-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathget_datasets.py
176 lines (124 loc) · 5.59 KB
/
get_datasets.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
from data.data_utils import MergedDataset
from data.cifar import get_cifar_10_datasets, get_cifar_100_datasets
from data.herbarium_19 import get_herbarium_datasets
from data.stanford_cars import get_scars_datasets
from data.imagenet import get_imagenet_100_datasets, get_imagenet_1k_datasets
from data.cub import get_cub_datasets
from data.fgvc_aircraft import get_aircraft_datasets
from copy import deepcopy
import pickle
import os
from config import osr_split_dir
get_dataset_funcs = {
'cifar10': get_cifar_10_datasets,
'cifar100': get_cifar_100_datasets,
'imagenet_100': get_imagenet_100_datasets,
'imagenet_1k': get_imagenet_1k_datasets,
'herbarium_19': get_herbarium_datasets,
'cub': get_cub_datasets,
'aircraft': get_aircraft_datasets,
'scars': get_scars_datasets
}
def get_datasets(dataset_name, train_transform, test_transform, args):
"""
:return: train_dataset: MergedDataset which concatenates labelled and unlabelled
test_dataset,
unlabelled_train_examples_test,
datasets
"""
#
if dataset_name not in get_dataset_funcs.keys():
raise ValueError
# Get datasets
get_dataset_f = get_dataset_funcs[dataset_name]
datasets = get_dataset_f(train_transform=train_transform, test_transform=test_transform,
train_classes=args.train_classes,
prop_train_labels=args.prop_train_labels,
split_train_val=False)
# Set target transforms:
target_transform_dict = {}
for i, cls in enumerate(list(args.train_classes) + list(args.unlabeled_classes)):
target_transform_dict[cls] = i
target_transform = lambda x: target_transform_dict[x]
for dataset_name, dataset in datasets.items():
if dataset is not None:
dataset.target_transform = target_transform
# Train split (labelled and unlabelled classes) for training
train_dataset = MergedDataset(labelled_dataset=deepcopy(datasets['train_labelled']),
unlabelled_dataset=deepcopy(datasets['train_unlabelled']))
test_dataset = datasets['test']
unlabelled_train_examples_test = deepcopy(datasets['train_unlabelled'])
unlabelled_train_examples_test.transform = test_transform
return train_dataset, test_dataset, unlabelled_train_examples_test, datasets
def get_class_splits(args):
# For FGVC datasets, optionally return bespoke splits
if args.dataset_name in ('scars', 'cub', 'aircraft'):
if hasattr(args, 'use_ssb_splits'):
use_ssb_splits = args.use_ssb_splits
else:
use_ssb_splits = False
# -------------
# GET CLASS SPLITS
# -------------
if args.dataset_name == 'cifar10':
args.image_size = 32
args.train_classes = range(5)
args.unlabeled_classes = range(5, 10)
elif args.dataset_name == 'cifar100':
args.image_size = 32
args.train_classes = range(80)
args.unlabeled_classes = range(80, 100)
elif args.dataset_name == 'herbarium_19':
args.image_size = 224
herb_path_splits = os.path.join(osr_split_dir, 'herbarium_19_class_splits.pkl')
with open(herb_path_splits, 'rb') as handle:
class_splits = pickle.load(handle)
args.train_classes = class_splits['Old']
args.unlabeled_classes = class_splits['New']
elif args.dataset_name == 'imagenet_100':
args.image_size = 224
args.train_classes = range(50)
args.unlabeled_classes = range(50, 100)
elif args.dataset_name == 'imagenet_1k':
args.image_size = 224
args.train_classes = range(500)
args.unlabeled_classes = range(500, 1000)
elif args.dataset_name == 'scars':
args.image_size = 224
if use_ssb_splits:
split_path = os.path.join(osr_split_dir, 'scars_osr_splits.pkl')
with open(split_path, 'rb') as handle:
class_info = pickle.load(handle)
args.train_classes = class_info['known_classes']
open_set_classes = class_info['unknown_classes']
args.unlabeled_classes = open_set_classes['Hard'] + open_set_classes['Medium'] + open_set_classes['Easy']
else:
args.train_classes = range(98)
args.unlabeled_classes = range(98, 196)
elif args.dataset_name == 'aircraft':
args.image_size = 224
if use_ssb_splits:
split_path = os.path.join(osr_split_dir, 'aircraft_osr_splits.pkl')
with open(split_path, 'rb') as handle:
class_info = pickle.load(handle)
args.train_classes = class_info['known_classes']
open_set_classes = class_info['unknown_classes']
args.unlabeled_classes = open_set_classes['Hard'] + open_set_classes['Medium'] + open_set_classes['Easy']
else:
args.train_classes = range(50)
args.unlabeled_classes = range(50, 100)
elif args.dataset_name == 'cub':
args.image_size = 224
if use_ssb_splits:
split_path = os.path.join(osr_split_dir, 'cub_osr_splits.pkl')
with open(split_path, 'rb') as handle:
class_info = pickle.load(handle)
args.train_classes = class_info['known_classes']
open_set_classes = class_info['unknown_classes']
args.unlabeled_classes = open_set_classes['Hard'] + open_set_classes['Medium'] + open_set_classes['Easy']
else:
args.train_classes = range(100)
args.unlabeled_classes = range(100, 200)
else:
raise NotImplementedError
return args