-
Notifications
You must be signed in to change notification settings - Fork 65
/
utils.py
126 lines (110 loc) · 5.66 KB
/
utils.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
import numpy as np
import os
from PIL import Image
from PIL import ImageOps
# from skimage import io
# from skimage import transform
# from skimage import util
def generate_random_strings(batch_size, seq_length, vector_dim):
return np.random.randint(0, 2, size=[batch_size, seq_length, vector_dim]).astype(np.float32)
def one_hot_encode(x, dim):
res = np.zeros(np.shape(x) + (dim, ), dtype=np.float32)
it = np.nditer(x, flags=['multi_index'])
while not it.finished:
res[it.multi_index][it[0]] = 1
it.iternext()
return res
def one_hot_decode(x):
return np.argmax(x, axis=-1)
def five_hot_decode(x):
x = np.reshape(x, newshape=np.shape(x)[:-1] + (5, 5))
def f(a):
return sum([a[i] * 5 ** i for i in range(5)])
return np.apply_along_axis(f, -1, np.argmax(x, axis=-1))
def baseN(num,b):
return ((num == 0) and "0" ) or ( baseN(num // b, b).lstrip("0") + "0123456789abcdefghijklmnopqrstuvwxyz"[num % b])
class OmniglotDataLoader:
def __init__(self, data_dir='./data', image_size=(20, 20), n_train_classses=1200, n_test_classes=423):
self.data = []
self.image_size = image_size
for dirname, subdirname, filelist in os.walk(data_dir):
if filelist:
self.data.append(
# [np.reshape(
# np.array(Image.open(dirname + '/' + filename).resize(image_size), dtype=np.float32),
# newshape=(image_size[0] * image_size[1])
# )
# for filename in filelist]
# [io.imread(dirname + '/' + filename).astype(np.float32) / 255 for filename in filelist]
[Image.open(dirname + '/' + filename).copy() for filename in filelist]
)
self.train_data = self.data[:n_train_classses]
self.test_data = self.data[-n_test_classes:]
def fetch_batch(self, n_classes, batch_size, seq_length,
type='train',
sample_strategy='random',
augment=True,
label_type='one_hot'):
if type == 'train':
data = self.train_data
elif type == 'test':
data = self.test_data
classes = [np.random.choice(range(len(data)), replace=False, size=n_classes) for _ in range(batch_size)]
if sample_strategy == 'random': # #(sample) per class may not be equal (sec 7)
seq = np.random.randint(0, n_classes, [batch_size, seq_length])
elif sample_strategy == 'uniform': # #(sample) per class are equal
seq = np.array([np.concatenate([[j] * int(seq_length / n_classes) for j in range(n_classes)])
for i in range(batch_size)])
for i in range(batch_size):
np.random.shuffle(seq[i, :])
self.rand_rotate_init(n_classes, batch_size)
seq_pic = [[self.augment(data[classes[i][j]][np.random.randint(0, len(data[classes[i][j]]))],
batch=i, c=j,
only_resize=not augment)
for j in seq[i, :]]
for i in range(batch_size)]
if label_type == 'one_hot':
seq_encoded = one_hot_encode(seq, n_classes)
seq_encoded_shifted = np.concatenate(
[np.zeros(shape=[batch_size, 1, n_classes]), seq_encoded[:, :-1, :]], axis=1
)
elif label_type == 'five_hot':
label_dict = [[[int(j) for j in list(baseN(i, 5)) + [0] * (5 - len(baseN(i, 5)))]
for i in np.random.choice(range(5 ** 5), replace=False, size=n_classes)]
for _ in range(batch_size)]
seq_encoded_ = np.array([[label_dict[b][i] for i in seq[b]] for b in range(batch_size)])
seq_encoded = np.reshape(one_hot_encode(seq_encoded_, dim=5), newshape=[batch_size, seq_length, -1])
seq_encoded_shifted = np.concatenate(
[np.zeros(shape=[batch_size, 1, 25]), seq_encoded[:, :-1, :]], axis=1
)
return seq_pic, seq_encoded_shifted, seq_encoded
def rand_rotate_init(self, n_classes, batch_size):
self.rand_rotate_map = np.random.randint(0, 4, [batch_size, n_classes])
def augment(self, image, batch, c, only_resize=False):
if only_resize:
image = ImageOps.invert(image.convert('L')).resize(self.image_size)
else:
rand_rotate = self.rand_rotate_map[batch, c] * 90 # rotate by 0, pi/2, pi, 3pi/2
image = ImageOps.invert(image.convert('L')) \
.rotate(rand_rotate + np.random.rand() * 22.5 - 11.25,
translate=np.random.randint(-10, 11, size=2).tolist()) \
.resize(self.image_size) # rotate between -pi/16 to pi/16, translate bewteen -10 and 10
np_image = np.reshape(np.array(image, dtype=np.float32),
newshape=(self.image_size[0] * self.image_size[1]))
max_value = np.max(np_image) # normalization is important
if max_value > 0.:
np_image = np_image / max_value
return np_image
# mat = transform.AffineTransform(translation=np.random.randint(-10, 11, size=2).tolist())
# return np.reshape(
# util.invert(
# transform.resize(
# transform.warp(
# transform.rotate(
# util.invert(image),
# angle=rand_rotate + np.random.rand() * 22.5 - 11.25
# ), mat
# ), output_shape=self.image_size
# )
# ), newshape=(self.image_size[0] * self.image_size[1])
# )