forked from i-deal/MLR-2.0
-
Notifications
You must be signed in to change notification settings - Fork 0
/
dataset_builder.py
executable file
·360 lines (301 loc) · 13.3 KB
/
dataset_builder.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
# prerequisites
import torch
import os
import numpy as np
from torchvision import datasets
from torchvision import transforms as torch_transforms
from torch.utils import data #.data import #DataLoader, Subset, Dataset
import random
from PIL import Image, ImageOps, ImageEnhance, __version__ as PILLOW_VERSION
colornames = ["red", "green", "blue", "purple", "yellow", "cyan", "orange", "brown", "pink", "white"]
colorrange = .08
colorvals = [
[1 - colorrange, colorrange * 1, colorrange * 1],
[colorrange * 1, 1 - colorrange, colorrange * 1],
[colorrange * 2, colorrange * 2, 1 - colorrange],
[1 - colorrange * 2, colorrange * 2, 1 - colorrange * 2],
[1 - colorrange, 1 - colorrange, colorrange * 2],
[colorrange, 1 - colorrange, 1 - colorrange],
[1 - colorrange, .5, colorrange * 2],
[.6, .4, .2],
[1 - colorrange, 1 - colorrange * 3, 1 - colorrange * 3],
[1-colorrange,1-colorrange,1-colorrange]
]
class Colorize_specific:
def __init__(self, col):
self.col = col
def __call__(self, img):
# col: an int index for which base color is being used
rgb = colorvals[self.col] # grab the rgb for this base color
r_color = rgb[0] + np.random.uniform() * colorrange * 2 - colorrange # generate a color randomly in the neighborhood of the base color
g_color = rgb[1] + np.random.uniform() * colorrange * 2 - colorrange
b_color = rgb[2] + np.random.uniform() * colorrange * 2 - colorrange
np_img = np.array(img, dtype=np.uint8)
np_img = np.dstack([np_img * r_color, np_img * g_color, np_img * b_color])
np_img = np_img.astype(np.uint8)
img = Image.fromarray(np_img, 'RGB')
return img
class No_Color_3dim:
def __init__(self):
self.x = None
def __call__(self, img):
np_img = np.array(img, dtype=np.uint8)
np_img = np.dstack([np_img, np_img, np_img])
np_img = np_img.astype(np.uint8)
img = Image.fromarray(np_img, 'RGB')
return img
class Translate:
def __init__(self, scale, loc, max_width, min_width = 28):
self.max_width = max_width
self.min_width = min_width
self.max_scale = max_width//2
self.pos = torch.zeros(2, max_width).cuda()
self.loc = loc
self.scale = scale
def __call__(self, img):
if self.scale == 0:
scale_val = (random.random()*4)
scale_dist = torch.zeros(10)
scale_dist[int(scale_val)] = 1
width = int(self.min_width + (self.max_width - self.min_width) * (scale_val / 10))
height = int(self.min_width + (self.max_width - self.min_width) * (scale_val/ 10))
resize = torch_transforms.Resize((width, height))
img = resize(img)
elif self.scale == 1:
scale_val = (random.random()*4) +4
scale_dist = torch.zeros(10)
scale_dist[int(scale_val)] = 1
width = int(self.min_width + (self.max_width - self.min_width) * (scale_val / 10))
height = int(self.min_width + (self.max_width - self.min_width) * (scale_val/ 10))
resize = torch_transforms.Resize((width, height))
img = resize(img)
else:
scale_dist = None
if self.loc == 1:
padding_left = int(random.uniform(0, (self.max_width // 2)-(img.size[0]//2))) #include center overlap region +
padding_right = self.max_width - img.size[0] - padding_left
padding_bottom = random.randint(0, self.max_width - img.size[0])
padding_top = self.max_width - img.size[0] - padding_bottom
elif self.loc == 2:
if img.size[0] >= self.max_width//2:
x = img.size[0]//2
else:
x = 0
padding_left = int(random.uniform((self.max_width // 2)-x, self.max_width - img.size[0])) #include center overlap region
padding_right = self.max_width - img.size[0] - padding_left
padding_bottom = random.randint(0, self.max_width - img.size[0])
padding_top = self.max_width - img.size[0] - padding_bottom
padding = (padding_left, padding_top, padding_right, padding_bottom)
pos = self.pos.clone()
pos[0][padding_left] = 1
pos[1][padding_bottom] = 1
#print(padding_left,padding_bottom)
return ImageOps.expand(img, padding), pos, scale_dist
class PadAndPosition:
def __init__(self, transform):
self.transform = transform
self.scale = transform.scale
def __call__(self, img):
new_img, position, scale_dist = self.transform(img)
if self.scale != -1:
return torch_transforms.ToTensor()(new_img), torch_transforms.ToTensor()(img), position, scale_dist #retinal, crop, position, scale
else:
return torch_transforms.ToTensor()(new_img), torch_transforms.ToTensor()(img), position
class ToTensor:
def __init__(self):
self.x = None
def __call__(self, img):
return torch_transforms.ToTensor()(img)
class Dataset(data.Dataset):
def __init__(self, dataset, transforms={}, train=True):
# initialize base dataset
if type(dataset) == str:
self.name = dataset
self.train = train
self.dataset = self._build_dataset(dataset, train)
else:
raise ValueError('invalid dataset input type')
# initialize retina
if 'retina' in transforms:
self.retina = transforms['retina']
if self.retina == True:
if 'retina_size' in transforms:
self.retina_size = transforms['retina_size']
else:
self.retina_size = 64
if 'location_targets' in transforms:
self.right_targets = transforms['location_targets']['right']
self.left_targets = transforms['location_targets']['left']
else:
self.right_targets = []
self.left_targets = []
else:
self.retina_size = None
self.right_targets = []
self.left_targets = []
else:
self.retina = False
self.retina_size = None
