-
Notifications
You must be signed in to change notification settings - Fork 2
/
datasets.py
360 lines (312 loc) · 14.7 KB
/
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
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
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
import numpy as np
import torch
import torchvision
from torch.utils.data import Dataset
from torchvision import transforms
import math
import random
import copy
import functools
from PIL import Image
import os
import glob
# from sample_dirichlet import clients_indices
class TensorDataset(Dataset):
def __init__(self, images, labels):
self.images = images.detach().float()
self.labels = labels.detach()
def __getitem__(self, index):
return self.images[index], self.labels[index]
def __len__(self):
return self.images.shape[0]
class DatasetSplit(Dataset):
def __init__(self, dataset, idxs):
self.dataset = dataset
self.idxs = list(idxs)
def __len__(self):
return len(self.idxs)
def __getitem__(self, item):
image, label = self.dataset[self.idxs[item]]
return image, label
class CustomSubset(torch.utils.data.Subset):
'''A custom subset class'''
def __init__(self, dataset, indices):
super().__init__(dataset, indices)
dataset.targets = torch.tensor(dataset.targets)
# print(dataset.targets)
self.targets = dataset.targets[indices]
# print(len(self.targets))
self.classes = dataset.classes
self.indices = indices
def __getitem__(self, idx):
x, y = self.dataset[self.indices[idx]]
return x, y
def __len__(self):
return len(self.indices)
EXTENSION = 'JPEG'
NUM_IMAGES_PER_CLASS = 500
CLASS_LIST_FILE = 'wnids.txt'
VAL_ANNOTATION_FILE = 'val_annotations.txt'
class TinyImageNet(Dataset):
"""
Ref: https://github.com/leemengtaiwan/tiny-imagenet/blob/master/TinyImageNet.py
Tiny ImageNet data set available from `http://cs231n.stanford.edu/tiny-imagenet-200.zip`.
Parameters
----------
root: string
Root directory including `train`, `test` and `val` subdirectories.
split: string
Indicating which split to return as a data set.
Valid option: [`train`, `test`, `val`]
transform: torchvision.transforms
A (series) of valid transformation(s).
in_memory: bool
Set to True if there is enough memory (about 5G) and want to minimize disk IO overhead.
"""
def __init__(self, root, split='train', transform=None, target_transform=None, in_memory=False):
self.root = os.path.expanduser(root)
self.split = split
self.transform = transform
self.target_transform = target_transform
self.in_memory = in_memory
self.split_dir = os.path.join(self.root, self.split)
self.image_paths = sorted(glob.iglob(os.path.join(self.split_dir, '**', '*.%s' % EXTENSION), recursive=True))
self.labels = {} # fname - label number mapping
self.images = [] # used for in-memory processing
# build class label - number mapping
with open(os.path.join(self.root, CLASS_LIST_FILE), 'r') as fp:
self.label_texts = sorted([text.strip() for text in fp.readlines()])
self.label_text_to_number = {text: i for i, text in enumerate(self.label_texts)}
if self.split == 'train':
for label_text, i in self.label_text_to_number.items():
for cnt in range(NUM_IMAGES_PER_CLASS):
self.labels['%s_%d.%s' % (label_text, cnt, EXTENSION)] = i
elif self.split == 'val':
with open(os.path.join(self.split_dir, VAL_ANNOTATION_FILE), 'r') as fp:
for line in fp.readlines():
terms = line.split('\t')
file_name, label_text = terms[0], terms[1]
self.labels[file_name] = self.label_text_to_number[label_text]
# get targets
self.targets = []
for index in range(len(self.image_paths)):
file_path = self.image_paths[index]
label_numeral = self.labels[os.path.basename(file_path)]
self.targets.append(label_numeral)
# read all images into torch tensor in memory to minimize disk IO overhead
if self.in_memory:
self.images = [self.read_image(path) for path in self.image_paths]
def __len__(self):
return len(self.image_paths)
def __getitem__(self, index):
