-
Notifications
You must be signed in to change notification settings - Fork 8
/
datasets.py
280 lines (219 loc) · 12.4 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
# https://www.youtube.com/watch?v=DuSqffoDojM -> 1:29:03
# -> https://gist.github.com/y0ast/f69966e308e549f013a92dc66debeeb4
from typing import Any, Callable, Optional
import torch
from torchvision.datasets import MNIST, SVHN, FashionMNIST
import numpy as np
import h5py
import os
import cv2
class Fast_MNIST(MNIST):
"""
Source as modified from: https://gist.github.com/y0ast/f69966e308e549f013a92dc66debeeb4
"""
def __init__(self, *args, **kwargs):
# deleting device key from kwargs (as cannot be passed to MNIST mother class)
device = kwargs['device']
del kwargs['device']
super().__init__(*args, **kwargs)
# Insert channel dim and scale data to [0,1]
self.data = self.data.unsqueeze(1).float().div(255)
# binarize MNIST by sampling from Bernoulli with probs that are pixel intensities
# done as in Salakhutdinov & Murray (2008)
# https://dl.acm.org/doi/pdf/10.1145/1390156.1390266?casa_token=mvJb8zb-U1EAAAAA:fH7RMyQbJi5vGtrUySEmv_Raf19T4PxWRlyoWXoKi4uMBeJJtxLMjryTMJEe-mTgIRciqBVlBOA
# n_examples, in_dim = self.data.size()[0], (self.data.size()[2] * self.data.size()[3])
# self.data = torch.flatten(self.data)
# bernoulli = torch.distributions.Bernoulli(self.data)
# self.data = bernoulli.sample()
# self.data = self.data.view(n_examples, in_dim)
# Normalize it with the usual MNIST mean and std
# self.data = self.data.sub_(0.1307).div_(0.3081)
# Put both data and targets on GPU in advance
self.data, self.targets = self.data.to(device), self.targets.to(device)
def __getitem__(self, index):
"""
Args:
index (int): Index
Returns:
tuple: (image, target) where target is index of the target class.
"""
img, target = self.data[index], self.targets[index]
return img, target
class Fast_SVHN(SVHN):
"""
Source as modified from: https://gist.github.com/y0ast/f69966e308e549f013a92dc66debeeb4
"""
def __init__(self, *args, **kwargs):
# deleting device key from kwargs (as cannot be passed to SVHN mother class)
device = kwargs['device']
del kwargs['device']
super().__init__(*args, **kwargs)
# convert numpy to torch
self.data = torch.tensor(self.data)
self.labels = torch.tensor(self.labels)
# Scale data to [-.5, .5]
# self.data = self.data.float().div(255) - .5
# Scale data to [0., 1.]
self.data = self.data.float().div(255)
# Normalize it with the usual MNIST mean and std
# self.data = self.data.sub_(0.1307).div_(0.3081)
# Put both data and targets on GPU in advance
# attribute is self.labels, see https://pytorch.org/docs/stable/_modules/torchvision/datasets/svhn.html#SVHN
self.data, self.labels = self.data.to(device), self.labels.to(device)
def __getitem__(self, index):
"""
Args:
index (int): Index
Returns:
tuple: (image, target) where target is index of the target class.
