-
Notifications
You must be signed in to change notification settings - Fork 16
/
utils_cifar.py
177 lines (137 loc) · 5.86 KB
/
utils_cifar.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
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
from tqdm import tqdm, trange
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
import pandas as pd
import scipy
import time
from observations import cifar10
from sklearn.calibration import calibration_curve
try:
import cPickle as pickle
except Exception as e:
import pickle
(x_train, y_train), (x_test, y_test) = cifar10(
'/vol/biomedic/users/np716/data/cifar_10/')
x_train = x_train.transpose(0, 2, 3, 1)
x_test = x_test.transpose(0, 2, 3, 1)
x_train_first = x_train[y_train < 5]
y_train_first = y_train[y_train < 5]
x_test_first = x_test[y_test < 5]
y_test_first = y_test[y_test < 5]
x_test_outlier = x_test[y_test >= 5]
# helper for rotation
def rotate(img, angle):
img = scipy.ndimage.rotate(img.reshape((28, 28)), angle, reshape=False)
return img.reshape((-1))
max_ent = np.sum(-1 * (np.ones(5) / 5.) * np.log((np.ones(5) / 5.)), -1)
def get_pred_df(data, session, ops, mode):
cols = ['prob', 'Prediction', 'sample_idx', 'unit']
df = pd.DataFrame(columns=cols)
probs = get_probs(data, session, ops, mode)
for sample_idx in range(probs.shape[1]): # per data sample
for class_idx in range(10): # per class ...
data = zip(
probs[:, sample_idx, class_idx],
[class_idx] * len(probs),
[sample_idx] * len(probs),
list(range(len(probs)))
)
new_df = pd.DataFrame(columns=cols, data=data)
df = pd.concat([df, new_df])
return df
def get_probs(data_inp, session, ops, mode, is_eval=True):
if mode == 'ensemble':
probs = np.stack([
session.run(prob, feed_dict={
ops['x']: data_inp, ops['is_eval']: is_eval})
for prob in ops['probs']])
elif mode == 'map' or mode == 'mle':
probs = session.run(ops['probs'], feed_dict={
ops['x']: data_inp, ops['is_eval']: is_eval})
probs = probs[np.newaxis, :]
else:
probs = np.zeros((100, len(data_inp), 5)) # ensemble, data, classes
batch_size = 1000
for b in range(len(data_inp) // batch_size):
start = b * batch_size
end = start + batch_size
for i in range(100):
probs[i, start:end] += session.run(
ops['probs'], feed_dict={
ops['x']: data_inp[start:end], ops['is_eval']: is_eval})
end = (len(data_inp) // batch_size) * batch_size
if end < len(data_inp):
start = end
for i in range(100):
probs[i, start:] += session.run(
ops['probs'], feed_dict={
ops['x']: data_inp[start:], ops['is_eval']: is_eval})
return probs
def build_adv_examples(images, labels, eps, session, ops, mode):
feed_dict = {ops['x']: images, ops['y']: labels, ops['adv_eps']: eps,
ops['is_eval']: True}
if mode == 'ensemble':
adv_data = np.mean(np.stack([
session.run(ad, feed_dict=feed_dict)
for ad in ops['adv_data']]), 0)
elif mode == 'map' or mode == 'mle':
adv_data = session.run(ops['adv_data'], feed_dict=feed_dict)
else:
adv_data = session.run(ops['adv_data'], feed_dict=feed_dict) / 100
for i in range(99):
adv_data += session.run(ops['adv_data'], feed_dict=feed_dict) / 100
return adv_data
def calc_entropy(probs): # shape = [sample, classes]
return np.sum(-1 * probs * np.log(np.maximum(probs, 1e-5)), -1)
def calc_ent_auc(ent):
hist, bin_edges = np.histogram(ent, density=True,
bins=np.arange(0, max_ent, max_ent / 500))
c_hist = np.cumsum(hist * np.diff(bin_edges))
return np.sum(np.diff(bin_edges) * c_hist)
def build_result_dict(session, ops, mode):
result_dict = {}
# calc test acc:
probs = get_probs(x_test_first, session, ops, mode)
mean_probs = probs.mean(0)
test_acc = np.mean(np.argmax(mean_probs, -1) == y_test_first)
test_entropy = calc_entropy(mean_probs)
test_ent_auc = calc_ent_auc(test_entropy)
test_cal_pos, test_cal_bins = calibration_curve(
np.ones(len(mean_probs)),
mean_probs[np.arange(len(mean_probs)), y_test_first],
normalize=False, n_bins=50)
result_dict['mean_probs'] = mean_probs
result_dict['test_acc'] = test_acc
result_dict['test_ent_auc'] = test_ent_auc
result_dict['test_entropy'] = test_entropy
result_dict['test_cal_pos'] = test_cal_pos
result_dict['test_cal_bins'] = test_cal_bins
# not mnist entropy
probs = get_probs(x_test_outlier, session, ops, mode)
mean_probs = probs.mean(0)
outlier_entropy = calc_entropy(mean_probs)
outlier_ent_auc = calc_ent_auc(outlier_entropy)
result_dict['outlier_mean_probs'] = mean_probs
result_dict['outlier_entropy'] = outlier_entropy
result_dict['outlier_ent_auc'] = outlier_ent_auc
# build adv examples and test performance
adv_df = pd.DataFrame(columns=['eps', 'acc', 'ent', 'ent_auc'])
result_dict['adv_examples'] = {}
for eps in np.linspace(0., 0.4, num=9):
adv_data = build_adv_examples(x_test_first[:100], y_test_first[:100],
eps, session, ops, mode)
result_dict['adv_examples'][eps] = adv_data
adv_data = np.pad(adv_data, ((0, 0), (4, 4), (4, 4), (0, 0)),
'constant')
adv_probs = get_probs(adv_data, session, ops, mode)
mean_adv_probs = adv_probs.mean(0)
adv_acc = np.mean(np.argmax(mean_adv_probs, -1) == y_test_first[:100])
adv_entropy = calc_entropy(mean_adv_probs)
adv_ent_auc = calc_ent_auc(adv_entropy)
adv_df.loc[len(adv_df)] = [eps, adv_acc, adv_entropy.mean(),
adv_ent_auc]
result_dict['adv_df'] = adv_df
return result_dict