-
Notifications
You must be signed in to change notification settings - Fork 7
/
datasets.py
296 lines (244 loc) · 12.9 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
import numpy as np
import torch
import torchvision
from torch.utils.data import Dataset
from torchvision import transforms
import copy
# Subset function
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
# Main data loader
class Data(object):
def __init__(self, args):
self.args = args
node_num = args.node_num
if args.dataset == 'cifar10':
# Data enhancement: None
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/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)
if args.dirichlet_alpha2:
groups, proportion = build_non_iid_by_dirichlet_hybrid(random_state=random_state, dataset=self.train_set, non_iid_alpha1=args.dirichlet_alpha,non_iid_alpha2=args.dirichlet_alpha2 ,num_classes=10, num_indices=num_indices, n_workers=node_num)
elif args.longtail_clients != 'none':
groups, proportion = build_non_iid_by_dirichlet_LT(random_state=random_state, dataset=self.train_set, lt_rho=args.longtail_clients, non_iid_alpha=args.dirichlet_alpha, num_classes=10, num_indices=num_indices, n_workers=node_num)
else:
groups, proportion = build_non_iid_by_dirichlet_new(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/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.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/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)
if args.dirichlet_alpha2:
groups, proportion = build_non_iid_by_dirichlet_hybrid(random_state=random_state, dataset=self.train_set, non_iid_alpha1=args.dirichlet_alpha,non_iid_alpha2=args.dirichlet_alpha2 ,num_classes=100, num_indices=num_indices, n_workers=node_num)
else:
groups, proportion = build_non_iid_by_dirichlet_new(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/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 == 'fmnist':
# Data enhancement
tra_transformer = transforms.Compose(
[
transforms.ToTensor(),
]
)
val_transformer = transforms.Compose(
[
transforms.ToTensor()
]
)
self.train_set = torchvision.datasets.FashionMNIST(
root="/home/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)
if args.dirichlet_alpha2:
groups, proportion = build_non_iid_by_dirichlet_hybrid(random_state=random_state, dataset=self.train_set, non_iid_alpha1=args.dirichlet_alpha,non_iid_alpha2=args.dirichlet_alpha2 ,num_classes=100, num_indices=num_indices, n_workers=node_num)
else:
groups, proportion = build_non_iid_by_dirichlet_new(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/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))])
### Dirichlet noniid functions ###
def build_non_iid_by_dirichlet_hybrid(
random_state = np.random.RandomState(0), dataset = 0, non_iid_alpha1 = 10, non_iid_alpha2 = 1, 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])
partition = random_state.dirichlet(np.repeat(non_iid_alpha1, n_workers), num_classes).transpose()
partition2 = random_state.dirichlet(np.repeat(non_iid_alpha2, n_workers/2), num_classes).transpose()
new_partition1 = copy.deepcopy(partition[:int(n_workers/2)])
sum_distr1 = np.sum(new_partition1, axis=0)
diag_mat = np.diag(1 - sum_distr1)
new_partition2 = np.dot(diag_mat, partition2.T).T
client_partition = np.vstack((new_partition1, new_partition2))
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
def build_non_iid_by_dirichlet_new(
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
def build_non_iid_by_dirichlet_LT(
random_state = np.random.RandomState(0), dataset = 0, lt_rho = 10.0, non_iid_alpha = 10, num_classes = 10, num_indices = 60000, n_workers = 10
):
# generate indicesbyclass list
indicesbyclass = {}
for i in range(num_classes):
indicesbyclass[i] = []
for idx, target in enumerate(dataset.targets):
indicesbyclass[int(target)].append(idx)
# calculate the image per class for LT
# reformulate the indicesbyclass according to the image per class
imb_factor = 1/float(lt_rho)
for _classes_idx in range(num_classes):
num = int(len(indicesbyclass[_classes_idx]) * (imb_factor**(_classes_idx / (num_classes - 1.0))))
random_state.shuffle(indicesbyclass[_classes_idx])
indicesbyclass[_classes_idx] = indicesbyclass[_classes_idx][:num]
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