-
Notifications
You must be signed in to change notification settings - Fork 6
/
routines.py
285 lines (244 loc) · 11.1 KB
/
routines.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
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
# -*- coding: utf-8 -*-
import os
import functools
import numpy as np
import tensorflow as tf
import pandas as pd
rng = np.random.seed(100)
from sklearn.model_selection import train_test_split
from sklearn import metrics
from skimage import io, transform, exposure, filters, color
def misc_seg_metrics(pred, ann):
""" Compute the spec, sens, acc between two ndarrays.
"""
#if len(np.unique(pred)) == 2:
acc = [((pred==j)*(ann==j)).sum(dtype='float')/(ann==j).sum() for j in np.unique(ann)]
acc.append((pred==ann).sum()/float(ann.size))
return acc
def misc_jaccard_index(pred, ann):
intersection = pred*ann
union = np.maximum(pred, ann)
return np.sum(intersection.astype(float))/np.sum(union)
def misc_metrics(gt, pr, ma=None, thrs= 0.5):
assert(gt.size == pr.size)
if ma is None:
ma = np.ones(gt.shape, dtype=bool)
ma = np.ravel(ma).astype(bool)
gt = np.ravel(gt)
pr = np.ravel(pr)
gt = gt[ma]
pr = pr[ma]
auc = metrics.roc_auc_score(gt, pr)
p = (pr>=thrs).astype(int)
mets = misc_seg_metrics(p, gt)
f1 = metrics.f1_score(gt, p)
jac = misc_jaccard_index(p, gt)
return [auc, mets[0], mets[1], mets[2], f1, jac]
def read_df(fpath, data_dir):
df_train = pd.read_csv(fpath)
x_paths = df_train['im_name'].map(lambda s: os.path.join(data_dir,s))
y_paths = df_train['gt_name'].map(lambda s: os.path.join(data_dir,s))
return x_paths, y_paths
### Training-Validation splits
def trainval_splits(ftrain, data_dir, validation_split=0.2, random_state=rng):
x_train_paths, y_train_paths = read_df(ftrain, data_dir)
return train_test_split(x_train_paths, y_train_paths,
test_size=validation_split,
random_state=random_state)
def _process_pathnames(fname, lname, resize=None):
img = io.imread(fname)
gt = io.imread(lname)
if gt.ndim < 3:
gt = np.expand_dims(gt, -1)
gt = gt[...,:1]
gt = (gt > 0).astype(int) # binarize the ground-truth
if resize is not None:
img = transform.resize(img, resize)
gt = transform.resize(gt, resize)
gt = gt >= filters.threshold_otsu(gt)
return img, gt
### Data augmentation routines
def shift_img(img, gt, width_shift_range, height_shift_range, rotate_range):
if width_shift_range or height_shift_range:
if width_shift_range:
width_shift_range = np.random.uniform(-width_shift_range * img.shape[1],
width_shift_range * img.shape[1])
if height_shift_range:
height_shift_range = np.random.uniform(-height_shift_range * img.shape[0],
height_shift_range * img.shape[0])
tr = transform.AffineTransform(translation=(width_shift_range, height_shift_range ))
img = transform.warp(img, tr, preserve_range=True)
gt = transform.warp(gt, tr, preserve_range=True)
if rotate_range :
if isinstance(rotate_range, np.ScalarType):
degre = np.random.uniform(-rotate_range,rotate_range)
else:
degre = np.random.uniform(rotate_range[0], rotate_range[1])
img = transform.rotate(img, degre, preserve_range=True)
gt = transform.rotate(gt, degre, preserve_range=True)
return img, gt
