-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval.py
197 lines (158 loc) · 7.52 KB
/
eval.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
import warnings
warnings.filterwarnings('ignore')
import torch
import numpy as np
from models import models, wideresnet, combinednet
from util import evaluation as evalutil
from util import gmm_utils
import util.dataloaders as dl
import util.misc as misc
import argparse
import os
from collections import defaultdict
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', default='MNIST')
parser.add_argument('--ood_data', default='tiny300k', choices=['imagenet', 'tiny300k', 'uniform', 'smooth'])
args = parser.parse_args()
torch.manual_seed(9999)
np.random.seed(9999)
assert args.dataset in ['MNIST', 'FMNIST', 'CIFAR10', 'SVHN', 'CIFAR100']
# Note the order of the methods matters
methods = ['plain', 'ddu', 'noneclass', 'oe', 'preload', 'oe_finetuning', 'doe_ft', 'energy_finetuning', 'preload_ft']
pretrained_path = './pretrained_models'
path_ood = ('/'+args.ood_data)
_, test_loader = dl.datasets_dict[args.dataset](train=False, augm_flag=False, val_size=1000)
targets = torch.cat([y for x, y in test_loader], dim=0).numpy()
num_classes = 100 if args.dataset == 'CIFAR100' else 10
depth = 16
widen_factor = 4
if args.dataset in ['MNIST', 'FMNIST']:
size_of_Tx = 1024
else:
size_of_Tx = 64 * widen_factor
ood_noise_names = ['UniformNoise', 'Noise', 'FarAway', 'FarAway2']
if args.ood_data == 'uniform':
ood_noise_names = [ood_noise_names[1], ood_noise_names[2], ood_noise_names[3]]
elif args.ood_data == 'smooth':
ood_noise_names = [ood_noise_names[0], ood_noise_names[2], ood_noise_names[3]]
ood_test_names = {
'MNIST': ['FMNIST', 'EMNIST', 'KMNIST', 'GrayCIFAR10'],
'FMNIST': ['MNIST', 'EMNIST', 'KMNIST', 'GrayCIFAR10'],
'CIFAR10': ['SVHN', 'LSUN', 'CIFAR100', 'FMNIST3D'],
'SVHN': ['CIFAR10', 'LSUN', 'CIFAR100', 'FMNIST3D'],
'CIFAR100': ['SVHN', 'LSUN', 'CIFAR10', 'FMNIST3D'],
}
ood_names = ood_test_names[args.dataset] + ood_noise_names
ood_test_loaders = {}
for ood_name in ood_test_names[args.dataset]:
ood_test_loaders[ood_name] = dl.datasets_dict[ood_name](train=False)
for ood_name in ood_noise_names:
ood_test_loaders[ood_name] = dl.datasets_dict[ood_name](dataset=args.dataset, train=False, size=10000)
tab_mmc = defaultdict(lambda: defaultdict(list))
tab_auroc = defaultdict(lambda: defaultdict(list))
tab_auprc = defaultdict(lambda: defaultdict(list))
tab_fpr95 = defaultdict(lambda: defaultdict(list))
tab_acc = defaultdict(list)
tab_cal = defaultdict(list)
def load_model(method, seed, feat_extractor=False):
path = pretrained_path
if 'preload' in method:
model = combinednet.CombinedNet(args.dataset, num_classes)
else:
n_classes = num_classes+1 if method == 'noneclass' else num_classes
if args.dataset in ['MNIST', 'FMNIST']:
model = models.LeNetMadry(n_classes, feature_extractor=feat_extractor)
else:
model = wideresnet.WideResNet(depth, widen_factor, n_classes, feature_extractor=feat_extractor)
method_name = 'plain' if method == 'ddu' else method
if method_name != 'plain':
path += path_ood
path += f'/{args.dataset}_{method_name}_{seed}.pt'
model.cuda()
model.load_state_dict(torch.load(path))
model.eval()
return model
def predict_(model, method, test_dl, test_feature_loader):
py = evalutil.predict(test_dl, model)
return py.cpu().numpy()
@torch.no_grad()
def evaluate(method, seed, verbose=True):
model = load_model(method, seed)
USE_NONE_CLASS = method in ['noneclass', 'preload', 'preload_ft']
py_in = predict_(model, method, test_loader, test_feature_loaders[seed-1])
