-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathutils.py
188 lines (146 loc) · 6.32 KB
/
utils.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
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.svm import LinearSVC, SVC
from sklearn.preprocessing import StandardScaler
import os
# generate data
def generate_dataset(X, Y, G, kernel, clean_ratio, attacker_split, test_split, random_seed):
"""
kernel: we use sklearn SVM to pre-process the data. kernel is linear or rbf
clean_ratio: fraction of data left after removing hard examples
attacker_split: fraction of data given to the attacker
test_split: train/test split ratio
"""
sc = StandardScaler()
X_scaled = sc.fit_transform(X)
n_clean = int(clean_ratio*Y.shape[0])
n_attacker = int(n_clean*attacker_split)
if kernel == 'linear':
model = LinearSVC(max_iter=5000)
else:
model = SVC(gamma='auto')
# Train the model, iterating on the data in batches of 32 samples
model.fit(X_scaled, Y)
loss = np.absolute(np.subtract(model.predict(X_scaled), Y))
index = np.argsort(loss, axis=0)
x_clean = X_scaled[index[:n_clean],:]
y_clean = Y[index[:n_clean]]
g_clean = G[index[:n_clean]]
x_noise = X_scaled[index[n_clean:],:]
y_noise = Y[index[n_clean:]]
g_noise = G[index[n_clean:]]
attacker_index = np.random.choice(y_clean.shape[0], n_attacker, replace = False)
x_attacker = x_clean[attacker_index]
y_attacker = y_clean[attacker_index]
g_attacker = g_clean[attacker_index]
x_clean = np.delete(x_clean, attacker_index, 0)
y_clean = np.delete(y_clean, attacker_index, 0)
g_clean = np.delete(g_clean, attacker_index, 0)
x_train,x_test,y_train,y_test,g_train,g_test = train_test_split(x_clean, y_clean, g_clean, test_size = test_split, random_state=random_seed)
data = {
'x_train': x_train,
'y_train': y_train,
'g_train': g_train,
'x_test': x_test,
'y_test': y_test,
'g_test': g_test,
'x_attacker': x_attacker,
'y_attacker': y_attacker,
'g_attacker': g_attacker,
'x_noise': x_noise,
'y_noise': y_noise,
'g_noise': g_noise
}
return data
# accuracy measure
def accuracy(y_true, y_pred):
if len(y_true)==0:
return 0
return 1 - sum([abs(y_true[i] - y_pred[i]) for i in range(len(y_true))])/len(y_true)
# fairness measure
def EO(s, y_pred, y_true):
y_pred_0_0 = [(1-y_pred[i]) for i in range(len(s)) if y_true[i] == 0 and s[i] == 0]
y_pred_1_0 = [(1-y_pred[i]) for i in range(len(s)) if y_true[i] == 0 and s[i] == 1]
y_pred_0_1 = [y_pred[i] for i in range(len(s)) if y_true[i] == 1 and s[i] == 0]
y_pred_1_1 = [y_pred[i] for i in range(len(s)) if y_true[i] == 1 and s[i] == 1]
loss_0_0 = np.mean(y_pred_0_0)
loss_1_0 = np.mean(y_pred_1_0)
loss_0_1 = np.mean(y_pred_0_1)
loss_1_1 = np.mean(y_pred_1_1)
return abs(loss_0_0 - loss_1_0), abs(loss_0_1 - loss_1_1)
# functions
def cross_entropy(y, t, tol=1e-12):
pred = np.clip(y, tol, 1-tol)
pred_n = np.clip(1-pred, tol, 1-tol)
return - (np.sum(np.multiply(t, np.log(pred)),axis=1) + np.sum(np.multiply(1-t, np.log(pred_n)),axis=1))
