[1ee192]: / core / layers.py

Download this file

285 lines (221 with data), 11.1 kB

  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
# -*- coding: utf-8 -*-
"""
Auxiliary functions and operations for network construction, some of which have
been deprecated for high-level modules in TensorFlow.
@author: Xinzhe Luo
"""
from __future__ import print_function, division, absolute_import, unicode_literals
import tensorflow as tf
import numpy as np
def weight_variable(shape, name="weight"):
fan_in, fan_out = shape[-2:]
low = -1*np.sqrt(6.0/(fan_in + fan_out)) # use 4 for sigmoid, 1 for tanh activation
high = 1*np.sqrt(6.0/(fan_in + fan_out))
return tf.Variable(tf.random_uniform(shape, minval=low, maxval=high, dtype=tf.float32), name=name)
def weight_variable_devonc(shape, name="weight_deconv"):
fan_in, fan_out = shape[-2:]
low = -1*np.sqrt(6.0/(fan_in + fan_out)) # use 4 for sigmoid, 1 for tanh activation
high = 1*np.sqrt(6.0/(fan_in + fan_out))
return tf.Variable(tf.random_uniform(shape, minval=low, maxval=high, dtype=tf.float32), name=name)
def bias_variable(shape, name="bias"):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial, name=name)
def conv2d(x, W, keep_prob_):
conv_2d = tf.nn.conv3d(x, W, strides=[1, 1, 1, 1], padding='SAME')
return tf.nn.dropout(conv_2d, keep_prob_)
def deconv2d(x, w, stride):
x_shape = tf.shape(x)
output_shape = tf.stack([x_shape[0], x_shape[1]*2, x_shape[2]*2, x_shape[3]//2])
return tf.nn.conv2d_transpose(x, w, output_shape, strides=[1, stride, stride, 1], padding='VALID')
def max_pool2d(x, n):
return tf.nn.max_pool(x, ksize=[1, n, n, 1], strides=[1, n, n, 1], padding='VALID')
'''
def batch_norm(x, train_phase):
x_norm = tf.layers.batch_normalization(x, axis=0, training=train_phase)
return x_norm
'''
def batch_norm(x, name_scope, training, size, epsilon=1e-3, decay=0.999):
"""
Assume 4d [batch_size, ny, nx, feature_size] tensor
size = output feature size
"""
with tf.variable_scope(name_scope):
scale = tf.get_variable('scale', [size], initializer=tf.constant_initializer(0.1))
offset = tf.get_variable('offset', [size])
pop_mean = tf.get_variable('pop_mean', [size], initializer=tf.zeros_initializer, trainable=False)
pop_var = tf.get_variable('pop_var', [size], initializer=tf.ones_initializer, trainable=False)
batch_mean, batch_var = tf.nn.moments(x, [0, 1, 2])
train_mean_op = tf.assign(pop_mean, pop_mean * decay + batch_mean * (1 - decay))
train_var_op = tf.assign(pop_var, pop_var * decay + batch_var * (1 - decay))
def batch_statistics():
with tf.control_dependencies([train_mean_op, train_var_op]):
return tf.nn.batch_normalization(x, batch_mean, batch_var, offset, scale, epsilon)
def population_statistics():
return tf.nn.batch_normalization(x, pop_mean, pop_var, offset, scale, epsilon)
return tf.cond(training, batch_statistics, population_statistics)
def crop_and_concat(x1, x2):
"""
Crop x1 to match the size of x2 and concatenate them.
"""
x1_shape = tf.shape(x1)
x2_shape = tf.shape(x2)
# offsets for the top left corner of the crop
offsets = [0, (x1_shape[1]-x2_shape[1])//2, (x1_shape[2]-x2_shape[2])//2, 0]
size = [-1, x2_shape[1], x2_shape[2], -1]
x1_crop = tf.slice(x1, offsets, size, name='crop')
crop_concat = tf.concat([x1_crop, x2], -1, name='crop_concat')
crop_concat.set_shape([None, None, None, x1.get_shape().as_list()[-1] + x2.get_shape().as_list()[-1]])
return crop_concat
def crop_and_add(x1, x2):
"""
Crop x1 to match the size of x2 and add them together.
