/
dsprites.py
268 lines (218 loc) · 9.93 KB
/
dsprites.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
# coding=utf-8
# Copyright 2018 The DisentanglementLib Authors. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""DSprites dataset and new variants with probabilistic decoders."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
from disentanglement_lib.data.ground_truth import ground_truth_data
from disentanglement_lib.data.ground_truth import util
import numpy as np
import PIL
from six.moves import range
from tensorflow.compat.v1 import gfile
DSPRITES_PATH = os.path.join(
os.environ.get("DISENTANGLEMENT_LIB_DATA", "."), "dsprites",
"dsprites_ndarray_co1sh3sc6or40x32y32_64x64.npz")
SCREAM_PATH = os.path.join(
os.environ.get("DISENTANGLEMENT_LIB_DATA", "."), "scream", "scream.jpg")
class DSprites(ground_truth_data.GroundTruthData):
"""DSprites dataset.
The data set was originally introduced in "beta-VAE: Learning Basic Visual
Concepts with a Constrained Variational Framework" and can be downloaded from
https://github.com/deepmind/dsprites-dataset.
The ground-truth factors of variation are (in the default setting):
0 - shape (3 different values)
1 - scale (6 different values)
2 - orientation (40 different values)
3 - position x (32 different values)
4 - position y (32 different values)
"""
def __init__(self, latent_factor_indices=None):
# By default, all factors (including shape) are considered ground truth
# factors.
if latent_factor_indices is None:
latent_factor_indices = list(range(6))
self.latent_factor_indices = latent_factor_indices
self.data_shape = [64, 64, 1]
# Load the data so that we can sample from it.
with gfile.Open(DSPRITES_PATH, "rb") as data_file:
# Data was saved originally using python2, so we need to set the encoding.
data = np.load(data_file, encoding="latin1", allow_pickle=True)
self.images = np.array(data["imgs"])
self.factor_sizes = np.array(
data["metadata"][()]["latents_sizes"], dtype=np.int64)
self.full_factor_sizes = [1, 3, 6, 40, 32, 32]
self.factor_bases = np.prod(self.factor_sizes) / np.cumprod(
self.factor_sizes)
self.state_space = util.SplitDiscreteStateSpace(self.factor_sizes,
self.latent_factor_indices)
@property
def num_factors(self):
return self.state_space.num_latent_factors
@property
def factors_num_values(self):
return [self.full_factor_sizes[i] for i in self.latent_factor_indices]
@property
def observation_shape(self):
return self.data_shape
def sample_factors(self, num, random_state):
"""Sample a batch of factors Y."""
return self.state_space.sample_latent_factors(num, random_state)
def sample_observations_from_factors(self, factors, random_state):
return self.sample_observations_from_factors_no_color(factors, random_state)
def sample_observations_from_factors_no_color(self, factors, random_state):
"""Sample a batch of observations X given a batch of factors Y."""
all_factors = self.state_space.sample_all_factors(factors, random_state)
indices = np.array(np.dot(all_factors, self.factor_bases), dtype=np.int64)
return np.expand_dims(self.images[indices].astype(np.float32), axis=3)
def _sample_factor(self, i, num, random_state):
return random_state.randint(self.factor_sizes[i], size=num)
class ColorDSprites(DSprites):
"""Color DSprites.
This data set is the same as the original DSprites data set except that when
sampling the observations X, the sprite is colored in a randomly sampled
color.
The ground-truth factors of variation are (in the default setting):
0 - shape (3 different values)
1 - scale (6 different values)
2 - orientation (40 different values)
3 - position x (32 different values)
4 - position y (32 different values)
"""
def __init__(self, latent_factor_indices=None):
DSprites.__init__(self, latent_factor_indices)
self.data_shape = [64, 64, 3]
def sample_observations_from_factors(self, factors, random_state):
no_color_observations = self.sample_observations_from_factors_no_color(
factors, random_state)
observations = np.repeat(no_color_observations, 3, axis=3)
color = np.repeat(
np.repeat(
random_state.uniform(0.5, 1, [observations.shape[0], 1, 1, 3]),
observations.shape[1],
axis=1),
observations.shape[2],
axis=2)
return observations * color
class NoisyDSprites(DSprites):
"""Noisy DSprites.
This data set is the same as the original DSprites data set except that when
sampling the observations X, the background pixels are replaced with random
noise.