self.right_targets = []
self.left_targets = []
# initialize colors
if 'colorize' in transforms:
self.colorize = transforms['colorize']
self.color_dict = {}
if self.colorize == True and 'color_targets' in transforms:
self.color_dict = {}
colors = {}
for color in transforms['color_targets']:
for target in transforms['color_targets'][color]:
colors[target] = color
self.color_dict = colors
else:
self.colorize = False
self.color_dict = {}
# initialize scaling
if 'scale' in transforms:
self.scale = transforms['scale']
if self.scale == True and 'scale_targets' in transforms:
self.scale_dict = {}
for scale in transforms['scale_targets']:
for target in transforms['scale_targets'][scale]:
self.scale_dict[target] = scale
else:
self.scale = False
# initialize skip connection
if 'skip' in transforms:
self.skip = transforms['skip']
if self.skip == True:
self.colorize = True
self.retina = False
else:
self.skip = False
self.no_color_3dim = No_Color_3dim()
self.totensor = ToTensor()
self.target_dict = {'mnist':[0,9], 'emnist':[10,35], 'fashion_mnist':[36,45], 'cifar10':[46,55]}
if dataset == 'emnist':
self.lowercase = list(range(0,10)) + list(range(36,63))
if os.path.exists('uppercase_ind_train.pt'):
if self.train == True:
self.indices = torch.load('uppercase_ind_train.pt')
else:
self.indices = torch.load('uppercase_ind_test.pt')
else:
print('indexing emnist dataset:')
self.indicies, self.indices = self._filter_indices()
print('indexing complete')
def _filter_indices(self):
base_dataset = datasets.EMNIST(root='./data', split='byclass', train=False, transform=torch_transforms.Compose([lambda img: torch_transforms.functional.rotate(img, -90),
lambda img: torch_transforms.functional.hflip(img)]), download=True)
indices_test = []
count = {target: 0 for target in list(range(10,36))}
print('starting indices collection')
for i in range(len(base_dataset)):
img, target = base_dataset[i]
if target not in self.lowercase and count[target] <= 6000:
indices_test += [i]
count[target] += 1
print(count)
#torch.save(indices_train, 'uppercase_ind_train.pt')
torch.save(indices_test, 'uppercase_ind_test.pt')
print('saved indices')
indices_train = torch.load('uppercase_ind_train.pt')
return indices_train, indices_test
def _build_dataset(self, dataset, train=True):
if dataset == 'mnist':
base_dataset = datasets.MNIST(root='./mnist_data/', train=train, transform = None, download=True)
elif dataset == 'emnist':
split = 'byclass'
# raw emnist dataset is rotated and flipped by default, the applied transforms undo that
base_dataset = datasets.EMNIST(root='./data', split=split, train=train, transform=torch_transforms.Compose([lambda img: torch_transforms.functional.rotate(img, -90),
lambda img: torch_transforms.functional.hflip(img)]), download=True)
elif dataset == 'fashion_mnist':
base_dataset = datasets.FashionMNIST('./fashionmnist_data/', train=train, transform = None, download=True)
elif dataset== 'cifar10':
base_dataset = datasets.CIFAR10(root='./data', train=train, download=True, transform=None)
elif os.path.exists(dataset):
pass
else:
raise ValueError(f'{dataset} is not a valid base dataset')
return base_dataset
def __len__(self):
return len(self.dataset)
def __getitem__(self, index):
image, target = self.dataset[index]
if self.name == 'emnist' and self.train == True:
image, target = self.dataset[self.indices[random.randint(0,len(self.indices)-1)]]
else:
target += self.target_dict[self.name][0]
col = None
transform_list = []
# append transforms according to transform attributes
# color
if self.colorize == True:
if target in self.color_dict:
col = self.color_dict[target]
transform_list += [Colorize_specific(col)]
else:
col = random.randint(0,9) # any
transform_list += [Colorize_specific(col)]
else:
col = -1
transform_list += [self.no_color_3dim]
# skip connection dataset
if self.skip == True:
transform_list += [torch_transforms.RandomRotation(90), torch_transforms.RandomCrop(size=28, padding= 8)]
# retina
if self.retina == True:
if self.scale == True:
if target in self.scale_dict:
scale = self.scale_dict[target]
else:
scale = random.randint(0,1)
else:
scale = -1
if target in self.left_targets:
translation = 1 # left
elif target in self.right_targets:
translation = 2 # right
else:
translation = random.randint(1,2) #any
translate = PadAndPosition(Translate(scale, translation, self.retina_size))
transform_list += [translate]
else:
scale = -1
translation = -1
transform_list += [self.totensor]
# labels
out_label = (target, col, translation, scale)
transform = torch_transforms.Compose(transform_list)
return transform(image), out_label
def get_loader(self, batch_size):
loader = torch.utils.data.DataLoader(dataset=self, batch_size=batch_size, shuffle=True, drop_last=True)
return loader
def all_possible_labels(self):
# return a list of all possible labels generated by this dataset in order: (shape identity, color, retina location)
dataset = self.name
start = self.target_dict[dataset][0]
end = self.target_dict[dataset][1] + 1
target_dict = {}
for i in range(start,end):
if self.colorize == True:
if i in self.color_dict:
col = [self.color_dict[i]]
else:
col = [0,9]
else:
col = [-1]
# retina
if self.retina == True:
if i in self.left_targets:
translation = [1]
elif i in self.right_targets:
translation = [2]
else:
translation = [1,2]
else:
translation = [-1]
# labels
target = [col, translation]
target_dict[i] = target
return target_dict