File size: 4,729 Bytes
8620e7b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c12c339
 
 
 
 
 
 
 
8620e7b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import datasets
import numpy as np
import os
from PIL import Image

_SHAPES3D_URL = "https://huggingface.co/datasets/randall-lab/shapes3d/resolve/main/shapes3d.npz"

class Shapes3D(datasets.GeneratorBasedBuilder):
    """Shapes3D dataset: 10x10x10x8x4x15 factor combinations, 64x64 RGB images."""

    VERSION = datasets.Version("1.0.0")

    def _info(self):
        return datasets.DatasetInfo(
            description=(
                "Shapes3D dataset: procedurally generated images of 3D shapes with 6 independent factors of variation. "
                "Commonly used for disentangled representation learning. "
                "Factors: floor hue (10), wall hue (10), object hue (10), scale (8), shape (4), orientation (15). "
                "Images are stored as the Cartesian product of the factors in row-major order."
            ),
            features=datasets.Features(
                {
                    "image": datasets.Image(),    # (64, 64, 3)
                    "index": datasets.Value("int32"),  # index of the image
                    "label": datasets.Sequence(datasets.Value("float64")),  # 6 factor values (continuous)
                    "label_index": datasets.Sequence(datasets.Value("int64")),  # 6 factor indices
                    "floor": datasets.Value("float64"),  # value of floor (0-1)
                    "wall": datasets.Value("float64"),   # value of wall (0-1)
                    "object": datasets.Value("float64"),  # value of object (0-1)
                    "scale": datasets.Value("float64"),  # value of scale (0.75-1.25)
                    "shape": datasets.Value("float64"),    # value of shape (0-3)
                    "orientation": datasets.Value("float64"),  # value of orientation (-30 to 30)
                    "floor_idx": datasets.Value("int32"),
                    "wall_idx": datasets.Value("int32"),
                    "object_idx": datasets.Value("int32"),
                    "scale_idx": datasets.Value("int32"),
                    "shape_idx": datasets.Value("int32"),
                    "orientation_idx": datasets.Value("int32"),
                }
            ),
            supervised_keys=("image", "label"),
            homepage="https://github.com/google-deepmind/3dshapes-dataset/",
            license="apache-2.0",
            citation="""@InProceedings{pmlr-v80-kim18b,
  title = {Disentangling by Factorising},
  author = {Kim, Hyunjik and Mnih, Andriy},
  booktitle = {Proceedings of the 35th International Conference on Machine Learning},
  pages = {2649--2658},
  year = {2018},
  editor = {Dy, Jennifer and Krause, Andreas},
  volume = {80},
  series = {Proceedings of Machine Learning Research},
  month = {10--15 Jul},
  publisher = {PMLR},
  pdf = {http://proceedings.mlr.press/v80/kim18b/kim18b.pdf},
  url = {https://proceedings.mlr.press/v80/kim18b.html}
}""",
        )

    def _split_generators(self, dl_manager):
        npz_path = dl_manager.download(_SHAPES3D_URL)

        return [
            datasets.SplitGenerator(
                name=datasets.Split.TRAIN,
                gen_kwargs={"npz_path": npz_path},
            ),
        ]

    def _generate_examples(self, npz_path):
        # Load npz
        data = np.load(npz_path)
        images = data["images"]  # (480000, 64, 64, 3)
        labels = data["labels"]  # (480000, 6)

        # Define factor sizes (from README / paper)
        factor_sizes = np.array([10, 10, 10, 8, 4, 15])
        factor_bases = np.cumprod([1] + list(factor_sizes[::-1]))[::-1][1:]

        def index_to_factors(index):
            factors = []
            for base, size in zip(factor_bases, factor_sizes):
                factor = (index // base) % size
                factors.append(int(factor))
            return factors

        # Iterate over images
        for idx in range(len(images)):
            img = images[idx]
            img_pil = Image.fromarray(img)

            label_value = labels[idx].tolist()
            label_index = index_to_factors(idx)

            yield idx, {
                "image": img_pil,
                "index": idx,
                "label": label_value,
                "label_index": label_index,
                "floor": label_value[0],
                "wall": label_value[1], 
                "object": label_value[2],
                "scale": label_value[3],
                "shape": label_value[4], 
                "orientation": label_value[5],
                "floor_idx": label_index[0],
                "wall_idx": label_index[1],
                "object_idx": label_index[2],
                "scale_idx": label_index[3],
                "shape_idx": label_index[4],
                "orientation_idx": label_index[5],
            }