-
Notifications
You must be signed in to change notification settings - Fork 0
/
pre_process.py
156 lines (142 loc) · 5.15 KB
/
pre_process.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
from PIL import Image
import numpy as np
from sklearn.cluster import KMeans
from sklearn.utils import shuffle
import os.path
import os
from itertools import chain
import ntpath
import multiprocessing as mp
from skimage.color import rgb2lab
from glob import glob
from math import cos, sin
def get_image_palette(img, nclusters):
"""
Extract tuple of (Image, Palette) in LAB space
"""
lab = rgb2lab(np.array(img))
palette = kmeans_get_palette(lab, nclusters)
return lab, palette
def get_augmented_image_palette(img, nclusters, angle):
"""
Return tuple of (Image, Palette) in LAB space
color shifted by the angle parameter
"""
lab = rgb2lab(img)
ch_a = lab[...,1]
ch_b = lab[...,2]
theta = np.deg2rad(angle)
rot = np.array([[cos(theta), -sin(theta)], [sin(theta), cos(theta)]])
hue_rotate = lambda ab: np.dot(rot, [ab[0], ab[1]])
ab = np.asarray(list(map(hue_rotate, zip(ch_a, ch_b)))).transpose((0, 2, 1))
lab = np.dstack((lab[...,0], ab[...,0], ab[...,1]))
palette = kmeans_get_palette(lab, nclusters)
return (lab, palette)
def kmeans_get_palette(img, nclusters):
"""
Extract color palette of nclusters x 3 rgb values
by performing a k-means clustering of the input image
"""
h, w, c = img.shape
km_im = np.array(img, dtype=np.float64).reshape(w * h, c)
image_array_sample = shuffle(km_im, random_state=0, n_samples=1_000)
kmeans = KMeans(n_clusters=nclusters, random_state=0).fit(image_array_sample)
codebook = kmeans.cluster_centers_
return codebook
def save_image_palette(image, palette, path_base, angle=0):
"""
Save preprocessed image and palette as .npz
:param nd array image: Input image as nd array
:param nd array palette: Input palette as nd array
:param str path_base: Base save path
:param int angle: Augmentation angle to use as a part of the save path
"""
if(angle != 0):
palette_file = os.path.join(path_base, str(angle).rjust(3, '0') + "_pal.npz")
image_file = os.path.join(path_base, str(angle).rjust(3, '0') + "_img.npz")
else:
palette_file = os.path.join(path_base, "pal.npz")
image_file = os.path.join(path_base, "img.npz")
if(not os.path.exists(path_base)):
os.makedirs(path_base)
np.savez_compressed(image_file, image)
np.savez_compressed(palette_file, palette)
def center_crop(img, w_target, h_target):
"""
Center-crop a PIL image according to the
target dimensions
"""
w_im, h_im = img.size
# preconditions:
if(w_target > w_im or h_target > h_im):
return None
# get the min dimension:
min_dim = min(w_im, h_im)
if w_im == min_dim:
h_new = float(w_target) / float(w_im) * float(h_im)
img = img.resize((int(w_target), int(h_new)))
# crop height:
d = int(0.5 * float(h_new - h_target))
top = d
bottom = d + h_target
img = img.crop((0, top, w_target, bottom))
else:
w_new = float(h_target) / float(h_im) * float(w_im)
img = img.resize((int(w_new), int(h_target)))
# crop width:
d = int(0.5 * float(w_new - w_target))
left = d
right = d + w_target
img = img.crop((left, 0, right, h_target))
return img
def preprocess_single(args):
"""
Pre-process a single raw input image:
1. Load input image
2. Center crop to match training set dims
3. Create a color-augmented version of the input by 40
rotations in LAB color space
4. Save the input and output image-palette pairs
"""
im_path, source_path, target_path, w, h, n_p = args
filename = os.path.splitext(ntpath.basename(im_path))[0]
print('input: ' + im_path + '...')
im = Image.open(im_path)
w_im, h_im = im.size
if(w_im < w or h_im < h):
return
angles = range(40, 360, 40)
im_crop = center_crop(im, w, h)
if im_crop == None:
return
im_crop = np.array(im_crop) / 255.
aug_pairs = map(lambda angle:
get_augmented_image_palette(im_crop, n_p, angle), angles)
# save image palette pairs (target)
for (im_aug, pal), ang in zip(aug_pairs, angles):
save_image_palette(im_aug, pal, os.path.join(target_path, filename), ang)
# save source image palette pair
im, palette = get_image_palette(im_crop, n_p)
save_image_palette(im, palette, os.path.join(source_path, filename))
def preprocess(dataset_path, source_path, target_path, w, h, n_p, num_workers):
"""
Preprocess raw input data in parallel
"""
im_paths_png = (chain.from_iterable(glob(os.path.join(x[0], '*.png')) for x in os.walk(dataset_path)))
im_paths_jpg = (chain.from_iterable(glob(os.path.join(x[0], '*.jpg')) for x in os.walk(dataset_path)))
im_paths = [*im_paths_png, *im_paths_jpg]
pool = mp.Pool(num_workers)
pool.map(preprocess_single, [(im_path, source_path, target_path, w, h, n_p) for im_path in im_paths])
pool.close()
if __name__ == '__main__':
w = 256
h = 256
n_p = 6
preprocess("./dataset/div2k",
source_path="./dataset/source",
target_path="./dataset/target",
w=w,
h=h,
n_p=6,
num_workers=16
)