def flip_img(img, gt, horizontal_flip, vertical_flip):
if horizontal_flip:
flip_prob = np.random.uniform(0.0, 1.0)
img, gt = (img, gt) if flip_prob >= 0.5 else (np.flip(img, 1), np.flip(gt, 1))
if vertical_flip:
flip_prob = np.random.uniform(0.0, 1.0)
img, gt = (img, gt) if flip_prob >= 0.5 else (np.flip(img, 0), np.flip(gt, 0))
return img, gt
def _process_img(img, gt, gamma=0,
clahe=False, gray=False, xyz=False, hed=False,
horizontal_flip=False, width_shift_range=0,
height_shift_range=0, vertical_flip=0, rotate_range=0):
img = exposure.rescale_intensity(img.astype(float), out_range=(0,1))
if gray:
img = color.rgb2gray(img)
if xyz:
img = color.rgb2xyz(img)
if hed:
img = color.rgb2hed(img)
img = exposure.rescale_intensity(img, out_range=(0,1))
if clahe:
img = exposure.equalize_adapthist(img)
if gamma:
img = exposure.adjust_gamma(img, gamma)
img = exposure.rescale_intensity(img, out_range=(0,1))
if img.ndim == 2:
img = np.expand_dims(img, -1)
img, gt = flip_img(img, gt, horizontal_flip, vertical_flip)
img, gt = shift_img(img, gt, width_shift_range, height_shift_range, rotate_range)
return img, gt
def image_generator(im_paths, gt_paths,
reader_fn=functools.partial(_process_pathnames),
preproc_fn=functools.partial(_process_img),
batch_size=1,
MAX_IM_QUEUE=20):
batch_x = []
batch_y = []
im_stack = dict()
while True:
for im_path, gt_path in zip(im_paths, gt_paths) :
hash_im = hash(im_path)
if not im_stack.has_key(hash_im):
img, gt = reader_fn(im_path, gt_path)
if len(im_stack.keys()) > MAX_IM_QUEUE:
im_stack.popitem()
im_stack[hash_im] = (img, gt)
else:
img, gt = im_stack[hash_im]
pr_im, pr_gt = preproc_fn(img, gt)
if len(batch_x) < batch_size:
batch_x.append(pr_im)
batch_y.append(pr_gt)
else:
ret = (np.array(batch_x), np.array(batch_y))
batch_x, batch_y = [pr_im], [pr_gt]
yield ret
def get_image_generator(x_train_paths, y_train_paths, batch_size=1,
width_shift_range=0, height_shift_range=0,
horizontal_flip=False,vertical_flip=False,
rotate_range=0, resize=None,
gamma=0, clahe=False, gray=False, xyz=False, hed=False,
MAX_IM_QUEUE=100):
prepro_cfg = dict(gamma=gamma, clahe=clahe,
gray=gray, xyz=xyz, hed=hed, horizontal_flip=horizontal_flip,
vertical_flip=vertical_flip, width_shift_range=width_shift_range,
height_shift_range=height_shift_range)
prepro_fn = functools.partial(_process_img, **prepro_cfg)
reader_cfg = dict(resize=resize)
reader_fn = functools.partial(_process_pathnames, **reader_cfg)
return image_generator(x_train_paths, y_train_paths, reader_fn=reader_fn,
preproc_fn=prepro_fn, batch_size=batch_size, MAX_IM_QUEUE=MAX_IM_QUEUE)
def fixed_patch_ids_creation(im_paths, gt_paths, spatial_shape=None,
p_stride=16, shuffle=True, per_label=0, mask=None):
all_ids = []
mask = mask if mask is not None else 1
for im_path, gt_path in zip(im_paths, gt_paths) :
if p_stride > 0:
ids = np.zeros(spatial_shape, dtype='int')
if ids.ndim == 2:
ids[0::p_stride, 0::p_stride] = 1
else:
ids[0::p_stride, 0::p_stride, 0::p_stride] = 1
ids = ids * mask
ids = np.array(np.nonzero(ids)).T
n = len(ids)
ap = np.c_[np.expand_dims([im_path]*n, -1), np.expand_dims([gt_path]*n, -1), ids]
all_ids.extend(ap)
if per_label >0:
