-
Notifications
You must be signed in to change notification settings - Fork 0
/
__main__.py
257 lines (200 loc) · 10.4 KB
/
__main__.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
"""
Copyright©[2023] Fraunhofer-Gesellschaft zur Foerderung der angewandten Forschung e.V. acting on behalf of its Fraunhofer-Institut für Kognitive Systeme IKS. All rights reserved.
This software is subject to the terms and conditions of the GNU GPLv2 (https://www.gnu.de/documents/gpl-2.0.de.html).
Contact: nicola.franco@iks.fraunhofer.de
"""
import torch
import utils, metrics
from pathlib import Path
from data_loader import prepare_data, prepare_training_data
from architectures.loader import load_predictor
import adversarial.aauc as internal_methods
import numpy as np
import argparse
def main(args):
""" Main function to run the experiments
Args:
args (argparse.Namespace): Arguments of the experiment
Returns:
None
"""
torch.manual_seed(1)
np.random.seed(1)
parameters = {}
models_path, output_path, dataset_path = utils.prepare_paths(args)
num_classes = 10 if args.dataset == 'cifar10' else 100
if args.experiment == 'atom': num_classes += 1
parameters['model'] = load_predictor(args, models_path)
id_loader, loaders = prepare_data(
dataset_path=dataset_path, batch_size=args.batch_size, dataset=args.dataset)
parameters['loader'] = id_loader
print(f'ID: Clean Accuracy {utils.compute_accuracy(**parameters):.2f}')
if args.finetune:
parameters['model'] = fine_tuning(args, parameters, dataset_path, models_path)
if args.experiment == 'atom': parameters['atom'] = True
if args.adv_robustness:
""" Compute the robustness of the model on in distribution data """
for eps in [2/255, 8/255]:
robustness = internal_methods.compute_adv_accuracy(
parameters['model'], loader=parameters['loader'], epsilon=eps,
version='rand' if args.experiment in ['distro', 'diffusion'] else 'standard',
output_path=output_path
)
print(f'\nID: Adversarial Robustness {robustness:.2f}\n')
if args.certify_robustness:
for sigma in [0.12, 0.25]:
robustness = internal_methods.compute_certified_accuracy(
parameters['model'], loader=parameters['loader'],
sigma=sigma, batch_size=args.batch_size, num_classes=num_classes
)
values = 'ID: Certified Robustness '
for r, value in robustness.items():
values += f'radius: {r} = {100*value:.2f}, '
print(values)
parameters['score_type'] = args.score
id_score = metrics.compute_score(**parameters)
del parameters['score_type']
if args.clean:
print(" Compute the clean AUC, AUPR, and FPR@95 ")
parameters['temperature'] = 1
parameters['score_type'] = args.score
compute_auroc(
args, parameters, loaders, id_score, output_path, num_classes, name='clean'
)
del parameters['temperature'], parameters['score_type']
if args.guar:
print(" Compute the guaranteed l-infinity norm AUC, AUPR, and FPR@95 ")
parameters['epsilon'] = 0.01
compute_auroc(
args, parameters, loaders, id_score, output_path, num_classes, name='guaranteed'
)
del parameters['epsilon']
if args.certify:
print(" Compute the guaranteed l-2 norm AUC, AUPR, and FPR@95 ")
certify_path = output_path/Path('certify')
certify_path.mkdir(parents=True, exist_ok=True)
compute_certify_auroc(
args, parameters, loaders, id_loader, certify_path, num_classes)
if args.adv:
print(" Compute the adversarial AUC, AUPR, and FPR@95 ")
parameters['epsilon'] = 0.01
parameters['num_classes'] = 10
parameters['temperature'] = 1
parameters['score_type'] = args.score
compute_auroc(
args, parameters, loaders, id_score, output_path, num_classes, name='adversarial'
)
del parameters['temperature'], parameters['num_classes'], parameters['score_type']
def fine_tuning(args, parameters, dataset_path, models_path):
parameters['loader'] = prepare_training_data(
dataset_path=dataset_path, batch_size=args.batch_size
)
if args.experiment in ['vos', 'logit'] and args.diff:
predictor = utils.fine_tune(**parameters)
print(f'ID: Clean Accuracy after fine-tuning {utils.compute_accuracy(**parameters):.2f}')
if args.experiment == 'vos':
torch.save(predictor.state_dict(),
models_path/Path('our/CIFAR10/vos/fine_tuned_vos_cifar10.pt')
)
elif args.experiment == 'logit':
torch.save(predictor.state_dict(),
models_path/Path('our/CIFAR10/logitnorm/fine_tuned_logit_cifar10.pt')
)
return predictor
def compute_certify_auroc(args, parameters, loaders, id_loader, output_path, num_classes=10):
""" Compute the AUC, AUPR, and FPR@95 for the guaranteed scores """
for sigma in [0.12]:
parameters['ranges'] = np.arange(0, 2.1, 0.25)
