-
Notifications
You must be signed in to change notification settings - Fork 1
/
data_aug.py
145 lines (124 loc) · 5.59 KB
/
data_aug.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
import numpy as np
import random
from genericpath import exists
from scipy.ndimage.filters import gaussian_filter
import numpy as np
import random
def shape_scale_constraints(mask_a, mask_b):
# shape and scale constraints
shape_factor = np.sum(np.abs(mask_a - mask_b)) / max(np.sum(mask_a), np.sum(mask_b))
scale_factor = min(np.sum(mask_a), np.sum(mask_b)) / max(np.sum(mask_a), np.sum(mask_b))
return shape_factor, scale_factor
def insmix(x_all, cent_sort, y_ist_sort, y_bnd_sort, sh, sw, crop_size):
"""
1. instance library, add 0 to 600
2. distance aware
3. shape constraints
"""
x_patch = x_all[:, sh:sh+crop_size, sw:sw+crop_size]
y_ist_patch = y_ist_sort[sh:sh+crop_size, sw:sw+crop_size]
y_gt_patch = y_ist_patch > 0
y_bnd_patch = y_bnd_sort[sh:sh+crop_size, sw:sw+crop_size] > 0
cent_patch = cent_sort[sh:sh+crop_size, sw:sw+crop_size]
max_add = 0.5 * len(list(cent_patch[cent_patch > 0]))
dis = np.random.randint(-40, 40)
idx_add_list = list(cent_sort[cent_sort > 0])
random.shuffle(idx_add_list)
if min(max_add, len(idx_add_list)) < 2:
num_add = 0
else:
num_add = np.random.randint(1, min(max_add, len(idx_add_list)))
cent_a_id_list = list(cent_patch[cent_patch > 0].astype(int))
cent_a_id_list.sort()
for i in range(num_add):
idx_add = idx_add_list[i]
mask_ist = np.array(np.where(y_ist_sort==idx_add))
mask_bnd = np.array(np.where(y_bnd_sort==idx_add))
mask_cent = np.array(np.where(cent_sort==idx_add))
# which nuclear to be augmented
nuc_id = cent_a_id_list[i % len(cent_a_id_list)]
center_point = np.where(cent_patch == nuc_id)
shift_h = mask_cent[0][0] - center_point[0][0]
shift_w = mask_cent[1][0] - center_point[1][0]
mask_ist_a = mask_ist.copy()
mask_ist_a[0] = mask_ist_a[0] - shift_h + dis
mask_ist_a[1] = mask_ist_a[1] - shift_w + dis
mask_bnd_a = mask_bnd.copy()
mask_bnd_a[0] = mask_bnd_a[0] - shift_h + dis
mask_bnd_a[1] = mask_bnd_a[1] - shift_w + dis
# remove out of bound
pos_ist = np.all(mask_ist_a >= 0, axis=0) * np.all(mask_ist_a < crop_size, axis=0)
pos_bnd = np.all(mask_bnd_a >= 0, axis=0) * np.all(mask_bnd_a < crop_size, axis=0)
mask_ist = mask_ist[:, pos_ist]
mask_ist_a = mask_ist_a[:, pos_ist]
mask_bnd = mask_bnd[:, pos_bnd]
mask_bnd_a = mask_bnd_a[:, pos_bnd]
mask_ist, mask_ist_a, mask_bnd, mask_bnd_a = map(tuple, [mask_ist, mask_ist_a, mask_bnd, mask_bnd_a])
# shape constraints
shape_constraints = 0.8
scale_constraints = 0.5
mask_a, mask_b = np.zeros_like(y_ist_patch), np.zeros_like(y_ist_patch)
mask_b_idx = np.array(mask_ist_a)
mask_b_idx = np.clip(mask_b_idx - dis, 0, crop_size-1)
mask_a[y_ist_patch==nuc_id] = 1
mask_b_idx = tuple(mask_b_idx)
mask_b[mask_b_idx] = 1
shape_factor, scale_factor = shape_scale_constraints(mask_a, mask_b)
# print(shape_factor, scale_factor)
if shape_factor > shape_constraints or scale_factor < scale_constraints:
continue
# mix
for i in range(x_all.shape[0]):
x_patch[i][mask_ist_a] = x_all[i][mask_ist]
y_gt_patch[mask_ist_a] = y_ist_sort[mask_ist]
y_bnd_patch[mask_ist_a] = 0
y_bnd_patch[mask_bnd_a] = y_bnd_sort[mask_bnd]
return x_patch, y_gt_patch, y_bnd_patch
def pashuffle(num, perc=0.5):
"""
num: number of patches
perc: percentage of patches to be shuffled
"""
num_shuffle = int(num * perc)
idx_shuffle = np.random.choice(num, num_shuffle, replace=False)
return idx_shuffle
def bg_shuffle(img, label, anchor_size, bg_shift_radio=0.5):
"""
background shift
img_shape: (c, h, w)
"""
img_out = img.copy()
x = np.arange(0, img.shape[1] - anchor_size, anchor_size)
y = np.arange(0, img.shape[2] - anchor_size, anchor_size)
xx, yy = np.meshgrid(x, y)
anchor_center = np.stack([xx, yy], axis=2).reshape(-1, 2)
boxes = np.concatenate([anchor_center, anchor_center + anchor_size], axis=1)
# remove the foreground patches
keep = np.ones_like(boxes, dtype=bool)
for i in range(boxes.shape[0]):
box = boxes[i]
if (label[box[0]:box[2], box[1]:box[3]] > 0).any():
keep[i] = False
boxes = boxes[keep].reshape(-1, 4)
idx_shuffle1 = pashuffle(boxes.shape[0], bg_shift_radio)
idx_shuffle2 = np.random.permutation(idx_shuffle1)
for i in range(idx_shuffle1.shape[0]):
box1 = boxes[idx_shuffle1[i]]
box2 = boxes[idx_shuffle2[i]]
img_out[:, box1[0]:box1[2], box1[1]:box1[3]] = img[:, box2[0]:box2[2], box2[1]:box2[3]]
return img_out
def elastic_transform(shape, alpha, sigma, random_state=None):
"""Elastic deformation of images as described in [Simard2003]_.
.. [Simard2003] Simard, Steinkraus and Platt, "Best Practices for
Convolutional Neural Networks applied to Visual Document Analysis", in
Proc. of the International Conference on Document Analysis and
Recognition, 2003.
"""
# This function get indices only.
if random_state is None:
random_state = np.random.RandomState(None)
dx = gaussian_filter((random_state.rand(*shape) * 2 - 1), sigma, mode="constant", cval=0) * alpha
dy = gaussian_filter((random_state.rand(*shape) * 2 - 1), sigma, mode="constant", cval=0) * alpha
x, y = np.meshgrid(np.arange(shape[0]), np.arange(shape[1]), indexing='ij')
indices = np.reshape(x+dx, (-1, 1)), np.reshape(y+dy, (-1, 1))
return indices