# Adding samples based on the classes distribution.
_, gt = _process_pathnames(im_path, gt_path, resize=spatial_shape)
cls_ids = []
for c in np.unique(gt):
search_area = np.nonzero((np.squeeze(gt) == c) * mask)
if len(search_area[0]) == 0:
continue
search_area = np.array(search_area).T
search_area = np.random.permutation(search_area)
cls_ids.append(search_area[:per_label])
cls_ids = np.concatenate([x for x in cls_ids])
n = len(cls_ids)
ap = np.c_[[im_path]*n, [gt_path]*n, cls_ids]
all_ids.extend(ap)
all_ids = np.array(all_ids)
if shuffle:
np.random.shuffle(all_ids)
return all_ids
class Patch_Sequence(tf.keras.utils.Sequence):
def __init__(self, fixed_patch_ids, p_shape=(32,32,3),
reader_fn=functools.partial(_process_pathnames),
preproc_fn=functools.partial(_process_img),
batch_size=32,
MAX_IM_QUEUE=20, unsup=False):
self.ids = fixed_patch_ids #
self.p_shape = p_shape
self.batch_size = batch_size
self.reader_fn = reader_fn
self.preproc_fn = preproc_fn
self.MAX_IM_QUEUE = MAX_IM_QUEUE
self.im_stack = {}
self.unsup = unsup
def __len__(self):
return int(np.ceil(len(self.ids) / float(self.batch_size)))
def __getitem__(self, idx):
cur_id = self.ids[idx * self.batch_size:(idx + 1) * self.batch_size]
batch_x = []
batch_y = []
for pos in cur_id:
pid, pim, pgt = pos[2:], pos[0], pos[1]
x_p, y_p = pid.astype(int)
hash_im = hash(pim)
if not self.im_stack.has_key(hash_im):
img, gt = self.reader_fn(pim, pgt)
img, gt = self.preproc_fn(img, gt)
img = np.pad(img, ((self.p_shape[0]//2,), (self.p_shape[1]//2,), (0,)), mode='reflect')
gt = np.pad(gt, ((self.p_shape[0]//2,), (self.p_shape[1]//2,), (0,)), mode='reflect')
if len(self.im_stack.keys()) > self.MAX_IM_QUEUE:
self.im_stack.popitem()
self.im_stack[hash_im] = (img, gt)
else:
img, gt = self.im_stack[hash_im]
patch = img[x_p:x_p+self.p_shape[0], y_p:y_p+self.p_shape[1]]
label = gt[x_p:x_p+self.p_shape[0], y_p:y_p+self.p_shape[1]]
batch_x.append(patch)
batch_y.append(label)
if self.unsup:
return np.array(batch_x), [np.array(batch_y), np.array(batch_x)]
return np.array(batch_x), np.array(batch_y)
def on_epoch_end(self, epoch=None, logs=None):
self.im_stack = {}
def get_patch_generator(dataset_ids, p_shape, batch_size=1, gamma=0.,
clahe=False, gray=False, xyz=False, hed=False,
width_shift_range=0, height_shift_range=0,
horizontal_flip=False,vertical_flip=False,
rotate_range=0, resize=None,
MIN_PATCH_STD=None, MAX_IM_QUEUE=100,
return_keras_queuer=False):
prepro_cfg = dict(gamma=gamma, horizontal_flip=horizontal_flip,
vertical_flip=vertical_flip, width_shift_range=width_shift_range,
height_shift_range=height_shift_range, clahe=clahe, gray=gray, xyz=xyz, hed=hed)
prepro_fn = functools.partial(_process_img, **prepro_cfg)
reader_cfg = dict(resize=resize)
reader_fn = functools.partial(_process_pathnames, **reader_cfg)
return Patch_Sequence(dataset_ids, p_shape=p_shape,
reader_fn=reader_fn, preproc_fn=prepro_fn,
batch_size=batch_size, MAX_IM_QUEUE=MAX_IM_QUEUE)