-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathAPPENDIX-D-HELPER.py
291 lines (209 loc) · 10.8 KB
/
APPENDIX-D-HELPER.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
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
import os
import time
import argparse
import numpy as np
from typing import Sequence
from dataset import fetch_data, DataTemplate
from dataset2 import fetch_data2, DataTemplate2
from eval import Evaluator
from model import LogisticRegression, NNLastLayerIF, MLPClassifier
from fair_fn import grad_ferm, grad_dp, loss_ferm, loss_dp
from utils import fix_seed, save2csv
import json
from robust_fn import grad_robust, calc_robust_acc
from robust_fn_nn import grad_robust_nn, calc_robust_acc_nn
import pickle
import random
import copy
import matplotlib.pyplot as plt
def parse_args():
parser = argparse.ArgumentParser(description='Influence Fairness')
parser.add_argument('--dataset', type=str, default="adult", help="name of the dataset")
parser.add_argument('--metric', type=str, default="eop", help="eop or dp")
parser.add_argument('--seed', type=float, default=42, help="random seed")
parser.add_argument('--save_model', type=str, default="n", help="y/n")
parser.add_argument('--type', type=str, default="util", help="util/fair/robust")
parser.add_argument('--strategy', type=str, default="dec", help="inc/dec/random")
parser.add_argument('--points_to_delete', type=int, default=500, help="points to delete")
parser.add_argument('--random_seed', type=int, default=42, help="seed for random strategy")
parser.add_argument('--only_pre', type=str, default="n", help="y/n")
parser.add_argument('--model_type', type=str, default="logreg", help="logreg/nn")
args = parser.parse_args()
return args
def pre_main(args):
tik = time.time()
if args.seed is not None:
fix_seed(args.seed)
with open('data/' + args.dataset + '/meta.json', 'r+') as f:
json_data = json.load(f)
json_data['train_path'] = './data/' + args.dataset + '/train.csv'
f.seek(0)
json.dump(json_data, f, indent=4)
f.truncate()
""" initialization"""
data: DataTemplate = fetch_data(args.dataset)
model = LogisticRegression(l2_reg=data.l2_reg)
if args.model_type == 'nn':
model = NNLastLayerIF(input_dim=data.dim, base_model_cls=MLPClassifier, l2_reg=1e-4)
val_evaluator, test_evaluator = Evaluator(data.s_val, "val"), Evaluator(data.s_test, "test")
""" vanilla training """
model.fit(data.x_train, data.y_train)
if args.dataset == "toy" and args.save_model == "y":
pickle.dump(model.model, open("toy/model.pkl", "wb"))
if args.metric == "eop":
ori_fair_loss_val = loss_ferm(model.log_loss, data.x_val, data.y_val, data.s_val)
elif args.metric == "dp":
pred_val, _ = model.pred(data.x_val)
ori_fair_loss_val = loss_dp(data.x_val, data.s_val, pred_val)
else:
raise ValueError
ori_util_loss_val = model.log_loss(data.x_val, data.y_val)
""" compute the influence and save data """
pred_train, _ = model.pred(data.x_train)
train_total_grad, train_indiv_grad = model.grad(data.x_train, data.y_train)
util_loss_total_grad, acc_loss_indiv_grad = model.grad(data.x_val, data.y_val)
if args.metric == "eop":
fair_loss_total_grad = grad_ferm(model.grad, data.x_val, data.y_val, data.s_val)
elif args.metric == "dp":
fair_loss_total_grad = grad_dp(model.grad_pred, data.x_val, data.s_val)
else:
raise ValueError
if args.model_type != 'nn':
robust_loss_total_grad = grad_robust(model, data.x_val, data.y_val)
else:
robust_loss_total_grad = grad_robust_nn(model, data.x_val, data.y_val)
hess = model.hess(data.x_train)
util_grad_hvp = model.get_inv_hvp(hess, util_loss_total_grad)
fair_grad_hvp = model.get_inv_hvp(hess, fair_loss_total_grad)
robust_grad_hvp = model.get_inv_hvp(hess, robust_loss_total_grad)
util_pred_infl = train_indiv_grad.dot(util_grad_hvp)
fair_pred_infl = train_indiv_grad.dot(fair_grad_hvp)
robust_pred_infl = train_indiv_grad.dot(robust_grad_hvp)
np.save('explainer/data/binaries/util_infl.npy', util_pred_infl)
np.save('explainer/data/binaries/fair_infl.npy', fair_pred_infl)
np.save('explainer/data/binaries/robust_infl.npy', robust_pred_infl)
_, pred_label_val = model.pred(data.x_val)
_, pred_label_test = model.pred(data.x_test)
val_res = val_evaluator(data.y_val, pred_label_val)
test_res = test_evaluator(data.y_test, pred_label_test)
if args.model_type != 'nn':
val_rob_acc = calc_robust_acc(model, data.x_val, data.y_val, 'val', 'pre')
test_rob_acc = calc_robust_acc(model, data.x_test, data.y_test, 'test', 'pre')
#######################################################
print("Validation set robustness accuracy -> ", val_rob_acc)
print("Test set robustness accuracy -> ", test_rob_acc)
#######################################################
val_res.update({'robust_acc': val_rob_acc})
test_res.update({'robust_acc': test_rob_acc})
else:
val_rob_acc = calc_robust_acc_nn(model, data.x_val, data.y_val, 'val', 'pre')
test_rob_acc = calc_robust_acc_nn(model, data.x_test, data.y_test, 'test', 'pre')
#######################################################
print("Validation set robustness accuracy -> ", val_rob_acc)