if method == 'ddu':
train_loader = dl.datasets_dict[args.dataset](train=True)
embeddings, labels = gmm_utils.get_embeddings(model, train_loader, size_of_Tx, torch.double, 'cuda', 'cuda')
gda, jitter_eps = gmm_utils.gmm_fit(embeddings, labels, num_classes)
train_log_probs_B_Y, train_labels = gmm_utils.gmm_evaluate(model, gda, train_loader, 'cuda', num_classes, 'cuda')
train_densities = torch.logsumexp(train_log_probs_B_Y, dim=-1)
train_min_density = train_densities.min().item()
in_densities = []
for x, y in test_loader:
x = x.cuda()
log_probs_B_Y = gmm_utils.gmm_forward(model, gda, x)
density = torch.logsumexp(log_probs_B_Y, dim=-1)-train_min_density
in_densities.append(density)
in_densities = torch.cat(in_densities)
elif method == 'energy_finetuning':
energy_in = evalutil.predict(test_loader, model, energy=True)
acc = evalutil.get_acc(py_in, targets)
ece, _ = evalutil.get_calib(py_in if not USE_NONE_CLASS else py_in[:, :-1], targets)
mmc = evalutil.get_mmc(py_in, USE_NONE_CLASS)
tab_mmc[method][args.dataset].append(mmc)
tab_acc[method].append(acc)
tab_cal[method].append(ece)
if verbose:
print(f'[In, {method}] Acc: {acc:.1f}; ECE: {ece:.1f}; MMC: {mmc:.3f}')
for ood_name, ood_test_loader in ood_test_loaders.items():
ood_test_feature_loader = misc.get_feature_loader(pretrained_plain_models[seed-1], ood_test_loader, batch_size=128, train=False)
py_out = predict_(model, method, ood_test_loader, ood_test_feature_loader)
mmc = evalutil.get_mmc(py_out, USE_NONE_CLASS)
if method == 'ddu':
out_densities = []
for x, y in ood_test_loader:
x = x.cuda()
log_probs_B_Y = gmm_utils.gmm_forward(model, gda, x)
density = torch.logsumexp(log_probs_B_Y, dim=-1)-train_min_density
out_densities.append(density)
out_densities = torch.cat(out_densities)
py_in = in_densities.cpu().numpy()
py_out = out_densities.cpu().numpy()
elif method == 'energy_finetuning':
energy_out = evalutil.predict(ood_test_loader, model, energy=True)
py_in = -energy_in.cpu().numpy()
py_out = -energy_out.cpu().numpy()
auroc = evalutil.get_auroc(py_in, py_out, USE_NONE_CLASS, method == 'ddu', method == 'energy_finetuning')
auprc = evalutil.get_aupr(py_in, py_out, USE_NONE_CLASS, method == 'ddu', method == 'energy_finetuning')
fpr95, _ = evalutil.get_fpr95(py_in, py_out, USE_NONE_CLASS, method == 'ddu', method == 'energy_finetuning')
tab_mmc[method][ood_name].append(mmc)
tab_auroc[method][ood_name].append(auroc)
tab_auprc[method][ood_name].append(auprc)
tab_fpr95[method][ood_name].append(fpr95)
if verbose:
print(f'[Out-{ood_name}, {method}] MMC: {mmc:.1f}; AUROC: {auroc:.1f}; '
+ f'AUPRC: {auprc:.1f} FPR@95: {fpr95:.1f}')
if verbose:
print()
pretrained_plain_models = [load_model('plain', seed, feat_extractor=True) for seed in range(1, 6)]
test_feature_loaders = [misc.get_feature_loader(pretrained_model, test_loader, batch_size=128, train=False) for pretrained_model in pretrained_plain_models]
print('----------', args.dataset, args.ood_data, '----------')
for method in methods:
print('Method:', method)
for seed in range(1, 6):
print('Seed:', seed)
evaluate(method, seed)
print()
print('-------------------------------------------')
print()
# Save results
dir_name = f'results/'
dir_name += f'{args.ood_data}/{args.dataset}'
if not os.path.exists(dir_name):
os.makedirs(dir_name)
np.save(f'{dir_name}/mmc', dict(tab_mmc))
np.save(f'{dir_name}/auroc', dict(tab_auroc))
np.save(f'{dir_name}/auprc', dict(tab_auprc))
np.save(f'{dir_name}/fpr95', dict(tab_fpr95))
np.save(f'{dir_name}/acc', dict(tab_acc))
np.save(f'{dir_name}/cal', dict(tab_cal))