-
Notifications
You must be signed in to change notification settings - Fork 200
/
augmix.py
256 lines (222 loc) · 9.98 KB
/
augmix.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
# coding=utf-8
# Copyright 2021 The Uncertainty Baselines Authors.
#
# 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.
"""Augmix utilities."""
import edward2 as ed
import numpy as np
import tensorflow as tf
import tensorflow_probability as tfp
tfd = tfp.distributions
# We use the convention of using mean = np.mean(train_images, axis=(0,1,2))
# and std = np.std(train_images, axis=(0,1,2)).
CIFAR10_MEAN = np.array([0.4914, 0.4822, 0.4465])
CIFAR10_STD = np.array([0.2470, 0.2435, 0.2616])
# Previously we used std = np.mean(np.std(train_images, axis=(1, 2)), axis=0)
# which gave std = tf.constant([0.2023, 0.1994, 0.2010], dtype=dtype), however
# we change convention to use the std over the entire training set instead.
def normalize_convert_image(input_image,
dtype,
mean=CIFAR10_MEAN,
std=CIFAR10_STD):
if input_image.dtype == tf.uint8:
input_image = tf.image.convert_image_dtype(input_image, dtype)
return ((input_image - tf.constant(mean, dtype=dtype)) /
tf.constant(std, dtype=dtype))
def augment_and_mix(image,
depth,
width,
prob_coeff,
augmenter,
dtype,
mean=CIFAR10_MEAN,
std=CIFAR10_STD,
seed=None):
"""Apply mixture of augmentations to image."""
if seed is None:
seed = tf.random.uniform((2,), maxval=int(1e10), dtype=tf.int32)
# We need three seeds, one for sampling from the Beta distribution, one for
# sampling from the Dirichlet distribution, one for sampling the depth, and a
# fourth seed to split for each individual RandAugment augmentation.
augment_seeds = tf.cast(tf.random.experimental.stateless_split(seed, num=4),
tf.int32)
# If seed is (2,), then sample returns a deterministically random sample.
mix_weight = tf.squeeze(tfd.Beta([prob_coeff], [prob_coeff]).sample(
[1], seed=augment_seeds[0]))
if width > 1:
branch_weights = tf.squeeze(tfd.Dirichlet([prob_coeff] * width).sample(
[1], seed=augment_seeds[1]))
else:
branch_weights = tf.constant([1.])
if depth < 0:
depth = tf.random.stateless_uniform([width],
augment_seeds[2],
minval=1,
maxval=4,
dtype=tf.dtypes.int32)
else:
depth = tf.constant([depth] * width)
mix = tf.cast(tf.zeros_like(image), tf.float32)
# Generate width * sum(depth) seeds for each individual augmentation.
distort_seeds = tf.random.experimental.stateless_split(
seed, num=width * tf.reduce_sum(depth))
seed_count = 0
for i in tf.range(width):
branch_img = tf.identity(image)
for _ in tf.range(depth[i]):
branch_img = augmenter.distort(branch_img, distort_seeds[seed_count])
seed_count += 1
branch_img = normalize_convert_image(branch_img, dtype, mean, std)
mix += branch_weights[i] * branch_img
return mix_weight * mix + (1 - mix_weight) * normalize_convert_image(
image, dtype, mean, std)
def do_augmix(image,
params,
augmenter,
dtype,
mean=CIFAR10_MEAN,
std=CIFAR10_STD,
seed=None):
"""Apply augmix augmentation to image."""
depth = params['augmix_depth']
width = params['augmix_width']
prob_coeff = params['augmix_prob_coeff']
count = params['aug_count']
if seed is None:
seed = tf.random.uniform((2,), maxval=int(1e10), dtype=tf.int32)
augment_seeds = tf.random.experimental.stateless_split(seed, num=count)
augmented = [
augment_and_mix(image, depth, width, prob_coeff, augmenter, dtype, mean,
std, seed=augment_seeds[c]) for c in range(count)
]
image = normalize_convert_image(image, dtype, mean, std)
return tf.stack([image] + augmented, 0)
def mixup(batch_size, aug_params, images, labels):
"""Applies Mixup regularization to a batch of images and labels.
[1] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, David Lopez-Paz
Mixup: Beyond Empirical Risk Minimization.
ICLR'18, https://arxiv.org/abs/1710.09412
`aug_params` can have the follwing fields:
augmix: whether or not to run AugMix.
mixup_alpha: the alpha to use in the Beta distribution.
aug_count: the number of augmentations to use in AugMix.
same_mix_weight_per_batch: whether to use the same mix coef over the batch.
use_truncated_beta: whether to sample from Beta_[0,1](alpha, alpha) or from
the truncated distribution Beta_[1/2, 1](alpha, alpha).
use_random_shuffling: Whether to pair images by random shuffling
(default is a deterministic pairing by reversing the batch).
Arguments:
batch_size: The input batch size for images and labels.
aug_params: Dict of data augmentation hyper parameters.
images: A batch of images of shape [batch_size, ...]
labels: A batch of labels of shape [batch_size, num_classes]
Returns:
A tuple of (images, labels) with the same dimensions as the input with
Mixup regularization applied.
