, max_depth=5)
c2 = ensemble.RandomForestClassifier(n_estimators=30, max_depth=5)
untrustworthy = [i for i, x in enumerate(inverse_vocabulary) if x.startswith('FAKE')]
train_idx, test_idx = tuple(cross_validation.ShuffleSplit(dirty_train_vectors.shape[0], 1, 0.2))[0]
train_acc1 = train_acc2 = test_acc1 = test_acc2 = 0
print 'Trying to find trees:'
sys.stdout.flush()
iteration = 0
found_tree = True
while np.abs(train_acc1 - train_acc2) > 0.001 or np.abs(test_acc1 - test_acc2) < 0.05:
iteration += 1
c1.fit(dirty_train_vectors[train_idx], train_labels[train_idx])
c2.fit(dirty_train_vectors[train_idx], train_labels[train_idx])
train_acc1 = accuracy_score(train_labels[test_idx], c1.predict(dirty_train_vectors[test_idx]))
train_acc2 = accuracy_score(train_labels[test_idx], c2.predict(dirty_train_vectors[test_idx]))
test_acc1 = accuracy_score(test_labels, c1.predict(test_vectors))
test_acc2 = accuracy_score(test_labels, c2.predict(test_vectors))
if iteration == 3000:
found_tree = False
break
if not found_tree:
print 'skipping iteration', Z
continue
print 'done'
print 'Train acc1:', train_acc1, 'Train acc2:', train_acc2
print 'Test acc1:', test_acc1, 'Test acc2:', test_acc2
sys.stdout.flush()
predictions = c1.predict(dirty_train_vectors)
predictions2 = c2.predict(dirty_train_vectors)
predict_probas = c1.predict_proba(dirty_train_vectors)[:,1]
predict_probas2 = c2.predict_proba(dirty_train_vectors)[:,1]
cv_preds1 = cross_validation.cross_val_predict(c1, dirty_train_vectors[train_idx], train_labels[train_idx], cv=5)
cv_preds2 = cross_validation.cross_val_predict(c2, dirty_train_vectors[train_idx], train_labels[train_idx], cv=5)
parzen1.fit(dirty_train_vectors[train_idx], cv_preds1)
parzen2.fit(dirty_train_vectors[train_idx], cv_preds2)
pp = []
pp2 = []
true_labels = []
iteration = 0
for i in test_idx:
if iteration % 50 == 0:
print iteration
sys.stdout.flush()
iteration += 1
pp.append(predict_probas[i])
pp2.append(predict_probas2[i])
true_labels.append(train_labels[i])
exp, mean = local.explain_instance(dirty_train_vectors[i], 1, c1.predict_proba, args.num_features)
exps1['lime'].append((exp, mean))
exp = parzen1.explain_instance(dirty_train_vectors[i], 1, c1.predict_proba, args.num_features, None)
mean = parzen1.predict_proba(dirty_train_vectors[i])[1]
exps1['parzen'].append((exp, mean))
exp = random.explain_instance(dirty_train_vectors[i], 1, None, args.num_features, None)
exps1['random'].append(exp)
exp = explainers.explain_greedy_martens(dirty_train_vectors[i], predictions[i], c1.predict_proba, args.num_features)
exps1['greedy'].append(exp)
# Classifier 2
exp, mean = local.explain_instance(dirty_train_vectors[i], 1, c2.predict_proba, args.num_features)
exps2['lime'].append((exp, mean))
exp = parzen2.explain_instance(dirty_train_vectors[i], 1, c2.predict_proba, args.num_features, None)
mean = parzen2.predict_proba(dirty_train_vectors[i])[1]
exps2['parzen'].append((exp, mean))
exp = random.explain_instance(dirty_train_vectors[i], 1, None, args.num_features, None)
exps2['random'].append(exp)
exp = explainers.explain_greedy_martens(dirty_train_vectors[i], predictions2[i], c2.predict_proba, args.num_features)
exps2['greedy'].append(exp)
out = {'true_labels' : true_labels, 'untrustworthy' : untrustworthy, 'train_acc1' : train_acc1, 'train_acc2' : train_acc2, 'test_acc1' : test_acc1, 'test_acc2' : test_acc2, 'exps1' : exps1, 'exps2': exps2, 'predict_probas1': pp, 'predict_probas2': pp2}
pickle.dump(out, open(os.path.join(args.output_folder, 'comparing_%s_%s_%d.pickle' % (dataset, args.num_features, Z + args.start_id)), 'w'))
if __name__ == "__main__":
main()