-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathrun_apps.py
227 lines (197 loc) · 11.3 KB
/
run_apps.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
"""
High-dimensional applications in Section 6 of our paper: https://arxiv.org/pdf/2010.10969.pdf.
Due to data privacy reasons, we omit Section 6.1: Clinical Action Prediction with the MIMIC-III database.
Default behavior when running this file::
All experiments are fully run from scratch, i.e. full inference is performed.
The config files from the `repro/` folder are used.
As BNN inference is approximate, results will slightly deviate from those in the paper, especially for VI methods.
There are three optional command-line arguments:
--debug :: run experiments in debug mode (only a couple iterations of inference)
--pretrained :: use our pre-trained posterior samples and AOCP parameters
--custom_config <filename.yaml> :: use your own config file --> <filename.yaml> if specified, else, `config.yaml` in the root directory
Note: The pretrained samples in `repro/` are not identical to the samples used to produce the actual tables in our paper.
They are separate runs of the same experiments to prove reproducibility of our results.
"""
import numpy as np
import torch
import logging
import argparse
from bnn import *
from bnn.utils import *
from data.dataloader import *
def recid_evaluate_high_risk(bnn):
""" Recidivism prediction task: evaluate high-risk fraction of both groups """
samples, _ = bnn.all_bayes_samples[-1]
preds = (bnn.predict(samples, bnn.X_train, return_probs=True).mean(dim=0) >= 0.5).float()
eval_data = torch.stack((bnn.X_train_race, preds)).t()
aa_highrisk = len(eval_data[(eval_data[:, 0] == 1.0) & (eval_data[:, 1] == 1.0)])
aa_total = len(eval_data[(eval_data[:, 0] == 1.0)])
nonaa_highrisk = len(eval_data[(eval_data[:, 0] == 0.0) & (eval_data[:, 1] == 1.0)])
nonaa_total = len(eval_data[(eval_data[:, 0] == 0.0)])
return aa_highrisk, aa_total, nonaa_highrisk, nonaa_total
def recidivism_task():
""" Section 6.2: Recidivism Prediction Task """
# With race as an explicit feature, baseline.
bnn = BNNSVGDBinaryClassifier(uid="recid", configfile=args.custom_config or "repro/recid.yaml")
bnn.load(**compas_dataset(csv_xfilename="data/compas_X.csv", csv_yfilename="data/compas_Y.csv"))
logging.info(f'[{bnn.uid}] Dataset <{bnn.dataset_name}>: {(100 * sum(bnn.Y_train) // bnn.N_train):.2f}% of training points are positive.')
if args.debug:
bnn.debug_mode()
if args.pretrained:
bnn.load_bayes_samples('repro/recid_svgd1.pt', 'svgd_baseline')
else:
bnn.infer()
logging.info(f'[{bnn.uid}] Baseline evaluation, with race as feature:')
acc, f1 = eval_accuracy_and_f1_score(bnn, is_binary=True, X_eval=bnn.X_train, Y_eval=bnn.Y_train)
logging.info(f'[{bnn.uid}] Train accuracy: {acc:.3f}.')
logging.info(f'[{bnn.uid}] Train F1 score: {f1:.3f}.')
aa_highrisk, aa_total, nonaa_highrisk, nonaa_total = recid_evaluate_high_risk(bnn)
logging.info(f'[{bnn.uid}] African American High-Risk Fraction: {(aa_highrisk / aa_total):.3f}.')
logging.info(f'[{bnn.uid}] Non-African American High-Risk Fraction: {(nonaa_highrisk / nonaa_total):.3f}.')
# With race as an explicit feature, OC-BNN.
cdomain = ()
for i in range(9):
if i == 1:
cdomain += (0., 1.)
else:
mm = bnn.X_train[:,i].mean().item()
ss = bnn.X_train[:,i].std().item()
cdomain += (mm - 0.2 * ss, mm + 0.2 * ss)
bnn.add_probabilistic_constraint(constrained_domain=cdomain, prob_func=lambda X: X[:,1], prior_type="gaussian_aocp")
if args.pretrained:
bnn.load_gaussian_aocp_parameters('repro/recid_aocp.pt')
else:
bnn.learn_gaussian_aocp()
bnn.update_config(use_ocbnn=True)
if args.pretrained:
bnn.load_bayes_samples('repro/recid_svgd2.pt', 'svgd_ocbnn')
else:
bnn.infer()
logging.info(f'[{bnn.uid}] OC-BNN evaluation, with race as feature:')
acc, f1 = eval_accuracy_and_f1_score(bnn, is_binary=True, X_eval=bnn.X_train, Y_eval=bnn.Y_train)
logging.info(f'[{bnn.uid}] Train accuracy: {acc:.3f}.')
logging.info(f'[{bnn.uid}] Train F1 score: {f1:.3f}.')
aa_highrisk, aa_total, nonaa_highrisk, nonaa_total = recid_evaluate_high_risk(bnn)
logging.info(f'[{bnn.uid}] African American High-Risk Fraction: {(aa_highrisk / aa_total):.3f}.')
logging.info(f'[{bnn.uid}] Non-African American High-Risk Fraction: {(nonaa_highrisk / nonaa_total):.3f}.')