"""
augmix = aug_params.get('augmix', False)
alpha = aug_params.get('mixup_alpha', 0.)
aug_count = aug_params.get('aug_count', 3)
same_mix_weight_per_batch = aug_params.get('same_mix_weight_per_batch', False)
use_truncated_beta = aug_params.get('use_truncated_beta', True)
use_random_shuffling = aug_params.get('use_random_shuffling', False)
if augmix and same_mix_weight_per_batch:
raise ValueError(
'Can only set one of `augmix` or `same_mix_weight_per_batch`.')
# 4 is hard-coding to aug_count=3. Fix this later!
if augmix:
mix_weight = ed.Beta(
alpha, alpha, sample_shape=[batch_size, aug_count + 1, 1])
elif same_mix_weight_per_batch:
mix_weight = ed.Beta(alpha, alpha, sample_shape=[1, 1])
mix_weight = tf.tile(mix_weight, [batch_size, 1])
else:
mix_weight = ed.Beta(alpha, alpha, sample_shape=[batch_size, 1])
if use_truncated_beta:
mix_weight = tf.maximum(mix_weight, 1. - mix_weight)
if augmix:
images_mix_weight = tf.reshape(mix_weight,
[batch_size, aug_count + 1, 1, 1, 1])
else:
images_mix_weight = tf.reshape(mix_weight, [batch_size, 1, 1, 1])
images_mix_weight = tf.cast(images_mix_weight, images.dtype)
if use_random_shuffling:
mixup_index = tf.random.shuffle(tf.range(batch_size))
else:
# Mixup on a single batch is implemented by taking a weighted sum with the
# same batch in reverse.
mixup_index = tf.reverse(tf.range(batch_size), axis=[0])
images_mix = (
images * images_mix_weight + tf.gather(images, mixup_index) *
(1. - images_mix_weight))
mix_weight = tf.cast(mix_weight, labels.dtype)
if augmix:
labels = tf.reshape(
tf.tile(labels, [1, aug_count + 1]), [batch_size, aug_count + 1, -1])
labels_mix = (
labels * mix_weight +
tf.gather(labels, mixup_index) * (1. - mix_weight))
labels_mix = tf.reshape(
tf.transpose(labels_mix, [1, 0, 2]), [batch_size * (aug_count + 1), -1])
else:
labels_mix = (
labels * mix_weight +
tf.gather(labels, mixup_index) * (1. - mix_weight))
return images_mix, labels_mix
def adaptive_mixup(batch_size, aug_params, images, labels):
"""Applies Confidence Adjusted Mixup (CAMixup) regularization.
[1] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, David Lopez-Paz
Mixup: Beyond Empirical Risk Minimization.
ICLR'18, https://arxiv.org/abs/1710.09412
Arguments:
batch_size: The input batch size for images and labels.
aug_params: Dict of data augmentation hyper parameters.
images: A batch of images of shape [batch_size, ...]
labels: A batch of labels of shape [batch_size, num_classes]
Returns:
A tuple of (images, labels) with the same dimensions as the input with
Mixup regularization applied.
"""
augmix = aug_params.get('augmix', False)
ensemble_size = aug_params['ensemble_size']
mixup_coeff = aug_params['mixup_coeff']
scalar_labels = tf.argmax(labels, axis=1)
alpha = tf.gather(mixup_coeff, scalar_labels, axis=-1) # 4 x Batch_size
# Need to filter out elements in alpha which equal to 0.
greater_zero_indicator = tf.cast(alpha > 0, alpha.dtype)
less_one_indicator = tf.cast(alpha < 1, alpha.dtype)
valid_alpha_indicator = tf.cast(
greater_zero_indicator * less_one_indicator, tf.bool)
sampled_alpha = tf.where(valid_alpha_indicator, alpha, 0.1)
mix_weight = tfd.Beta(sampled_alpha, sampled_alpha).sample()
mix_weight = tf.where(valid_alpha_indicator, mix_weight, alpha)
mix_weight = tf.reshape(mix_weight, [ensemble_size * batch_size, 1])
mix_weight = tf.clip_by_value(mix_weight, 0, 1)
mix_weight = tf.maximum(mix_weight, 1. - mix_weight)
images_mix_weight = tf.reshape(mix_weight,
[ensemble_size * batch_size, 1, 1, 1])
images_mix_weight = tf.cast(images_mix_weight, images.dtype)
# Mixup on a single batch is implemented by taking a weighted sum with the
# same batch in reverse.
if augmix:
images_shape = tf.shape(images)
images = tf.reshape(
tf.transpose(images, [1, 0, 2, 3, 4]),
[-1, images_shape[2], images_shape[3], images_shape[4]])
else:
images = tf.tile(images, [ensemble_size, 1, 1, 1])
labels = tf.tile(labels, [ensemble_size, 1])
images_mix = (
images * images_mix_weight + images[::-1] * (1. - images_mix_weight))
labels_mix = labels * mix_weight + labels[::-1] * (1. - mix_weight)
return images_mix, labels_mix