"""
x1_shape = tf.shape(x1)
x2_shape = tf.shape(x2)
# offsets for the top left corner of the crop
offsets = [0, (x1_shape[1]-x2_shape[1])//2, (x1_shape[2]-x2_shape[2])//2, 0]
size = [-1, x2_shape[1], x2_shape[2], -1]
x1_crop = tf.slice(x1, offsets, size, name='crop')
return tf.add(x1_crop, x2, name='crop_add')
def pad_and_concat(x1, x2):
"""
Pad x2 to match the size of x1 and concatenate them.
"""
x1_shape = tf.shape(x1)
x2_shape = tf.shape(x2)
offsets = [0, (x1_shape[1]-x2_shape[1])//2, (x1_shape[2]-x2_shape[2])//2, 0]
paddings = [[0, 0],
[offsets[1], x1_shape[1]-x2_shape[1]-offsets[1]],
[offsets[2], x1_shape[2]-x2_shape[2]-offsets[2]],
[0, 0]]
x2_pad = tf.pad(x2, paddings, name='pad')
pad_concat = tf.concat([x1, x2_pad], -1, name='pad_concat')
pad_concat.set_shape([None, None, None, x1.get_shape().as_list()[-1] + x2.get_shape().as_list()[-1]])
return pad_concat
def pad_and_add(x1, x2):
"""
Pad x2 to match the size of x1 and add them together.
"""
x1_shape = tf.shape(x1)
x2_shape = tf.shape(x2)
offsets = [0, (x1_shape[1]-x2_shape[1])//2, (x1_shape[2]-x2_shape[2])//2, 0]
paddings = [[0, 0],
[offsets[1], x1_shape[1]-x2_shape[1]-offsets[1]],
[offsets[2], x1_shape[2]-x2_shape[2]-offsets[2]],
[0, 0]]
x2_pad = tf.pad(x2, paddings, name='pad')
return tf.add(x1, x2_pad, name='pad_add')
def crop_to_tensor(x1, x2):
"""
Crop tensor x1 to match the shape of x2.
"""
x1_shape = tf.shape(x1)
x2_shape = tf.shape(x2)
# offsets for the top left corner of the crop
offsets = [0, (x1_shape[1]-x2_shape[1])//2, (x1_shape[2]-x2_shape[2])//2, 0]
size = [-1, x2_shape[1], x2_shape[2], -1]
x1_crop = tf.slice(x1, offsets, size, name='crop')
x1_crop.set_shape([None, None, None, x1.get_shape().as_list()[-1]])
return x1_crop
def crop_to_tensor_3d(x1, x2):
"""
Crop tensor x1 to match the shape of x2.
"""
x1_shape = tf.shape(x1)
x2_shape = tf.shape(x2)
# offsets for the top left corner of the crop
offsets = [0, (x1_shape[1]-x2_shape[1])//2, (x1_shape[2]-x2_shape[2])//2, (x1_shape[3]-x2_shape[3])//2, 0]
size = [-1, x2_shape[1], x2_shape[2], x2_shape[3], -1]
x1_crop = tf.slice(x1, offsets, size, name='crop')
x1_crop.set_shape([None, None, None, None, x1.get_shape().as_list()[-1]])
return x1_crop
def pad_to_tensor(x1, x2):
"""
Pad tensor x1 to match the shape of x2.
