-
Notifications
You must be signed in to change notification settings - Fork 0
/
datasets.py
330 lines (272 loc) · 9.86 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
'''
Datasets
'''
import random
import os
from math import comb
from pathlib import Path
import numpy as np
from PIL import Image
import torch
from torchvision import datasets
from torch.utils.data import Dataset
celeba_attrs = {
'0':'5_o_Clock_Shadow',
'1':'Arched_Eyebrows',
'2':'Attractive',
'3':'Bags_Under_Eyes',
'4':'Bald',
'5':'Bangs',
'6':'Big_Lips',
'7':'Big_Nose',
'8':'Black_Hair',
'9':'Blond_Hair',
'10':'Blurry',
'11':'Brown_Hair',
'12':'Bushy_Eyebrows',
'13':'Chubby',
'14':'Double_Chin',
'15':'Eyeglasses',
'16':'Goatee',
'17':'Gray_Hair',
'18':'Heavy_Makeup',
'19':'High_Cheekbones',
'20':'Male',
'21':'Mouth_Slightly_Open',
'22':'Mustache',
'23':'Narrow_Eyes',
'24':'No_Beard',
'25':'Oval_Face',
'26':'Pale_Skin',
'27':'Pointy_Nose',
'28':'Receding_Hairline',
'29':'Rosy_Cheeks',
'30':'Sideburns',
'31':'Smiling',
'32':'Straight_Hair',
'33':'Wavy_Hair',
'34':'Wearing_Earrings',
'35':'Wearing_Hat',
'36':'Wearing_Lipstick',
'37':'Wearing_Necklace',
'38':'Wearing_Necktie',
'39':'Young'
}
def get_splits(t, margin_cands, k, nclasses, nsubclasses, classes, subclasses):
assert k >= 2
assert (nclasses == 1 or nsubclasses == 1) and nclasses * nsubclasses == k
margins = random.sample(margin_cands, k - 1)
margins.append(t)
margins.insert(0, 0) # first parameter is index
seg = []
for i in range(len(margins) - 1):
seg.append(margins[i + 1] - margins[i])
classes = random.sample(classes, nclasses)
subclasses = random.sample(subclasses, nsubclasses)
if nclasses == 1:
classes = classes * k
elif nsubclasses == 1:
subclasses = subclasses * k
return [seg, classes, subclasses]
def cumulatize(splits):
n, _, k = splits.shape
result = [{} for i in range(n)]
snapshot = {}
for i in range(n):
result[i] = snapshot.copy()
for j, class_key in enumerate(splits[i][1:].transpose()):
class_key = tuple(class_key)
snapshot[class_key] = snapshot.get(class_key, 0) + splits[i][0][j]
return result
def discretize(seg, t):
k = 0
while t >= seg[k]:
t -= seg[k]
k += 1
return k
def absolute(seg):
result = [0]*(len(seg)+1)
for i in range(1, len(result)):
result[i] = result[i-1] + seg[i-1]
return result
# time-series datasets
class TS(Dataset):
def __init__(self, datapath, dataset, split, n_max, t_max, classes, transform):
is_train = split == 'train'
if dataset == 'mnist':
self.dataset = datasets.MNIST(
root=datapath,
download=True,
train=is_train,
transform=transform)
elif dataset == 'cifar10':
self.dataset = datasets.CIFAR10(
root=datapath,
download=True,
train=is_train,
transform=transform)
elif dataset == 'cifar100':
self.dataset = datasets.CIFAR100(
root=datapath,
download=True,
train=is_train,
transform=transform)
elif dataset == 'celeba':
self.dataset = datasets.CelebA(
root=datapath,
download=True,
split=split,
transform=transform)
else:
raise Exception("Dataset not found")
self.dims = self.dataset[0][0].size()
self.N = n_max
self.T = t_max
min_t = 2
assert self.T > 2*min_t
margin_cands = list(range(min_t, self.T - min_t))
if dataset != 'celeba':
indices = np.arange(len(self.dataset))
targets = np.array(self.dataset.targets)
indices_by_c = {}
for c in classes:
indices_by_c[(c, 0)] = indices[targets == c]
self.indices_by_c = indices_by_c
self.splits = np.asarray(
[get_splits(self.T, margin_cands, 2, 2, 1, classes, [0])
for _ in range(self.N)])
self.splits_cum = cumulatize(self.splits)
else:
subclasses = np.unique(self.dataset.attr)
indices_by_c = {}
for subc in subclasses:
hits = np.where(self.dataset.attr == subc)
for c in classes:
indices_by_c[(c, subc)] = hits[0][hits[1] == c]
for c_key in indices_by_c:
random.shuffle(indices_by_c[c_key])
self.indices_by_c = indices_by_c
self.splits = np.asarray(
[get_splits(self.T, margin_cands, 2, 1, 2, classes, [0, 1])
for _ in range(self.N)])
self.splits_cum = cumulatize(self.splits)
def __len__(self):
return self.N * self.T
def __getitem__(self, idx):
n = idx // self.T
t = idx % self.T
N, _, K = self.splits.shape