The ground-truth factors of variation are (in the default setting):
0 - shape (3 different values)
1 - scale (6 different values)
2 - orientation (40 different values)
3 - position x (32 different values)
4 - position y (32 different values)
"""
def __init__(self, latent_factor_indices=None):
DSprites.__init__(self, latent_factor_indices)
self.data_shape = [64, 64, 3]
def sample_observations_from_factors(self, factors, random_state):
no_color_observations = self.sample_observations_from_factors_no_color(
factors, random_state)
observations = np.repeat(no_color_observations, 3, axis=3)
color = random_state.uniform(0, 1, [observations.shape[0], 64, 64, 3])
return np.minimum(observations + color, 1.)
class ScreamDSprites(DSprites):
"""Scream DSprites.
This data set is the same as the original DSprites data set except that when
sampling the observations X, a random patch of the Scream image is sampled as
the background and the sprite is embedded into the image by inverting the
color of the sampled patch at the pixels of the sprite.
The ground-truth factors of variation are (in the default setting):
0 - shape (3 different values)
1 - scale (6 different values)
2 - orientation (40 different values)
3 - position x (32 different values)
4 - position y (32 different values)
"""
def __init__(self, latent_factor_indices=None):
DSprites.__init__(self, latent_factor_indices)
self.data_shape = [64, 64, 3]
with gfile.Open(SCREAM_PATH, "rb") as f:
scream = PIL.Image.open(f)
scream.thumbnail((350, 274, 3))
self.scream = np.array(scream) * 1. / 255.
def sample_observations_from_factors(self, factors, random_state):
no_color_observations = self.sample_observations_from_factors_no_color(
factors, random_state)
observations = np.repeat(no_color_observations, 3, axis=3)
for i in range(observations.shape[0]):
x_crop = random_state.randint(0, self.scream.shape[0] - 64)
y_crop = random_state.randint(0, self.scream.shape[1] - 64)
background = (self.scream[x_crop:x_crop + 64, y_crop:y_crop + 64] +
random_state.uniform(0, 1, size=3)) / 2.
mask = (observations[i] == 1)
background[mask] = 1 - background[mask]
observations[i] = background
return observations
# Object colors generated using
# >> seaborn.husl_palette(n_colors=6, h=0.1, s=0.7, l=0.7)
OBJECT_COLORS = np.array(
[[0.9096231780824386, 0.5883403686424795, 0.3657680693481871],
[0.6350181801577739, 0.6927729880940552, 0.3626904230371999],
[0.3764832455369271, 0.7283900430001952, 0.5963114605342514],
[0.39548987063404156, 0.7073922557810771, 0.7874577552076919],
[0.6963644829189117, 0.6220697032672371, 0.899716387820763],
[0.90815966835861, 0.5511103319168646, 0.7494337214212151]])
BACKGROUND_COLORS = np.array([
(0., 0., 0.),
(.25, .25, .25),
(.5, .5, .5),
(.75, .75, .75),
(1., 1., 1.),
])
class AbstractDSprites(DSprites):
"""DSprites variation for abstract reasoning task.
Rotation is not considered a ground-truth factor and we sample random colors
both for the object and the background.
The ground-truth factors of variation are (in the default setting):
0 - background color (5 different values)
1 - object color (6 different values)
2 - shape (3 different values)
3 - scale (6 different values)
4 - position x (32 different values)
5 - position y (32 different values)
"""
def __init__(self):
# We retain all original factors except shape.
DSprites.__init__(self, [1, 2, 4, 5])
self.data_shape = [64, 64, 3]
@property
def num_factors(self):
return 2 + self.state_space.num_latent_factors
@property
def factors_num_values(self):
return ([BACKGROUND_COLORS.shape[0], OBJECT_COLORS.shape[0]] +
[self.full_factor_sizes[i] for i in self.latent_factor_indices])
def sample_factors(self, num, random_state):
"""Sample a batch of factors Y."""
colors = np.zeros((num, 2), dtype=np.int64)
colors[:, 0] = random_state.randint(BACKGROUND_COLORS.shape[0], size=num)
colors[:, 1] = random_state.randint(OBJECT_COLORS.shape[0], size=num)
other_factors = self.state_space.sample_latent_factors(num, random_state)
return np.concatenate([colors, other_factors], axis=-1)
def sample_observations_from_factors(self, factors, random_state):
mask = self.sample_observations_from_factors_no_color(
factors[:, 2:], random_state)
background_color = BACKGROUND_COLORS[factors[:, 0]]
object_color = OBJECT_COLORS[factors[:, 1]]
# Add axis for height and width.
background_color = np.expand_dims(np.expand_dims(background_color, 1), 1)
object_color = np.expand_dims(np.expand_dims(object_color, 1), 1)
return mask * object_color + (1. - mask) * background_color