results, columns = utils.init_certify_results(parameters['ranges'])
parameters['sigma'] = sigma
parameters['loader'] = id_loader
parameters['batch_size'] = args.batch_size
parameters['num_classes'] = num_classes
id_certify_scores = metrics.compute_certify_score(**parameters)
for dataset_name, loader in loaders.items():
parameters['loader'] = loader
ood_certify_scores = metrics.compute_certify_score(**parameters)
for r, result in results.items():
results[r]['Dataset'].append(dataset_name)
id_score = np.array(id_certify_scores[r])
ood_score = np.array(ood_certify_scores[r])
if np.count_nonzero(id_score == 0) < len(id_score) or \
np.count_nonzero(ood_score == 0) < len(ood_score):
measures = metrics.get_measures(id_score, ood_score)
results[r], output = utils.store_results(
dataset_name=dataset_name,
results=result,
measures=measures,
cols=columns)
print(f'Sigma: {sigma}, Radius: {r}, ' + output)
if r == '0.0':
np.save(
output_path/f'{sigma}_{dataset_name}',
np.array([id_score, ood_score])
)
for r, result in results.items():
path = output_path/f'{str(sigma)}_{str(r)}'
if result['AUC']:
utils.store_average(
results=result, cols=columns, output_path=path)
del parameters['ranges'], parameters['sigma'], parameters['batch_size']
def compute_auroc(args, parameters, loaders, id_score, output_path, num_classes, name:str = 'clean'):
""" Compute the clean AUC, AUPR, and FPR@95 """
results, columns = utils.init_results()
for dataset_name, loader in loaders.items():
parameters['loader'] = loader
# compute AUROC
results['Dataset'].append(dataset_name)
if name == 'clean':
ood_score = metrics.compute_score(**parameters)
elif name == 'guaranteed':
if args.experiment in ['good']:
parameters['num_classes'] = num_classes
ood_score = metrics.get_conf_ibp_good(**parameters)
elif args.experiment in ['prood', 'distro']:
ood_score = metrics.get_conf_ibp(**parameters)
else:
parameters['num_classes'] = num_classes
ood_score = metrics.get_conf_ibp_general(**parameters)
elif name == 'adversarial':
parameters['num_classes'] = num_classes
ood_score = internal_methods.get_conf_lb(**parameters)
else:
raise NotImplementedError(f'{name} is not implemented')
measures = metrics.get_measures(id_score, ood_score)
results, output = utils.store_results(
dataset_name=dataset_name, results=results, measures=measures, cols=columns)
print(output)
utils.store_average(results=results, cols=columns, output_path=output_path/Path(name))
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Compute the AUC, AUPR, and FPR@95 for the DDS models')
parser.add_argument('--all', type=utils.str2bool, nargs='?',
const=True, default=False, help='Compute clean, guaranteed and adversarial AUC, AUFPR, FPR95')
parser.add_argument('--clean', type=utils.str2bool, nargs='?',
const=True, default=False, help='Compute clean AUC, AUFPR, FPR95')
parser.add_argument('--guar', type=utils.str2bool, nargs='?',
const=True, default=False, help='Compute guaranteed AUC, AUFPR, FPR95')
parser.add_argument('--adv', type=utils.str2bool, nargs='?',
const=True, default=False, help='Compute adversarial AUC, AUFPR, FPR95')
parser.add_argument('--certify', type=utils.str2bool, nargs='?',
const=True, default=False, help='Compute certify AUC, AUFPR, FPR95')
parser.add_argument('--adv_robustness', type=utils.str2bool, nargs='?',
const=True, default=False, help='Compute adversarial robustness on the ID dataset')
parser.add_argument('--certify_robustness', type=utils.str2bool, nargs='?',
const=True, default=False, help='Compute certified robustness on the ID dataset')
parser.add_argument('--diff', type=utils.str2bool, nargs='?',
const=True, default=False, help='Add diffusion model')
parser.add_argument('--finetune', type=utils.str2bool, nargs='?',
const=True, default=False, help='Fine-tune the model')
parser.add_argument('--experiment', type=str, default='vos',
choices=['distro', 'vos', 'logit', 'oe', 'prood', 'good', 'plain', 'acet', 'atom', 'diffusion'],
help='Experiment to run')
parser.add_argument('--dataset', type=str, default='cifar10',
choices=['cifar10', 'cifar100'], help='Dataset to use')
parser.add_argument('--score', type=str, default='softmax',
choices=['softmax', 'energy', 'logit'], help='Score to use')
parser.add_argument('--batch_size', type=int, default=128, help='Batch size')
parser.add_argument('--standardized', type=bool, default=False, help='Do standardized tests (with all models having similar model and normalization)')
args = parser.parse_args()
print(args)
main(args)