# Without race as feature, baseline.
bnn = BNNSVGDBinaryClassifier(uid="recid-no-race", configfile=args.custom_config or "repro/recid-no-race.yaml")
bnn.load(**compas_dataset(csv_xfilename="data/compas_X.csv", csv_yfilename="data/compas_Y.csv", with_race=False))
if args.debug:
bnn.debug_mode()
if args.pretrained:
bnn.load_bayes_samples('repro/recid-no-race_svgd1.pt', 'svgd_baseline')
else:
bnn.infer()
logging.info(f'[{bnn.uid}] Baseline evaluation, with race as feature:')
acc, f1 = eval_accuracy_and_f1_score(bnn, is_binary=True, X_eval=bnn.X_train, Y_eval=bnn.Y_train)
logging.info(f'[{bnn.uid}] Train accuracy: {acc:.3f}.')
logging.info(f'[{bnn.uid}] Train F1 score: {f1:.3f}.')
aa_highrisk, aa_total, nonaa_highrisk, nonaa_total = recid_evaluate_high_risk(bnn)
logging.info(f'[{bnn.uid}] African American High-Risk Fraction: {(aa_highrisk / aa_total):.3f}.')
logging.info(f'[{bnn.uid}] Non-African American High-Risk Fraction: {(nonaa_highrisk / nonaa_total):.3f}.')
# Without race as feature, OC-BNN.
bnn.add_probabilistic_constraint(constrained_domain=cdomain[:-2], prob_func=lambda X: X[:,1], prior_type="gaussian_aocp")
if args.pretrained:
bnn.load_gaussian_aocp_parameters('repro/recid-no-race_aocp.pt')
else:
bnn.learn_gaussian_aocp()
bnn.update_config(use_ocbnn=True)
if args.pretrained:
bnn.load_bayes_samples('repro/recid-no-race_svgd2.pt', 'svgd_ocbnn')
else:
bnn.infer()
logging.info(f'[{bnn.uid}] OC-BNN evaluation, with race as feature:')
acc, f1 = eval_accuracy_and_f1_score(bnn, is_binary=True, X_eval=bnn.X_train, Y_eval=bnn.Y_train)
logging.info(f'[{bnn.uid}] Train accuracy: {acc:.3f}.')
logging.info(f'[{bnn.uid}] Train F1 score: {f1:.3f}.')
aa_highrisk, aa_total, nonaa_highrisk, nonaa_total = recid_evaluate_high_risk(bnn)
logging.info(f'[{bnn.uid}] African American High-Risk Fraction: {(aa_highrisk / aa_total):.3f}.')
logging.info(f'[{bnn.uid}] Non-African American High-Risk Fraction: {(nonaa_highrisk / nonaa_total):.3f}.')
def credit_evaluate_recourse(bnn):
""" Credit scoring task: evaluate effort of recourse on young adults. """
samples, _ = bnn.all_bayes_samples[-1]
preds = bnn.predict(samples, bnn.X_test, return_probs=True).mean(dim=0).argmax(dim=1)
is_young = (bnn.X_test[:, 1] * bnn.X_train_std[1] + bnn.X_train_mean[1] < 35).nonzero().squeeze(dim=1)
X_test_young = bnn.X_test[is_young]
Y_test_young = bnn.Y_test[is_young]
preds_young = preds[is_young]
gt_ruul_neg = X_test_young[(Y_test_young == 0).nonzero(), 0].mean()
gt_ruul_pos = X_test_young[(Y_test_young == 1).nonzero(), 0].mean()
preds_ruul_neg = X_test_young[(preds_young == 0).nonzero(), 0].mean()
preds_ruul_pos = X_test_young[(preds_young == 1).nonzero(), 0].mean()
return gt_ruul_neg, gt_ruul_pos, preds_ruul_neg, preds_ruul_pos
def credit_scoring_task():
""" Section 6.3: Credit Scoring Task """
bnn = BNNBBBClassifier(uid='credit', configfile=args.custom_config or "repro/credit.yaml")
bnn.load(**credit_dataset(csv_filename="data/give_me_some_credit.csv"))
logging.info(f'[{bnn.uid}] Dataset <{bnn.dataset_name}> blind dataset: {bnn.X_train_blind.shape[0]} training points.')
logging.info(f'[{bnn.uid}] Dataset <{bnn.dataset_name}>: {(100 * sum(bnn.Y_train) // bnn.N_train):.2f}% of training points are positive.')
logging.info(f'[{bnn.uid}] Dataset <{bnn.dataset_name}>: {(100 * sum(bnn.Y_train_blind) // bnn.X_train_blind.shape[0]):.2f}% of blind training points are positive.')
logging.info(f'[{bnn.uid}] Dataset <{bnn.dataset_name}>: {(100 * sum(bnn.Y_test) // bnn.N_test):.2f}% of test points are positive.')
