-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
170 lines (135 loc) · 7.72 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
import numpy as np
import torch
from copy import deepcopy
from sklearn.metrics import classification_report, accuracy_score, precision_recall_fscore_support, roc_auc_score, \
average_precision_score
from fairlearn.metrics import (
MetricFrame, equalized_odds_difference, equalized_odds_ratio,
selection_rate, demographic_parity_difference, demographic_parity_ratio,
false_positive_rate, false_negative_rate,
false_positive_rate_difference, false_negative_rate_difference,
equalized_odds_difference)
def prepare_data(dataloader, model, device):
x_, y_, a_ = [], [], []
for batch_idx, (images, labels) in enumerate(dataloader):
gender = labels[:, 20]
a_.append(gender)
hair = labels[:, 9]
y_.append(hair)
X = model.get_features(images.to(device)).detach().cpu()
x_.append(X)
return torch.cat(x_), torch.cat(y_), torch.cat(a_)
def eo_constraint(p, y, a):
fpr = torch.abs(torch.sum(p * (1 - y) * a) / (torch.sum(a) + 1e-5) - torch.sum(p * (1 - y) * (1 - a)) / (
torch.sum(1 - a) + 1e-5))
fnr = torch.abs(torch.sum((1 - p) * y * a) / (torch.sum(a) + 1e-5) - torch.sum((1 - p) * y * (1 - a)) / (
torch.sum(1 - a) + 1e-5))
return fpr, fnr
def di_constraint(p, a):
di = -1 * torch.min((torch.sum(a * p) / torch.sum(a)) / (torch.sum((1 - a) * p) / torch.sum((1 - a))),
(torch.sum((1 - a) * p) / torch.sum((1 - a))) / (torch.sum(a * p) / torch.sum(a)))
return di
def dp_constraint(p, a):
dp = torch.abs((torch.sum(a * p) / torch.sum(a)) - (torch.sum((1 - a) * p) / torch.sum((1 - a))))
return dp
def ae_constraint(criterion, log_softmax, y, a):
loss_p = criterion(log_softmax[a == 1], y[a == 1])
loss_n = criterion(log_softmax[a == 0], y[a == 0])
return torch.abs(loss_p - loss_n)
def mmf_constraint(criterion, log_softmax, y, a):
# loss_p = criterion(log_softmax[a == 1], y[a == 1])
# loss_n = criterion(log_softmax[a == 0], y[a == 0])
# return torch.max(loss_p, loss_n)
y_p_a = y + a
y_m_a = y - a
if len(y[y_p_a == 2]) > 0:
loss_1 = criterion(log_softmax[y_p_a == 2], y[y_p_a == 2]) # (1, 1)
else:
loss_1 = torch.tensor(0.0).cuda()
if len(y[y_p_a == 0]) > 0:
loss_2 = criterion(log_softmax[y_p_a == 0], y[y_p_a == 0]) # (0, 0)
else:
loss_2 = torch.tensor(0.0).cuda()
if len(y[y_m_a == 1]) > 0:
loss_3 = criterion(log_softmax[y_m_a == 1], y[y_m_a == 1]) # (1, 0)
else:
loss_3 = torch.tensor(0.0).cuda()
if len(y[y_m_a == -1]) > 0:
loss_4 = criterion(log_softmax[y_m_a == -1], y[y_m_a == -1]) # (0, 1)
else:
loss_4 = torch.tensor(0.0).cuda()
return torch.max(torch.max(loss_1, loss_2), torch.max(loss_3, loss_4))
def disparity_impact_difference(y, pred, sensitive_features):
return demographic_parity_difference(y, pred, sensitive_features=sensitive_features)
def disparity_impact_ratio(y, pred, sensitive_features):
return demographic_parity_ratio(y, pred, sensitive_features=sensitive_features)
def accuracy_equality_difference(y, pred, sensitive_features):
misclassification_rate_p = sum(y[sensitive_features == 1] != pred[sensitive_features == 1]) / sum(
sensitive_features == 1)
misclassification_rate_n = sum(y[sensitive_features == 0] != pred[sensitive_features == 0]) / sum(
sensitive_features == 0)
return abs(misclassification_rate_p - misclassification_rate_n)
def accuracy_equality_ratio(y, pred, sensitive_features):
misclassification_rate_p = sum(y[sensitive_features == 1] != pred[sensitive_features == 1]) / sum(
sensitive_features == 1)
