|
1
|
|
|
from fuel.datasets import H5PYDataset |
|
2
|
|
|
from fuel.transformers.defaults import uint8_pixels_to_floatX |
|
3
|
|
|
from fuel.utils import find_in_data_path |
|
4
|
|
|
|
|
5
|
|
|
|
|
6
|
|
|
class CelebA(H5PYDataset): |
|
7
|
|
|
"""The CelebFaces Attributes Dataset (CelebA) dataset. |
|
8
|
|
|
|
|
9
|
|
|
CelebA is a large-scale face |
|
10
|
|
|
attributes dataset with more than 200K celebrity images, each |
|
11
|
|
|
with 40 attribute annotations. The images in this dataset cover |
|
12
|
|
|
large pose variations and background clutter. CelebA has large |
|
13
|
|
|
diversities, large quantities, and rich annotations, including: |
|
14
|
|
|
|
|
15
|
|
|
* 10,177 number of identities |
|
16
|
|
|
* 202,599 number of face images |
|
17
|
|
|
* 5 landmark locations per image |
|
18
|
|
|
* 40 binary attributes annotations per image. |
|
19
|
|
|
|
|
20
|
|
|
The dataset can be employed as the training and test sets for |
|
21
|
|
|
the following computer vision tasks: |
|
22
|
|
|
|
|
23
|
|
|
* face attribute recognition |
|
24
|
|
|
* face detection |
|
25
|
|
|
* landmark (or facial part) localization |
|
26
|
|
|
|
|
27
|
|
|
Parameters |
|
28
|
|
|
---------- |
|
29
|
|
|
which_format : {'aligned_cropped, '64'} |
|
30
|
|
|
Either the aligned and cropped version of CelebA, or |
|
31
|
|
|
a 64x64 version of it. |
|
32
|
|
|
which_sets : tuple of str |
|
33
|
|
|
Which split to load. Valid values are 'train', 'valid' and |
|
34
|
|
|
'test' corresponding to the training set (162,770 examples), the |
|
35
|
|
|
validation set (19,867 examples) and the test set (19,962 |
|
36
|
|
|
examples). |
|
37
|
|
|
|
|
38
|
|
|
""" |
|
39
|
|
|
_filename = 'celeba_{}.hdf5' |
|
40
|
|
|
default_transformers = uint8_pixels_to_floatX(('features',)) |
|
41
|
|
|
|
|
42
|
|
|
def __init__(self, which_format, which_sets, **kwargs): |
|
43
|
|
|
self.which_format = which_format |
|
44
|
|
|
super(CelebA, self).__init__( |
|
45
|
|
|
file_or_path=find_in_data_path(self.filename), |
|
46
|
|
|
which_sets=which_sets, **kwargs) |
|
47
|
|
|
|
|
48
|
|
|
@property |
|
49
|
|
|
def filename(self): |
|
50
|
|
|
return self._filename.format(self.which_format) |
|
51
|
|
|
|
mila-udem /
fuel