file_path = self.image_paths[index]
if self.in_memory:
img = self.images[index]
else:
img = self.read_image(file_path)
if self.split == 'test':
return img
else:
# file_name = file_path.split('/')[-1]
return img, self.labels[os.path.basename(file_path)]
def __repr__(self):
fmt_str = 'Dataset ' + self.__class__.__name__ + '\n'
fmt_str += ' Number of datapoints: {}\n'.format(self.__len__())
tmp = self.split
fmt_str += ' Split: {}\n'.format(tmp)
fmt_str += ' Root Location: {}\n'.format(self.root)
tmp = ' Transforms (if any): '
fmt_str += '{0}{1}\n'.format(tmp, self.transform.__repr__().replace('\n', '\n' + ' ' * len(tmp)))
tmp = ' Target Transforms (if any): '
fmt_str += '{0}{1}'.format(tmp, self.target_transform.__repr__().replace('\n', '\n' + ' ' * len(tmp)))
return fmt_str
def read_image(self, path):
img = Image.open(path)
img = img.convert('RGB')
return self.transform(img) if self.transform else img
class Data(object):
# self.trainset, self.testset
def __init__(self, args):
self.args = args
node_num = args.node_num
if args.dataset == 'cifar10':
# Data enhancement
# tra_transformer = transforms.Compose(
# [
# transforms.RandomCrop(32, padding=4),
# transforms.RandomHorizontalFlip(),
# transforms.ToTensor(),
# transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
# ]
# )
# val_transformer = transforms.Compose(
# [
# transforms.ToTensor(),
# transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
# ]
# )
tra_transformer = transforms.Compose(
[
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
]
)
val_transformer = transforms.Compose(
[
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
]
)
self.train_set = torchvision.datasets.CIFAR10(
root="/home/lzx/Dataset/cifar/", train=True, download=False, transform=tra_transformer
)
if args.iid == 0: # noniid
random_state = np.random.RandomState(int(args.random_seed))
num_indices = len(self.train_set)
groups, proportion = build_non_iid_by_dirichlet(random_state=random_state, dataset=self.train_set, non_iid_alpha=args.dirichlet_alpha, num_classes=10, num_indices=num_indices, n_workers=node_num)
self.train_loader = groups
self.groups = groups
self.proportion = proportion
else:
data_num = [int(50000/node_num) for _ in range(node_num)]
splited_set = torch.utils.data.random_split(self.train_set, data_num)
self.train_loader = splited_set
self.test_set = torchvision.datasets.CIFAR10(
root="/home/lzx/Dataset/cifar/", train=False, download=False, transform=val_transformer
)
self.test_loader = torch.utils.data.random_split(self.test_set, [int(len(self.test_set))])
elif args.dataset == 'cifar100':
# Data enhancement
# tra_transformer = transforms.Compose(
# [
# transforms.RandomCrop(32, padding=4),
# transforms.RandomHorizontalFlip(),
# transforms.ToTensor(),
# transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
# ]
# )
# val_transformer = transforms.Compose(
# [
# transforms.ToTensor(),
# transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
# ]
# )
tra_transformer = transforms.Compose(
[
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
]
)
val_transformer = transforms.Compose(
[
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
]
)
self.train_set = torchvision.datasets.CIFAR100(
root="/home/lzx/Dataset/cifar/", train=True, download=False, transform=tra_transformer
)
if args.iid == 0: # noniid
random_state = np.random.RandomState(int(args.random_seed))
num_indices = len(self.train_set)
groups, proportion = build_non_iid_by_dirichlet(random_state=random_state, dataset=self.train_set, non_iid_alpha=args.dirichlet_alpha, num_classes=100, num_indices=num_indices, n_workers=node_num)
self.train_loader = groups
self.groups = groups
self.proportion = proportion
else:
data_num = [int(50000/node_num) for _ in range(node_num)]
splited_set = torch.utils.data.random_split(self.train_set, data_num)
self.train_loader = splited_set
self.test_set = torchvision.datasets.CIFAR100(