seg = self.splits[n][0]
k = discretize(seg, t)
c_key = tuple(self.splits[n][1:].transpose()[k])
if c_key not in self.splits_cum[n]:
index = t
else:
index = self.splits_cum[n][c_key] + t
if index >= len(self.indices_by_c[c_key]):
random.shuffle(self.indices_by_c[c_key])
index = index % len(self.indices_by_c[c_key])
sample, _ = self.dataset[self.indices_by_c[c_key][index]]
label = n * K + k
return sample, label
def get_x_n(self, n):
x = torch.empty(size=(self.T,) + self.dims)
for t in range(self.T):
x[t], _ = TS.__getitem__(self, n * self.T + t)
return x
# contrastive dataset based on TS dataset
class CON(TS):
def __init__(self, p_max, **kwargs):
super().__init__(**kwargs)
self.P = p_max
N, _, K = self.splits.shape
self.K = K
def __len__(self):
return self.N * self.P
def __getitem__(self, idx):
n = idx // self.P
N, _, K = self.splits.shape
seg = self.splits[n][0] # segments
seg_cum = absolute(seg) # cumulative segments
npros = np.asarray([comb(s, 2) for s in seg])
npros = npros / np.sum(npros)
is_positive = idx % self.P <= self.P // 2
if is_positive:
k = random.choices(range(K), weights=npros, k=1)[0]
chosen_ts = random.sample(range(seg_cum[k], seg_cum[k+1]), K)
else:
chosen_ts = [random.sample(range(seg_cum[k], seg_cum[k+1]), 1)[0]
for k in range(0, len(seg_cum)-1)]
samples = torch.empty(size=(self.K,) + self.dims)
for k, t in enumerate(chosen_ts):
samples[k] = super().__getitem__(n * self.T + t)[0]
return samples, 1 - int(is_positive) # positive sample is labeled 0
class clevr_change_contrastive_explicit(Dataset):
def __init__(self, name, T, P, transform_configuration):
self.dir1 = path.join(dataset_dir, 'clevr', name, 'outputnsc_images/')
self.dir2 = path.join(dataset_dir, 'clevr', name, 'outputsc_images/')
self.transform = transform_configuration
all_filenames = os.listdir(self.dir1)
image_1 = Image.open(self.dir1 + all_filenames[0]).convert('RGB')
self.data_dim = self.transform(image_1).shape
cps = {}
for filename in all_filenames:
_, _, i, t = filename.split('_') # i-th image at time t
t = int(t[:-4])
i = int(i)
cps[i] = cps.get(i, 0) + 1
self.cps = cps
self.n, self.T, self.P = len(cps), T, P
def __len__(self):
return self.n * self.P
def __getitem__(self, item):
i, t = item // self.T, item % self.T
if t < self.cps[i]:
label = 2 * i
file_name = path.join(self.dir1, 'CLEVR_nonsemantic_' + str(i).zfill(6) + '_' + str(t) + '.png')
img = Image.open(file_name).convert('RGB')
else:
label = 2 * i + 1
file_name = path.join(self.dir2, 'CLEVR_semantic_' + str(i).zfill(6) + '_' + str(t - self.cps[i]) + '.png')
img = Image.open(file_name).convert('RGB')
return self.transform(img), label
def get_time_series_sample(self, n):
X = torch.empty(size=(self.T,) + self.data_dim)
for t in range(self.T):
X[t] = self.__getitem__(self.T * n + t)[0]
return X
class clevr_change(Dataset):
def __init__(self, name, T, transform_configuration):
self.dir1 = path.join(dataset_dir, 'clevr', name, 'outputnsc_images/')
self.dir2 = path.join(dataset_dir, 'clevr', name, 'outputsc_images/')
self.transform = transform_configuration
all_filenames = os.listdir(self.dir1)
image_1 = Image.open(self.dir1 + all_filenames[0]).convert('RGB')
self.data_dim = self.transform(image_1).shape
cps = {}
for filename in all_filenames:
_, _, i, t = filename.split('_') # i-th image at time t
t = int(t[:-4])
i = int(i)
cps[i] = cps.get(i, 0) + 1
self.cps = cps
self.n = len(cps)
self.T = T
def __len__(self):
return self.n * self.T
def __getitem__(self, item):
i, t = item // self.T, item % self.T
if t < self.cps[i]:
label = 2 * i
file_name = path.join(self.dir1, 'CLEVR_nonsemantic_' + str(i).zfill(6) + '_' + str(t) + '.png')
img = Image.open(file_name).convert('RGB')
else:
label = 2 * i + 1
file_name = path.join(self.dir2, 'CLEVR_semantic_' + str(i).zfill(6) + '_' + str(t - self.cps[i]) + '.png')
img = Image.open(file_name).convert('RGB')
return self.transform(img), label
def get_time_series_sample(self, n):
X = torch.empty(size=(self.T,) + self.data_dim)
for t in range(self.T):
X[t] = self.__getitem__(self.T * n + t)[0]
return X