print("Test set robustness accuracy -> ", test_rob_acc)
#######################################################
val_res.update({'robust_acc': val_rob_acc})
test_res.update({'robust_acc': test_rob_acc})
np.save('trn.npy', np.append(data.x_train, data.y_train.reshape((-1,1)), 1))
return val_res, test_res
def post_main(args):
tik = time.time()
if args.seed is not None:
fix_seed(args.seed)
with open('data/' + args.dataset + '/meta.json', 'r+') as f:
json_data = json.load(f)
json_data['train_path'] = './data/' + args.dataset + '/train.csv'
f.seek(0)
json.dump(json_data, f, indent=4)
f.truncate()
""" initialization"""
data: DataTemplate2 = fetch_data2(args.dataset)
model = LogisticRegression(l2_reg=data.l2_reg)
if args.model_type == 'nn':
model = NNLastLayerIF(input_dim=data.dim, base_model_cls=MLPClassifier, l2_reg=1e-4)
val_evaluator, test_evaluator = Evaluator(data.s_val, "val"), Evaluator(data.s_test, "test")
""" vanilla training """
model.fit(data.x_train, data.y_train)
if args.dataset == "toy" and args.save_model == "y":
pickle.dump(model.model, open("toy/model.pkl", "wb"))
_, pred_label_val = model.pred(data.x_val)
_, pred_label_test = model.pred(data.x_test)
val_res = val_evaluator(data.y_val, pred_label_val)
test_res = test_evaluator(data.y_test, pred_label_test)
if args.model_type != 'nn':
val_rob_acc = calc_robust_acc(model, data.x_val, data.y_val, 'val', 'post')
test_rob_acc = calc_robust_acc(model, data.x_test, data.y_test, 'test', 'post')
#######################################################
print("Validation set robustness accuracy -> ", val_rob_acc)
print("Test set robustness accuracy -> ", test_rob_acc)
#######################################################
val_res.update({'robust_acc': val_rob_acc})
test_res.update({'robust_acc': test_rob_acc})
else:
val_rob_acc = calc_robust_acc_nn(model, data.x_val, data.y_val, 'val', 'post')
test_rob_acc = calc_robust_acc_nn(model, data.x_test, data.y_test, 'test', 'post')
#######################################################
print("Validation set robustness accuracy -> ", val_rob_acc)
print("Test set robustness accuracy -> ", test_rob_acc)
#######################################################
val_res.update({'robust_acc': val_rob_acc})
test_res.update({'robust_acc': test_rob_acc})
return val_res, test_res
def deletion_process(args):
X_org = np.load('trn.npy')
num_to_del = int(args.points_to_delete)
num_features = X_org.shape[1]
I = np.load('explainer/data/binaries/'+args.type+'_infl.npy')
if args.strategy == 'inc':
indices_to_delete = I.argsort()[::-1][-num_to_del:][::-1].tolist() #INC by deleting bad points
elif args.strategy == 'dec':
indices_to_delete = I.argsort()[-num_to_del:][::-1].tolist() #DEC by deleting good points
elif args.strategy == 'random':
random.seed(int(args.random_seed)) #42, 1, 35, 999, 5454
indices_to_delete = random.sample(range(0,len(I)), int(0.1*len(I)))[:num_to_del] #RANDOM
X_new = []
for i in range(X_org.shape[0]):
if i in indices_to_delete:
continue
X_new.append(X_org[i])
X_new = np.array(X_new)
X = X_new
np.save('2trn.npy', X)
return indices_to_delete
if __name__ == "__main__":
args = parse_args()
pre_val_res, pre_test_res = pre_main(args) #Run pre code
influence_test_results = {'fair': [pre_test_res[args.metric]]} #Initialize results dict (infl)
random_test_results = {42: copy.deepcopy(influence_test_results), #Initialize results dict (random)
1: copy.deepcopy(influence_test_results),
35: copy.deepcopy(influence_test_results),
999: copy.deepcopy(influence_test_results),
5454: copy.deepcopy(influence_test_results),
}
x_ax = np.linspace(0, int(0.05*len(np.load('explainer/data/binaries/'+args.type+'_infl.npy'))), num=11) #Set up x axis range according to data
if args.only_pre == 'y': #If only pre needed, exit
exit(0)
args.strategy, args.type = 'inc', 'fair' #EOP, increase fairness
for i,num_i in enumerate(x_ax[1:]):
args.points_to_delete = int(num_i)
fair_del_idx = deletion_process(args)
post_val_res, post_test_res = post_main(args)
influence_test_results['fair'].append(post_test_res[args.metric])
args.strategy = 'random' #Random baseline experiments
for seedval in [42,1,35,999,5454]:
args.random_seed = seedval
for i,num_i in enumerate(x_ax[1:]):
args.points_to_delete = int(num_i)
#print('-->>', args.points_to_delete, args.random_seed)
deletion_process(args)
post_val_res, post_test_res = post_main(args)
random_test_results[args.random_seed]['fair'].append(post_test_res[args.metric])
with open('appendix-d-outputs/influence-'+args.dataset+'-'+args.model_type+'.pkl', "wb") as output_file:
pickle.dump(influence_test_results, output_file)
with open('appendix-d-outputs/random-'+args.dataset+'-'+args.model_type+'.pkl', "wb") as output_file:
pickle.dump(random_test_results, output_file)
# Save influence values
fair_infl_vals = np.load('explainer/data/binaries/fair_infl.npy')
deleted_fair_influences = {'fair':fair_infl_vals[fair_del_idx]}
with open('appendix-d-outputs/fair-'+args.dataset+'-'+args.model_type+'.pkl', "wb") as output_file:
pickle.dump(deleted_fair_influences, output_file)