misclassification_rate_n = sum(y[sensitive_features == 0] != pred[sensitive_features == 0]) / sum(
sensitive_features == 0)
return min(misclassification_rate_p / (misclassification_rate_n + 1e-6),
misclassification_rate_n / (misclassification_rate_p + 1e-6))
def max_min_fairness(y, pred, sensitive_features):
# classification_rate_p = sum(y[sensitive_features == 1] == pred[sensitive_features == 1]) / sum(
# sensitive_features == 1)
# classification_rate_n = sum(y[sensitive_features == 0] == pred[sensitive_features == 0]) / sum(
# sensitive_features == 0)
# return min(classification_rate_p, classification_rate_n)
y_p_a = y + sensitive_features
y_m_a = y - sensitive_features
classification_rate_1 = sum(y[y_p_a == 2] == pred[y_p_a == 2]) / sum(y_p_a == 2)
classification_rate_2 = sum(y[y_p_a == 0] == pred[y_p_a == 0]) / sum(y_p_a == 0)
classification_rate_3 = sum(y[y_m_a == 1] == pred[y_m_a == 1]) / sum(y_m_a == 1)
classification_rate_4 = sum(y[y_m_a == -1] == pred[y_m_a == -1]) / sum(y_m_a == -1)
return min(min(classification_rate_1, classification_rate_2), min(classification_rate_3, classification_rate_4))
def print_fpr_fnr_sensitive_features(y_true, y_pred, x_control, sensitive_attrs):
for s in sensitive_attrs:
s_attr_vals = x_control[s]
print("|| s || FPR. || FNR. ||")
for s_val in sorted(list(set(s_attr_vals))):
y_true_local = y_true[s_attr_vals == s_val]
y_pred_local = y_pred[s_attr_vals == s_val]
# acc = float(sum(y_true_local==y_pred_local)) / len(y_true_local)
# fp = sum(np.logical_and(y_true_local == 0.0, y_pred_local == +1.0)) # something which is -ve but is misclassified as +ve
# fn = sum(np.logical_and(y_true_local == +1.0, y_pred_local == 0.0)) # something which is +ve but is misclassified as -ve
# tp = sum(np.logical_and(y_true_local == +1.0, y_pred_local == +1.0)) # something which is +ve AND is correctly classified as +ve
# tn = sum(np.logical_and(y_true_local == 0.0, y_pred_local == 0.0)) # something which is -ve AND is correctly classified as -ve
# all_neg = sum(y_true_local == 0.0)
# all_pos = sum(y_true_local == +1.0)
# fpr = float(fp) / float(fp + tn)
# fnr = float(fn) / float(fn + tp)
# tpr = float(tp) / float(tp + fn)
# tnr = float(tn) / float(tn + fp)
fpr = false_positive_rate(y_true_local, y_pred_local)
fnr = false_negative_rate(y_true_local, y_pred_local)
if isinstance(s_val, float): # print the int value of the sensitive attr val
s_val = int(s_val)
print("|| %s || %0.2f || %0.2f ||" % (s_val, fpr, fnr))
def print_clf_stats(out_train, out_finetune, out_test, pred_train, pred_finetune, pred_test, y_train, a_train,
y_finetune, a_finetune, y_test, a_test, sensitive_attrs):
train_acc, finetune_acc, test_acc = accuracy_score(y_train, pred_train), accuracy_score(y_finetune,
pred_finetune), accuracy_score(
y_test, pred_test)
train_auc, finetune_auc, test_auc = roc_auc_score(y_train, out_train), roc_auc_score(y_finetune,
out_finetune), roc_auc_score(
y_test, out_test)
for s_attr in sensitive_attrs:
print("*** Train ***")
print("Accuracy: %0.3f, AUC: %0.3f" % (train_acc, train_auc))
print_fpr_fnr_sensitive_features(y_train, pred_train, a_train, sensitive_attrs)
print("\n")
print("*** Finetune ***")
print("Accuracy: %0.3f, AUC: %0.3f" % (finetune_acc, finetune_auc))
print_fpr_fnr_sensitive_features(y_finetune, pred_finetune, a_finetune, sensitive_attrs)
print("\n")
print("*** Test ***")
print("Accuracy: %0.3f, AUC: %0.3f" % (test_acc, test_auc))
print_fpr_fnr_sensitive_features(y_test, pred_test, a_test, sensitive_attrs)
print("\n")