-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathood_base.py
175 lines (124 loc) · 7.78 KB
/
ood_base.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
from distutils.command.config import config
import enum
from tabnanny import verbose
from tkinter.messagebox import NO
import torch
import argparse
from torch._C import device
from confidnet.models import get_model
from confidnet.utils.misc import load_yaml
from confidnet.loaders import get_loader
import torch.nn.functional as F
from torch.utils.data import DataLoader
import numpy as np
from rot_learner.dataset import RotDataset
from rot_learner.dataset_trans import TransDataset
from rot_learner.util import eval_res, retrieve_data
from rot_learner.data_util import preprocess_config
from tqdm import tqdm
import copy
def get_weighted_res(all_scores, args, device, metrics=['auroc'], choose_smaller=False, cached_models=None, return_scores=False):
# train_scores, val_scores, test_scores = all_scores
val_scores = all_scores['val']
test_scores = all_scores['test']
model, features_extractor, dloader,\
trainset_data, trainset_targets,\
valset_data, valset_targets,\
dataset, class_num = preprocess_config(args.config_path, args, device)
if dataset == 'stl10':
# TRANS = [(0, 0), (0, -32), (0, 32), (-32, 0), (32, 0)]
TRANS = [(0, 0), (0, -32), (0, 32)]
# TRANS = [(0, 0), (-32, 0), (32, 0)]
# TRANS = [(0, 0), (-32, 0), (32, 0)]
PRED_ROTS = [0, 90]
else:
TRANS = [(0, 0), (0, -8), (0, 8), (-8, 0), (8, 0)]
PRED_ROTS = [0, 90, 180, 270]
TRANS_TASK_NUM = len(TRANS)
PRED_TASK_NUM = len(PRED_ROTS)
train_ds = RotDataset(trainset_data, PRED_ROTS, trainset_targets, dataset=dataset, train_mode=True)
val_ds = RotDataset(valset_data, PRED_ROTS, valset_targets, dataset=dataset, train_mode=False)
test_ds = RotDataset(dloader.all_test_data, PRED_ROTS, dloader.all_test_targets, dataset=dataset, train_mode=False)
train_trans_ds = TransDataset(trainset_data, TRANS, trainset_targets, dataset=dataset, train_mode=True)
val_trans_ds = TransDataset(valset_data, TRANS, valset_targets, dataset=dataset, train_mode=False)
test_trans_ds = TransDataset(dloader.all_test_data, TRANS, dloader.all_test_targets, dataset=dataset, train_mode=False)
# print('common idx', list(set(dloader.train_idx).intersection(dloader.val_idx)))
print('train_ds', len(train_ds))
print('val_ds', len(val_ds))
trainloader = DataLoader(train_ds, batch_size=300, shuffle=True)
valloader = DataLoader(val_ds, batch_size=100, shuffle=False)
# train_valloader = DataLoader(train_val_ds, batch_size=100, shuffle=True)
testloader = DataLoader(test_ds, batch_size=128, shuffle=False, pin_memory=False, num_workers=3)
# test_trainloader = DataLoader(test_train_ds, batch_size=128, shuffle=False, pin_memory=False, num_workers=3)
trans_trainloader = DataLoader(train_trans_ds, batch_size=128, shuffle=True)
trans_valloader = DataLoader(val_trans_ds, batch_size=128, shuffle=False)
trans_testloader = DataLoader(test_trans_ds, batch_size=128, shuffle=False)
# rot predict model
from rot_learner import main, main_cinic10, main_stl10
if dataset == 'cinic10':
main = main_cinic10
if dataset == 'stl10':
main = main_stl10
if cached_models is None:
net = main.train(features_extractor, len(PRED_ROTS), 512, args.ssl_epoch, trainloader, device)
trans_net = main.train(copy.deepcopy(features_extractor), TRANS_TASK_NUM, 512, args.ssl_epoch, trans_trainloader, device)
else:
net = cached_models['rot_net']
trans_net = cached_models['trans_net']
# class_num = int(config_args['data']['num_classes'])
# eval test rot
cls_res, aug_res = main.test(net, model, testloader, class_num, len(PRED_ROTS), device)
