-
Notifications
You must be signed in to change notification settings - Fork 12
/
data_holder.py
223 lines (198 loc) · 12.3 KB
/
data_holder.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
from torch.utils.data import DataLoader
from torch.utils.data.sampler import SubsetRandomSampler
from data_io.fewshot import FewShot, extract_fewshot_data
from data_io.noisy import extract_noisy_data
from data_io.semi import extract_semi_data
from augmentation.domain_gen import domain_gen
from augmentation.type import aug_type
class DataHolder(object):
def __init__(self, dataset_class, config):
self.config = config
self.dataset_class = dataset_class
train_data_transform = aug_type(self.config['train_aug_type'], self.config)
valid_data_transform = aug_type(self.config['valid_aug_type'], self.config)
self.train_dataset = dataset_class(data_path=self.config['data_path'],
learning_mode=self.config['learning_mode'],
phase='train',
data_transform=train_data_transform,
num_task=self.config['num_task'])
self.valid_dataset = dataset_class(data_path=self.config['data_path'],
learning_mode=self.config['learning_mode'],
phase='test',
data_transform=valid_data_transform,
num_task=self.config['num_task'])
self.refer_dataset = extract_fewshot_data(self.train_dataset, self.config['ref_num'])
self.vis_dataset = extract_fewshot_data(self.valid_dataset, self.config['vis_num'])
self.class_name = self.train_dataset.class_name
self.train_loaders, self.valid_loaders = [], []
self.refer_loaders, self.vis_loaders = [], []
self.chosen_train_loaders, self.chosen_valid_loaders = [], []
self.chosen_vis_loaders = []
# only for transfer
self.chosen_transfer_train_loaders, self.chosen_transfer_valid_loaders = [], []
self.chosen_transfer_vis_loaders = []
# vanilla training
for i in range(self.config['num_task']):
train_task_data_list = self.train_dataset.sample_indices_in_task
train_loader = DataLoader(self.train_dataset,
batch_size=self.config['train_batch_size'],
num_workers=self.config['num_workers'],
sampler=SubsetRandomSampler(train_task_data_list[i]),
drop_last=False)
self.train_loaders.append(train_loader)
valid_task_data_list = self.valid_dataset.sample_indices_in_task
valid_loader = DataLoader(self.valid_dataset,
batch_size=self.config['valid_batch_size'],
num_workers=self.config['num_workers'],
shuffle=False,
sampler=SubsetRandomSampler(valid_task_data_list[i]),
drop_last=False)
self.valid_loaders.append(valid_loader)
refer_task_data_list = self.refer_dataset.sample_indices_in_task
refer_loader = DataLoader(self.train_dataset,
batch_size=self.config['ref_num'],
num_workers=self.config['num_workers'],
sampler=SubsetRandomSampler(refer_task_data_list[i]),
drop_last=False)
self.refer_loaders.append(refer_loader)
vis_task_data_list = self.vis_dataset.sample_indices_in_task
vis_loader = DataLoader(self.vis_dataset,
batch_size=self.config['valid_batch_size'],
num_workers=self.config['num_workers'],
sampler=SubsetRandomSampler(vis_task_data_list[i]),
drop_last=False)
self.vis_loaders.append(vis_loader)
def create(self):
if self.config['vanilla']:
self.create_vanilla()
if self.config['fewshot']:
self.create_fewshot()
if self.config['noisy']:
self.create_noisy()
if self.config['semi']:
self.create_semi()
if self.config['continual']:
self.create_continual()
if self.config['transfer']:
self.create_transfer()
def create_vanilla(self):
self.create_chosen_dataloaders()
def create_fewshot(self):
self.train_fewshot_dataset = extract_fewshot_data(self.train_dataset, self.config['fewshot_exm'])
# capture few-shot images
fewshot_images = []
fewshot_task_data_list = self.train_fewshot_dataset.sample_indices_in_task
for i in range(self.config['num_task']):
img_list = []
for idx in fewshot_task_data_list[i]:
img_list.append(self.train_fewshot_dataset[idx])
fewshot_images.append(img_list)
# data augumentation
if self.config['fewshot_data_aug']:
fewshot_images_dg = []
for i in range(self.config['num_task']):
data_gen_dataset = domain_gen(self.config, fewshot_images[i])
fewshot_images_dg.append(data_gen_dataset)
fewshot_images = fewshot_images_dg
# back to normal training
train_fewshot_loaders = []
for i in range(self.config['num_task']):
fewshot_dg_datset = FewShot(fewshot_images[i])
train_fewshot_loader = DataLoader(fewshot_dg_datset,
batch_size=self.config['train_batch_size'],
num_workers=self.config['num_workers'])
train_fewshot_loaders.append(train_fewshot_loader)
self.train_loaders = train_fewshot_loaders
self.create_chosen_dataloaders()
def create_noisy(self):
self.train_noisy_dataset, self.valid_noisy_dataset, self.noisy_dataset = extract_noisy_data(self.train_dataset,
self.valid_dataset,