root="/home/lzx/Dataset/cifar/", train=False, download=False, transform=val_transformer
)
self.test_loader = torch.utils.data.random_split(self.test_set, [int(len(self.test_set))])
elif args.dataset == 'tinyimagenet':
# Data enhancement
tra_transformer = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
val_transformer = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
self.train_set = TinyImageNet("/home/lzx/data/tiny-imagenet-200/", 'train', tra_transformer, in_memory=False)
if args.iid == 0: # noniid
random_state = np.random.RandomState(int(args.random_seed))
num_indices = len(self.train_set)
groups, proportion = build_non_iid_by_dirichlet(random_state=random_state, dataset=self.train_set, non_iid_alpha=args.dirichlet_alpha, num_classes=200, num_indices=num_indices, n_workers=node_num)
self.train_loader = groups
self.groups = groups
self.proportion = proportion
else:
data_num = [int(100000/node_num) for _ in range(node_num)]
splited_set = torch.utils.data.random_split(self.train_set, data_num)
self.train_loader = splited_set
self.test_set = TinyImageNet("/home/lzx/data/tiny-imagenet-200/", 'val', val_transformer, in_memory=False)
self.test_loader = torch.utils.data.random_split(self.test_set, [int(len(self.test_set))])
elif args.dataset == 'fmnist':
# Data enhancement
tra_transformer = transforms.Compose(
[
transforms.ToTensor(),
]
)
val_transformer = transforms.Compose(
[
transforms.ToTensor()
]
)
self.train_set = torchvision.datasets.FashionMNIST(
root="/home/lzx/Dataset/FashionMNIST", train=True, download=False, transform=tra_transformer
)
if args.iid == 0: # noniid
random_state = np.random.RandomState(int(args.random_seed))
num_indices = len(self.train_set)
groups, proportion = build_non_iid_by_dirichlet(random_state=random_state, dataset=self.train_set, non_iid_alpha=args.dirichlet_alpha, num_classes=100, num_indices=num_indices, n_workers=node_num)
self.train_loader = groups
self.groups = groups
self.proportion = proportion
else:
data_num = [int(60000/node_num) for _ in range(node_num)]
splited_set = torch.utils.data.random_split(self.train_set, data_num)
self.train_loader = splited_set
self.test_set = torchvision.datasets.FashionMNIST(
root="/home/lzx/Dataset/FashionMNIST", train=False, download=False, transform=val_transformer
)
self.test_loader = torch.utils.data.random_split(self.test_set, [int(len(self.test_set))])
def build_non_iid_by_dirichlet(
random_state = np.random.RandomState(0), dataset = 0, non_iid_alpha = 10, num_classes = 10, num_indices = 60000, n_workers = 10
):
indicesbyclass = {}
for i in range(num_classes):
indicesbyclass[i] = []
for idx, target in enumerate(dataset.targets):
indicesbyclass[int(target)].append(idx)
for i in range(num_classes):
random_state.shuffle(indicesbyclass[i])
client_partition = random_state.dirichlet(np.repeat(non_iid_alpha, n_workers), num_classes).transpose()
for i in range(len(client_partition)):
for j in range(len(client_partition[i])):
client_partition[i][j] = int(round(client_partition[i][j]*len(indicesbyclass[j])))
client_partition_index = copy.deepcopy(client_partition)
for i in range(len(client_partition)):
for j in range(len(client_partition[i])):
if i == 0:
client_partition_index[i][j] = client_partition_index[i][j]
elif i == len(client_partition) - 1:
client_partition_index[i][j] = len(indicesbyclass[j])
else:
client_partition_index[i][j] = client_partition_index[i-1][j] + client_partition_index[i][j]
dict_users = {}
for i in range(n_workers):
dict_users[i] = []
for i in range(len(client_partition)):
for j in range(len(client_partition[i])):
if i == 0:
dict_users[i].extend(indicesbyclass[j][:int(client_partition_index[i][j])])
else:
dict_users[i].extend(indicesbyclass[j][int(client_partition_index[i-1][j]) : int(client_partition_index[i][j])])
for i in range(len(dict_users)):
random_state.shuffle(dict_users[i])
return dict_users, client_partition