val_cls_res, val_aug_res = main.test(net, model, valloader, class_num, len(PRED_ROTS), device)
# train_val_cls_res, train_val_aug_res = main.test(net, model, train_valloader, class_num, len(PRED_ROTS), device)
# train_cls_res, train_aug_res = main.test(net, model, test_trainloader, class_num, len(PRED_ROTS), device)
pred_scores, pred_corrects, gt_labels, pred_labels, pred_scores_all, pred_scores_for_gts, pred_features, pred_logits, pred_loss_all = cls_res
rot_scores, rot_accs, rot_scores_full, rot_scores_all, aug_features, aug_logits, all_rot_corrects, aug_loss_all = aug_res
val_pred_scores, val_pred_corrects, val_gt_labels, val_pred_labels, val_pred_scores_all, val_pred_scores_for_gts, val_pred_features, val_pred_logits, val_pred_loss_all = val_cls_res
val_rot_scores, val_rot_accs, val_rot_scores_full, val_rot_scores_all, val_aug_features, val_aug_logits, val_all_rot_corrects, val_aug_loss_all = val_aug_res
trans_cls_res, trans_aug_res = main.test(trans_net, model, trans_testloader, class_num, TRANS_TASK_NUM, device)
trans_val_cls_res, trans_val_aug_res = main.test(trans_net, model, trans_valloader, class_num, TRANS_TASK_NUM, device)
# trans_pred_scores, trans_pred_corrects, trans_gt_labels, trans_pred_labels, trans_pred_scores_all, trans_pred_scores_for_gts, trans_pred_features, trans_pred_logits, trans_pred_loss_all = trans_cls_res
trans_scores, trans_accs, trans_scores_full, trans_scores_all, trans_aug_features, trans_aug_logits, all_trans_corrects, trans_aug_loss_all = trans_aug_res
# val_pred_scores, val_pred_corrects, val_gt_labels, val_pred_labels, val_pred_scores_all, val_pred_scores_for_gts, val_pred_features, val_pred_logits, val_pred_loss_all = trans_val_cls_res
val_trans_scores, val_trans_accs, val_trans_scores_full, val_trans_scores_all, val_trans_aug_features, val_trans_aug_logits, val_all_trans_corrects, val_trans_aug_loss_all = trans_val_aug_res
def cal_entropy(pred_vector):
return - np.sum( pred_vector * np.log(pred_vector), -1)
def find_best_avg(val_info, test_info , trials=50, metric='auroc'):
pred_scores, rot_scores, trans_scores, pred_corrects = val_info
test_scores, test_rot_scores, test_trans_scores = test_info
best_auroc = 0
best_config = np.zeros(2)
if metric == 'erm':
pass
else:
for i in tqdm(range(trials)):
for j in range(trials):
# scores = i /trials * pred_scores + (1 - i/trials) * rot_scores
if 1 - ( (i + j) / trials) < 0:
continue
scores = i / trials * rot_scores + j / trials * trans_scores + (1 - (i+j)/trials) * pred_scores
# set false later
res = eval_res(scores, pred_corrects, verbose=False)
auroc = res[metric]
criteria = auroc < best_auroc if choose_smaller else auroc > best_auroc
if criteria:
best_auroc = auroc
best_config = (i / trials, j / trials)
new_scores = best_config[0] * test_rot_scores + best_config[1] * test_trans_scores\
+ (1-best_config[0]-best_config[1]) * test_scores
return new_scores, best_config
all_new_res = {}
for metric in metrics:
new_scores, score_config = find_best_avg(
(np.array(val_scores), np.array(val_rot_scores_full)[:, 0], np.array(val_trans_scores_full)[:, 0], np.array(val_pred_corrects)),
(np.array(test_scores), np.array(rot_scores_full)[:, 0], np.array(trans_scores_full)[:, 0]),
metric=metric
)
all_new_res[metric] = {**eval_res(new_scores, pred_corrects), 'config': copy.deepcopy(score_config)}
print('baseline (without weighted)')
eval_res(test_scores, pred_corrects)
if return_scores:
return all_new_res, {'rot_net': net, 'trans_net': trans_net},\
{'pred': np.array(pred_scores), 'rot': np.array(rot_scores_full)[:, 0], 'trans': np.array(trans_scores_full)[:, 0]}
else:
return all_new_res, {'rot_net': net, 'trans_net': trans_net}