-
Notifications
You must be signed in to change notification settings - Fork 3
/
utils.py
198 lines (146 loc) · 6.49 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
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
########################################################################################################################
# @author Oriol Aranda (https://github.com/oriolaranda/)
# @date Oct 2021
########################################################################################################################
import tensorflow as tf
from ray import tune
###############################
# Metrics and Loss functions
###############################
def dice_coeff(seg_true, seg_pred, smooth=0.01):
seg_true_flat = tf.keras.layers.Flatten(data_format='channels_first', dtype=tf.float32)(seg_true)
seg_pred_flat = tf.keras.layers.Flatten(data_format='channels_first', dtype=tf.float32)(seg_pred)
intersection = tf.multiply(seg_true_flat, seg_pred_flat, name="intersection")
intersect = tf.reduce_sum(intersection)
dice = tf.divide(2 * intersect + smooth, tf.reduce_sum(seg_true_flat) + tf.reduce_sum(seg_pred_flat) + smooth)
return dice
def dice_loss(seg_true, seg_pred):
return 1 - dice_coeff(seg_true, seg_pred)
def loss_metric(loss, metric):
@tf.autograph.experimental.do_not_convert
def loss_func(seg_true, seg_pred):
return loss(seg_true, seg_pred)
@tf.autograph.experimental.do_not_convert
def metric_func(seg_true, seg_pred):
return metric(seg_true, seg_pred)
return loss_func, metric_func
##############
# Data utils
##############
@tf.autograph.experimental.do_not_convert
def parse_raw_sample(data_shape):
def _parse_raw_sample(sample):
parse_dic = {
'img': tf.io.FixedLenFeature([], tf.string),
'label': tf.io.FixedLenFeature([], tf.string)
}
sample_message = tf.io.parse_single_example(sample, parse_dic)
img = sample_message['img'] # get byte string
label = sample_message['label']
img = tf.io.parse_tensor(img, out_type=tf.float32) # restore 2D array from byte string
label = tf.io.parse_tensor(label, out_type=tf.float32)
img = tf.ensure_shape(img, (4, *tuple(data_shape)))
label = tf.ensure_shape(label, (1, *tuple(data_shape)))
return img, label
return _parse_raw_sample
def data_generator(filenames, config, shuffle_size, train=False):
raw_dataset = tf.data.TFRecordDataset(filenames, # buffer_size=config['tfrecord_buffer_size'],
num_parallel_reads=config['tfrecord_num_parallel_reads'])
parsed_dataset = raw_dataset.map(parse_raw_sample(config['input_shape']),
num_parallel_calls=config['map_num_parallel_calls'])
if train:
parsed_dataset = parsed_dataset.shuffle(shuffle_size).repeat().batch(config['batch_size'])
else:
parsed_dataset = parsed_dataset.repeat(1).batch(config['batch_size'])
parsed_dataset = parsed_dataset.prefetch(config['prefetch_buffer_size'])
return parsed_dataset
##############
# Callbacks
##############
class TuneReporter(tf.keras.callbacks.Callback):
"""Tune Callback for Keras."""
def __init__(self, reporter=None, freq="epoch", logs=None):
super(TuneReporter, self).__init__()
def on_epoch_end(self, epoch, logs=None):
tune.report(keras_info=logs, loss=logs['loss'], dice_coefficient=logs['metric_func'])
def on_test_end(self, logs=None):
tune.report(keras_info=logs, val_loss=logs['loss'],
val_dice_coefficient=logs['metric_func'])
################################################
# Data Augmentation and Preprocessing functions
################################################
def _resize_by_axis_trilinear(images, size_0, size_1, ax):
"""
Resize image bilinearly to [size_0, size_1] except axis ax.
:param image: a tensor 4-D with shape
[batch, d0, d1, d2, channels]
:param size_0: size 0
:param size_1: size 1
:param ax: axis to exclude from the interpolation
"""
resized_list = []
# unstack the image in 2d cases
unstack_list = tf.unstack(images, axis=ax)
for i in unstack_list:
resized_list.append(tf.image.resize(i, [size_0, size_1])) # resize bilinearly
stack_img = tf.stack(resized_list, axis=ax)
return stack_img
def resize_trilinear(images, size):
"""
Resize images to size using trilinear interpolation.
:param images: A tensor 5-D with shape
[batch, d0, d1, d2, channels]
:param size: A 1-D int32 Tensor of 3 elements: new_d0, new_d1,
new_d2. The new size for the images.
"""
assert size.shape[0] == 3
x_ = tf.expand_dims(images, axis=0)
x_ = tf.transpose(x_, [0, 2, 3, 4, 1])
resized = _resize_by_axis_trilinear(x_, size[0], size[1], 3)
resized = _resize_by_axis_trilinear(resized, size[0], size[2], 1)
resized = tf.transpose(resized, [0, 4, 1, 2, 3])
resized = tf.squeeze(resized, axis=[0])
return resized
def resize_image(x, y, resize_shape):
x_ = resize_trilinear(x, tf.constant(list(resize_shape)))
y_ = resize_trilinear(y, tf.constant(list(resize_shape)))
return x_, y_
def data_augmentation(flip=False, noise=False, rotation=False, p=0.1):
"""Apply data augmentation: add noise, flip nad rotation"""
print("Data Augmentation:", flip, noise, p)
def _apply_transforms(*sample):
probs = tf.random.uniform([5])
if noise and probs[0] < p:
sample = _add_noise(*sample)
if flip:
if probs[1] < p:
sample = _flip_up_down(*sample)
if probs[2] < p:
sample = _flip_left_right(*sample) # the most sense flip
if probs[3] < p:
sample = _flip_front_back(*sample)
if rotation and probs[4] < p:
pass
img, label = sample
return img, label
return _apply_transforms
@tf.function
def _flip(img, label, axis):
flipped_img = tf.reverse(img, axis=[axis])
flipped_label = tf.reverse(label, axis=[axis])
return flipped_img, flipped_label
def _flip_up_down(img, seg):
return _flip(img, seg, 3)
def _flip_left_right(img, label):
return _flip(img, label, 1)
def _flip_front_back(img, label):
return _flip(img, label, 2)
@tf.function
def _add_noise(img, label):
mean, std = tf.nn.moments(img, axes=[0, 1, 2, 3])
noise = tf.random.normal(shape=img.shape, mean=0.0, stddev=0.2, dtype=tf.float32)
# non_zero = img > 0
# _img = tf.where(non_zero, img + noise, img)
_img = img + noise
st_img = tf.image.per_image_standardization(_img)
return st_img, label