"""
x1_shape = tf.shape(x1)
x2_shape = tf.shape(x2)
offsets = [0, (x2_shape[1] - x1_shape[1]) // 2, (x2_shape[2] - x1_shape[2]) // 2, 0]
paddings = [[0, 0],
[offsets[1], x2_shape[1] - x1_shape[1] - offsets[1]],
[offsets[2], x2_shape[2] - x1_shape[2] - offsets[2]],
[0, 0]]
x1_pad = tf.pad(x1, paddings, name='pad')
return x1_pad
def pixel_wise_softmax(output_map):
"""
deprecated function for tf.nn.softmax
"""
exponential_map = tf.exp(output_map)
evidence = tf.add(exponential_map, tf.reverse(exponential_map, [False, False, False, True]))
return tf.divide(exponential_map, evidence, name="pixel_wise_softmax")
def pixel_wise_softmax_2(output_map):
"""
deprecated function for tf.nn.softmax
"""
exponential_map = tf.exp(output_map)
sum_exp = tf.reduce_sum(exponential_map, -1, keepdims=True)
tensor_sum_exp = tf.tile(sum_exp, tf.stack([1, 1, 1, tf.shape(output_map)[-1]]))
return tf.clip_by_value(tf.divide(exponential_map, tensor_sum_exp), 1e-10, 1.0)
def cross_entropy_map(labels, probs):
"""
Compute the element-wise cross-entropy map by clipping the values of softmax probabilities to avoid Nan loss.
:param labels: ground-truth value using one-hot representation
:param probs: probability map as the output of softmax
:return: A tensor of the same shape as lables and of the same shape as probs with the cross entropy loss.
"""
return tf.reduce_sum(- labels * tf.log(tf.clip_by_value(probs, 1e-10, 1.0)), axis=-1, name="cross_entropy_map")
def balance_weight_map(flat_labels):
"""
:param flat_labels: masked ground truth tensor in shape [-1, n_class]
:return the balance weight map in 1-D tensor
"""
n = tf.shape(flat_labels)[0]
return tf.reduce_sum(tf.multiply(flat_labels, tf.tile(1 / tf.reduce_sum(flat_labels, axis=0, keepdims=True),
[n, 1])), axis=-1, name='balance_weight_map')
def gaussian_noise_layer(input_layer, std):
"""
Apply Gaussian noise to the input.
:param input_layer: Inputs.
:param std: Standard deviation for the noise.
:return: Blurred features.
"""
noise = tf.random_normal(shape=tf.shape(input_layer), mean=0.0, stddev=std, dtype=tf.float32)
return input_layer + noise
def linear_additive_upsample(input_tensor, new_size=2, n_split=4):
"""
Apply linear additive up-sampling layer, described in paper Wojna et al., The devil is in the decoder,
https://arxiv.org/abs/1707.05847.
:param input_tensor: Input tensor.
:param new_size: The factor of up-sampling.
:param n_split: The n_split consecutive channels are added together.
:return: Linearly additively upsampled feature maps.
"""
with tf.name_scope('linear_additive_upsample'):
n_channels = input_tensor.get_shape().as_list()[-1]
input_dim = input_tensor.get_shape().ndims
assert n_split > 0 and n_channels % n_split == 0, "Number of feature channels should be divisible by n_splits."
if input_dim == 4:
upsample = tf.keras.layers.UpSampling2D(size=new_size, name='upsample')(input_tensor)
elif input_dim == 5:
upsample = tf.keras.layers.UpSampling3D(size=new_size, name='upsample')(input_tensor)
else:
raise TypeError('Incompatible input spatial rank: %d' % input_dim)
split = tf.split(upsample, n_split, axis=-1)
split_tensor = tf.stack(split, axis=-1)
output_tensor = tf.reduce_sum(split_tensor, axis=-1, name='output_tensor')
return output_tensor
def residual_additive_upsample(inputs, filter_size, strides, feature_size, n_split, regularizer, train_pahse, trainable,
name_or_scope='residual_additive_upsample'):
n_channel = inputs.get_shape().as_list()[-1]
assert n_channel == feature_size * n_split, "The number of input channels must be the product of output feature " \
"size and the number of splits."
with tf.variable_scope(name_or_scope):
deconv = tf.layers.conv2d_transpose(inputs, filters=feature_size, kernel_size=filter_size, strides=strides,
padding='same', use_bias=False,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
kernel_regularizer=regularizer, trainable=trainable, name='deconv')
bn = tf.layers.batch_normalization(deconv, training=train_pahse, trainable=trainable, name='bn')
relu = tf.nn.relu(bn, name='relu')
upsample = linear_additive_upsample(inputs, strides, n_split)
return tf.add(relu, upsample, name='res_upsample')