"""
img, label = self.data[index], self.labels[index]
return img, label
class Fast_3DShapes(torch.utils.data.Dataset):
"""
Source as modified from: https://gist.github.com/y0ast/f69966e308e549f013a92dc66debeeb4
# Dataset from: https://github.com/deepmind/3d-shapes
# Using loading and sampling script from: https://github.com/deepmind/3d-shapes/blob/master/3dshapes_loading_example.ipynb
"""
def __init__(self, train, device,
train_frac=1.,
factors_variation_dict={'floor_hue': list(range(10)), 'wall_hue': list(range(10)), 'object_hue': list(range(10)), 'scale': list(range(8)), 'shape': list(range(4)), 'orientation': list(range(15))},
factors_label_list=['floor_hue', 'wall_hue', 'object_hue', 'scale', 'shape', 'orientation'],
seed=0,
data_folder=None):
# make the train-test split deterministic
torch.manual_seed(seed)
np.random.seed(seed)
# deleting device key from kwargs (as cannot be passed to SVHN mother class)
self.DATA_RAW_DIR = 'data/Fast_3DShapes/raw' if data_folder is None else os.path.join(data_folder, 'Fast_3DShapes')
self.DATA_PROC_DIR = 'data/Fast_3DShapes/processed'
self.RAW_FILE_NAME = '3dshapes.h5'
# check if dataset is available
if not os.path.exists(os.path.join(self.DATA_RAW_DIR, self.RAW_FILE_NAME)):
exit("3DShapes not downloaded yet! Download the file 3dshapes.h5 from https://console.cloud.google.com/storage/browser/3d-shapes;tab=objects?prefix=&forceOnObjectsSortingFiltering=false (see https://github.com/deepmind/3d-shapes for further instructions) and put this file into data/Fast_3DShapes/raw")
raw_file_path = os.path.join(self.DATA_RAW_DIR, self.RAW_FILE_NAME)
# from loading script
dataset = h5py.File(raw_file_path, 'r')
# print(dataset.keys())
data = dataset['images'] # array shape [480000,64,64,3], uint8 in range(256)
labels = dataset['labels'] # array shape [480000,6], float64
image_shape = data.shape[1:] # [64,64,3]
label_shape = labels.shape[1:] # [6]
n_samples = labels.shape[0] # 10*10*10*8*4*15=480000
_FACTOR_TO_COLUMN_INDEX = {'floor_hue': 0, 'wall_hue': 1, 'object_hue': 2, 'scale': 3, 'shape': 4, 'orientation': 5}
_FACTOR_TO_ALLOWED_VALUES = {'floor_hue': list(range(10)), 'wall_hue': list(range(10)), 'object_hue': list(range(10)), 'scale': list(range(8)), 'shape': list(range(4)), 'orientation': list(range(15))}
_NUM_VALUES_PER_FACTOR = {'floor_hue': 10, 'wall_hue': 10, 'object_hue': 10,
'scale': 8, 'shape': 4, 'orientation': 15}
# convert labels to numpy array
labels = labels[()]
labels = labels.astype(float)
# convert float to int scales
labels[:, 0] = labels[:, 0] * 10 # [0, 1, ..., 9]
labels[:, 1] = labels[:, 1] * 10 # [0, 1, ..., 9]
labels[:, 2] = labels[:, 2] * 10 # [0, 1, ..., 9]
# label in column index 3 is very weird -> manual remapping required
labels[:, 3] = np.round(labels[:, 3], 2) # first round, since values are very precise floats
remap = {0.75: 0.0, 0.82: 1.0, 0.89: 2.0, 0.96: 3.0, 1.04: 4.0, 1.11: 5.0, 1.18: 6.0, 1.25: 7.0}
label_3 = np.copy(labels[:, 3])
for k, v in remap.items():
label_3[labels[:, 3] == k] = v
labels[:, 3] = label_3
# shape is already on int scale # # [0, 1, ..., 3]
labels[:, 5] = np.round(labels[:, 5], 2) # first round, since values are very precise floats
remap = {-30.: 0, -25.71: 1, -21.43: 2, -17.14: 3, -12.86: 4, -8.57: 5, -4.29: 6, 0.: 7, 4.29: 8, 8.57: 9, 12.86: 10, 17.14: 11, 21.43: 12, 25.71: 13, 30.: 14}
label_5 = np.copy(labels[:, 5])
for k, v in remap.items():
label_5[labels[:, 5] == k] = v
labels[:, 5] = label_5 # [0, 1, ..., 15]
# make labels an int, because
# since 3 in labels[:, 0] is actually 3.0000000000004, even though not correctly displayed
labels = labels.astype(int)
# some asserts
assert _FACTOR_TO_COLUMN_INDEX.keys() == factors_variation_dict.keys()
assert all(x in _FACTOR_TO_COLUMN_INDEX.keys() for x in factors_label_list)
# assert that values of variation for each factor are correctly chosen
for (key, value) in factors_variation_dict.items():
assert all(x in _FACTOR_TO_ALLOWED_VALUES[key] for x in value)