noisy_ratio=self.config['noisy_ratio'],
noisy_overlap=self.config['noisy_overlap'])
train_task_data_list = self.train_noisy_dataset.sample_indices_in_task
valid_task_data_list = self.valid_noisy_dataset.sample_indices_in_task
train_noisy_loaders, valid_noisy_loaders = [], []
for i in range(self.config['num_task']):
train_noisy_loader = DataLoader(self.train_noisy_dataset,
batch_size=self.config['train_batch_size'],
num_workers=self.config['num_workers'],
sampler=SubsetRandomSampler(train_task_data_list[i]))
train_noisy_loaders.append(train_noisy_loader)
valid_noisy_loader = DataLoader(self.valid_noisy_dataset,
batch_size=self.config['valid_batch_size'],
num_workers=self.config['num_workers'],
shuffle=False,
sampler=SubsetRandomSampler(valid_task_data_list[i]))
valid_noisy_loaders.append(valid_noisy_loader)
self.train_loaders = train_noisy_loaders
self.valid_loaders = valid_noisy_loaders
self.create_chosen_dataloaders()
def create_semi(self):
self.train_semi_dataset, self.valid_semi_dataset, self.semi_dataset = extract_semi_data(self.train_dataset,
self.valid_dataset,
anomaly_num=self.config['semi_anomaly_num'],
anomaly_overlap=self.config['semi_overlap'])
train_task_data_list = self.train_semi_dataset.sample_indices_in_task
valid_task_data_list = self.valid_semi_dataset.sample_indices_in_task
train_semi_loaders, valid_semi_loaders = [], []
for i in range(self.config['num_task']):
train_semi_loader = DataLoader(self.train_semi_dataset,
batch_size=self.config['train_batch_size'],
num_workers=self.config['num_workers'],
sampler=SubsetRandomSampler(train_task_data_list[i]))
train_semi_loaders.append(train_semi_loader)
valid_semi_loader = DataLoader(self.valid_semi_dataset,
batch_size=self.config['valid_batch_size'],
num_workers=self.config['num_workers'],
shuffle=False,
sampler=SubsetRandomSampler(valid_task_data_list[i]))
valid_semi_loaders.append(valid_semi_loader)
self.train_loaders = train_semi_loaders
self.valid_loaders = valid_semi_loaders
self.create_chosen_dataloaders()
def create_continual(self):
self.create_chosen_dataloaders()
def create_transfer(self):
self.train_transfer_dataset = extract_fewshot_data(self.train_dataset, self.config['transfer_target_sample_num'])
train_transfer_task_data_list = self.train_transfer_dataset.sample_indices_in_task
self.train_transfer_loaders = []
for i in range(self.config['num_task']):
train_transfer_loader = DataLoader(self.train_transfer_dataset,
batch_size=self.config['train_batch_size'],
num_workers=self.config['num_workers'],
sampler=SubsetRandomSampler(train_transfer_task_data_list[i]),
drop_last=False)
self.train_transfer_loaders.append(train_transfer_loader)
self.create_chosen_transfer_dataloaders()
def create_chosen_dataloaders(self):
if self.config['model'] == 'dra':
for idx in self.config['train_task_id']:
self.chosen_train_loaders.append([self.train_loaders[idx], self.refer_loaders[idx]])
for idx in self.config['valid_task_id']:
self.chosen_valid_loaders.append([self.valid_loaders[idx], self.refer_loaders[idx]])
self.chosen_vis_loaders.append([self.vis_loaders[idx], self.refer_loaders[idx]])
else:
for idx in self.config['train_task_id']:
self.chosen_train_loaders.append(self.train_loaders[idx])
for idx in self.config['valid_task_id']:
self.chosen_valid_loaders.append(self.valid_loaders[idx])
self.chosen_vis_loaders.append(self.vis_loaders[idx])
def create_chosen_transfer_dataloaders(self):
if self.config['model'] == 'dra':
for idx in self.config['train_task_id']: # for step 1, train source task
self.chosen_train_loaders.append([self.train_loaders[idx], self.refer_loaders[idx]])
self.chosen_valid_loaders.append([self.valid_loaders[idx], self.refer_loaders[idx]])
self.chosen_vis_loaders.append([self.vis_loaders[idx], self.refer_loaders[idx]])
for idx in self.config['valid_task_id']: # for step 2, train target task
self.chosen_transfer_train_loaders.append([self.train_transfer_loaders[idx], self.refer_loaders[idx]])
self.chosen_transfer_valid_loaders.append([self.valid_loaders[idx], self.refer_loaders[idx]])
self.chosen_transfer_vis_loaders.append([self.vis_loaders[idx], self.refer_loaders[idx]])
else:
for idx in self.config['train_task_id']: # for step 1, train source task
self.chosen_train_loaders.append(self.train_loaders[idx])
self.chosen_valid_loaders.append(self.valid_loaders[idx])
self.chosen_vis_loaders.append(self.vis_loaders[idx])
for idx in self.config['valid_task_id']: # for step 2, train target task
self.chosen_transfer_train_loaders.append(self.train_transfer_loaders[idx])
self.chosen_transfer_valid_loaders.append(self.valid_loaders[idx])
self.chosen_transfer_vis_loaders.append(self.vis_loaders[idx])