if args.debug:
bnn.debug_mode()
# Baseline inference and evaluation on full dataset.
if args.pretrained:
bnn.load_bayes_samples('repro/credit_bbb1.pt', 'bbb_baseline')
else:
bnn.infer()
logging.info(f'[{bnn.uid}] Baseline evaluation on full dataset:')
acc, f1 = eval_accuracy_and_f1_score(bnn)
logging.info(f'[{bnn.uid}] Test accuracy: {acc:.3f}.')
logging.info(f'[{bnn.uid}] Test F1 score: {f1:.3f}.')
gt_ruul_neg, gt_ruul_pos, preds_ruul_neg, preds_ruul_pos = credit_evaluate_recourse(bnn)
logging.info(f'[{bnn.uid}] Effort of recourse: {preds_ruul_pos - preds_ruul_neg:.3f}.')
logging.info(f'[{bnn.uid}] Ground-truth effort of recourse: {gt_ruul_pos - gt_ruul_neg:.3f}.')
# Calculating constrained domain boundaries for OC-BNNs.
# Define constrained region as: young with high RUUL. For other features, sample around training distribution mean.
age_lb = bnn.X_train[:, 1].min().item()
age_ub = (35 - bnn.X_train_mean[1]) / bnn.X_train_std[1]
Y_train_young = bnn.Y_train[(bnn.X_train[:,1] * bnn.X_train_std[1] + bnn.X_train_mean[1] < 35).nonzero().squeeze(dim=1)]
X_train_young = bnn.X_train[(bnn.X_train[:,1] * bnn.X_train_std[1] + bnn.X_train_mean[1] < 35).nonzero().squeeze(dim=1)]
ruul_lb = np.quantile(X_train_young[:, 0], 0.6)
ruul_ub = np.quantile(X_train_young[:, 0], 1.0)
cdomain = (ruul_lb, ruul_ub, age_lb, age_ub)
for i in range(2, 10):
dim_mean = bnn.X_train[bnn.X_train[:,1] * bnn.X_train_std[1] + bnn.X_train_mean[1] < 35].mean(dim=0)[i].item()
dim_std = bnn.X_train[bnn.X_train[:,1] * bnn.X_train_std[1] + bnn.X_train_mean[1] < 35].std(dim=0)[i].item()
cdomain = cdomain + (dim_mean - 0.2 * dim_std, dim_mean + 0.2 * dim_std)
# Baseline inference and evaluation on blind dataset.
bnn.X_train, bnn.Y_train = bnn.X_train_blind, bnn.Y_train_blind
bnn.N_train = bnn.X_train.shape[0]
if args.pretrained:
bnn.load_bayes_samples('repro/credit_bbb2.pt', 'bbb_blind_baseline')
else:
bnn.infer()
logging.info(f'[{bnn.uid}] Baseline evaluation on blind dataset:')
acc, f1 = eval_accuracy_and_f1_score(bnn)
logging.info(f'[{bnn.uid}] Test accuracy: {acc:.3f}.')
logging.info(f'[{bnn.uid}] Test F1 score: {f1:.3f}.')
gt_ruul_neg, gt_ruul_pos, preds_ruul_neg, preds_ruul_pos = credit_evaluate_recourse(bnn)
logging.info(f'[{bnn.uid}] Effort of recourse: {preds_ruul_pos - preds_ruul_neg:.3f}.')
logging.info(f'[{bnn.uid}] Ground-truth effort of recourse: {gt_ruul_pos - gt_ruul_neg:.3f}.')
# OC-BNN inference and evaluation.
bnn.add_deterministic_constraint(constrained_domain=cdomain, forbidden_classes=[1], prior_type="positive_dirichlet_cocp")
bnn.update_config(use_ocbnn=True)
if args.pretrained:
bnn.load_bayes_samples('repro/credit_bbb3.pt', 'bbb_ocbnn')
else:
bnn.infer()
logging.info(f'[{bnn.uid}] OC-BNN evaluation:')
acc, f1 = eval_accuracy_and_f1_score(bnn)
logging.info(f'[{bnn.uid}] Test accuracy: {acc:.3f}.')
logging.info(f'[{bnn.uid}] Test F1 score: {f1:.3f}.')
gt_ruul_neg, gt_ruul_pos, preds_ruul_neg, preds_ruul_pos = credit_evaluate_recourse(bnn)
logging.info(f'[{bnn.uid}] Effort of recourse: {preds_ruul_pos - preds_ruul_neg:.3f}.')
logging.info(f'[{bnn.uid}] Ground-truth effort of recourse: {gt_ruul_pos - gt_ruul_neg:.3f}.')
if __name__ == "__main__":
logging.basicConfig(level=logging.INFO, format="%(name)s-%(levelname)s: %(message)s")
logging.info("Running OC-BNN library...")
parser = argparse.ArgumentParser(description='Process command-line arguments.')
parser.add_argument('--debug', action='store_true')
parser.add_argument('--pretrained', action='store_true')
parser.add_argument('--custom_config', nargs='?', const='config.yaml')
args = parser.parse_args()
recidivism_task()
credit_scoring_task()
logging.info("Completed.")