# each label factor chosen must have at least 2 values in the corresponding factors_variation_dict
for factor in factors_label_list:
assert len(factors_variation_dict[factor]) > 1
# at least one label chosen
assert len(factors_label_list) > 0
# assert 'scale' not in factors_label_list # because there is an issue with it
# choose the data
chosen_conjunction = None
for i, (key, value) in enumerate(factors_variation_dict.items()):
chosen = np.in1d(labels[:, _FACTOR_TO_COLUMN_INDEX[key]], value)
if i == 0:
chosen_conjunction = chosen
else:
chosen_conjunction = np.logical_and(chosen_conjunction, chosen)
# chosen_conjunction = chosen_conjunction.flatten()
data = data[chosen_conjunction, :, :, :]
labels = labels[chosen_conjunction, :]
# choose the label
label_column_indices = []
for factor in factors_label_list:
label_column_indices.append(_FACTOR_TO_COLUMN_INDEX[factor])
labels = labels[:, label_column_indices]
# scale down the images from (64x64) to (32x32)
scaled_images = []
for i in range(data.shape[0]):
img = data[i]
img = cv2.resize(img, dsize=(32, 32)) # see https://docs.opencv.org/master/da/d6e/tutorial_py_geometric_transformations.html
scaled_images.append(img)
data = np.asarray(scaled_images)
# Scale data to [0., 1.]
data = data / 255
# split the data into train and test
# this is currently done in a very simple way: just take the first train_frac*100 percent for train, rest for test
# but it could be more complex: e.g. selecting only among the "pure variation"/non-label columns of labels to ensure that every label value is in both train and test
# assuming that the factors_label_list shall contain rather few items compared to factors_variation_dict.keys(), this should not make a large difference
# take a random sample
train_indices = np.random.choice(data.shape[0], int(data.shape[0] * train_frac), replace=False)
test_indices = np.setdiff1d(np.arange(data.shape[0]), train_indices)
assert np.array_equal(np.setdiff1d(np.concatenate((train_indices, test_indices)), np.arange(data.shape[0])), np.array([]).astype(int))
if train:
data = data[train_indices]
labels = labels[train_indices]
else:
data = data[test_indices]
labels = labels[test_indices]
# convert data to numpy (after slicing, sizing etc. is all done to be minimum size)
print("converting to numpy. this can take more than one minute, depending on how much variation is selected...")
data = data[()]
# data = data.astype(float)
# channel first -> transpose dimensions
data = np.transpose(data, (0, 3, 1, 2))
# convert numpy to torch
data = torch.tensor(data)
labels = torch.tensor(labels)
# cast data to FLoatTensor
data = data.float()
# Put both data and targets on GPU in advance
# attribute is self.labels, see https://pytorch.org/docs/stable/_modules/torchvision/datasets/svhn.html#SVHN
data, labels = data.to(device), labels.to(device)
# assign as instance variables
self.data = data
self.labels = labels
def __getitem__(self, index):
"""
Args:
index (int): Index
Returns:
tuple: (image, target) where target is index of the target class.
"""
img, label = self.data[index], self.labels[index]
return img, label
def __len__(self):
"""
"""
return self.data.shape[0]
if __name__ == "__main__":
dataset = Fast_3DShapes(train=True, device='cuda:1', train_frac = 1.,
factors_variation_dict={'floor_hue': list(range(4)), 'wall_hue': list(range(4)), 'object_hue': list(range(4)), 'scale': list(range(4)), 'shape': list(range(4)), 'orientation': list(range(5))}, \
factors_label_list=['floor_hue', 'wall_hue', 'object_hue', 'scale', 'shape', 'orientation'])
# dataset = Fast_3DShapes(train=True, device='cuda:1', train_frac = 1.,
# factors_variation_dict={'floor_hue': list(range(10)), 'wall_hue': list(range(10)), 'object_hue': list(range(10)), 'scale': list(range(8)), 'shape': list(range(4)), 'orientation': list(range(15))},
# factors_label_list=['floor_hue', 'wall_hue', 'object_hue', 'scale', 'shape', 'orientation'])
_FACTOR_TO_COLUMN_INDEX = {'floor_hue': 0, 'wall_hue': 1, 'object_hue': 2, 'scale': 3, 'shape': 4, 'orientation': 5}