# evaluate penalized loss
def eval_loss(A, b, x, y, g, x_reg, y_reg, g_reg, L, num_points):
s = 1 - (np.dot(x, A) + b) * (2*y-1)
loss = s*(s>=0)
loss2 = (1-(2*y-1)*(np.dot(x, A) + b))/2
loss2_reg = (1-(2*y_reg-1)*(np.dot(x_reg, A) + b))/2
idx00 = np.logical_and(g_reg.flatten()==0, y_reg.flatten()==0)
idx01 = np.logical_and(g_reg.flatten()==0, y_reg.flatten()==1)
idx10 = np.logical_and(g_reg.flatten()==1, y_reg.flatten()==0)
idx11 = np.logical_and(g_reg.flatten()==1, y_reg.flatten()==1)
_s00 = np.sum(loss2_reg[idx00])
_s01 = np.sum(loss2_reg[idx01])
_s10 = np.sum(loss2_reg[idx10])
_s11 = np.sum(loss2_reg[idx11])
_c00 = np.sum(idx00)
_c01 = np.sum(idx01)
_c10 = np.sum(idx10)
_c11 = np.sum(idx11)
for i in range(len(loss)):
s00 = _s00
s01 = _s01
s10 = _s10
s11 = _s11
c00 = _c00
c01 = _c01
c10 = _c10
c11 = _c11
if g[i][0]==0 and y[i][0]==0:
s00 = _s00 + num_points*loss2[i][0]
c00 = _c00 + num_points
elif g[i][0]==0 and y[i][0]==1:
s01 = _s01 + num_points*loss2[i][0]
c01 = _c01 + num_points
elif g[i][0]==1 and y[i][0]==0:
s10 = _s10 + num_points*loss2[i][0]
c10 = _c10 + num_points
elif g[i][0]==1 and y[i][0]==1:
s11 = _s11 +num_points*loss2[i][0]
c11 = _c11 + num_points
reg = L*np.abs(s00/c00-s10/c10) + L*np.abs(s01/c01-s11/c11)
loss[i][0] += reg
return loss
# gradient
# note: write 2y-1 to convert to +/-1 prediction to use Hinge loss
def gradient(A, b, x_loss, y_loss, x_reg, y_reg, g_reg, L, r, n, num_points):
s = 1 - (np.dot(x_loss, A) + b) * (2*y_loss-1)
dfA_loss = np.sum((-x_loss*(2*y_loss-1))*(s>=0), axis=0).reshape((-1,1))
dfb_loss = np.sum(-(2*y_loss-1)*(s>=0))
idx00 = np.logical_and(g_reg.flatten()==0, y_reg.flatten()==0)
idx01 = np.logical_and(g_reg.flatten()==0, y_reg.flatten()==1)
idx10 = np.logical_and(g_reg.flatten()==1, y_reg.flatten()==0)
idx11 = np.logical_and(g_reg.flatten()==1, y_reg.flatten()==1)
loss2_reg = (1-(2*y_reg-1)*(np.dot(x_reg, A) + b))/2
s00 = np.sum(loss2_reg[idx00])
s01 = np.sum(loss2_reg[idx01])
s10 = np.sum(loss2_reg[idx10])
s11 = np.sum(loss2_reg[idx11])
c00 = np.sum(idx00)
c01 = np.sum(idx01)
c10 = np.sum(idx10)
c11 = np.sum(idx11)
dfA_reg_0 = (L*np.sign(s00/c00-s10/c10)*(np.mean(-(2*y_reg[idx00]-1)*x_reg[idx00]/2, axis=0) - np.mean(-(2*y_reg[idx10]-1)*x_reg[idx10]/2, axis=0))).reshape((-1,1))
dfb_reg_0 = L*np.sign(s00/c00-s10/c10)*(-sum(2*y_reg[idx00]-1)/2/len(idx00) - -sum(2*y_reg[idx10]-1)/2/len(idx10))
dfA_reg_1 = (L*np.sign(s01/c01-s11/c11)*(np.mean(-(2*y_reg[idx01]-1)*x_reg[idx01]/2, axis=0) - np.mean(-(2*y_reg[idx11]-1)*x_reg[idx11]/2, axis=0))).reshape((-1,1))
dfb_reg_1 = L*np.sign(s01/c01-s11/c11)*(-sum(2*y_reg[idx01]-1)/2/len(idx01) - -sum(2*y_reg[idx11]-1)/2/len(idx11))
dA_reg = (2*A).reshape((-1,1))
db_reg = 2*b
return {
'dA': ((dfA_loss+num_points*dfA_reg_0+num_points*dfA_reg_1)/n+r*dA_reg).reshape((-1,1)) ,
'db': (dfb_loss+num_points*dfb_reg_0+num_points*dfb_reg_1)/n#+r*db_reg
}