-
Notifications
You must be signed in to change notification settings - Fork 10
/
unet_model_helpers.py
335 lines (266 loc) · 13.2 KB
/
unet_model_helpers.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
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
import tensorflow as tf
def output_layer(inputs, depth):
"""
convert the unscaled inputs to label maps
Args:
inputs (4d tensor [float]) : unscaled inputs
Returns:
output (4d tensor [int]) : label map
"""
with tf.variable_scope('output_layer'):
#if it is binary classification
if depth == 1:
#scale the unscaled inputs to 0 - 1
probs = tf.nn.sigmoid(inputs)
#convert the probability map to label map by simply thresholding
output = tf.cast(probs >= 0.5, tf.uint8)
else:
#scale the unscaled inputs to 0 - 1
probs = tf.nn.softmax(inputs)
#convert the probability map to label map by selecting index of the highest probability
#note: expand_dims function is used to change shape of "output" from [batch_size, height, width] to [batch_size, height, width, 1]
output = tf.expand_dims(tf.argmax(probs, axis = -1), -1)
return output
def calc_loss(logits, labels, depth):
"""
flatten logits and labels to 2D tensors, where dimensions are [batch_size x height x width, depth]
calculate the loss with the following formula:
loss = [cross entropy loss] + [intersection over union loss]
Args:
logits (matrix [float]) : unscaled output generated by the network, dims: [batch_size, height, width, depth]
labels (matrix [float]) : groun-truth, dims: [batch_size, height, width, depth]
depth (int) : depth of the label layer (1 for binary classification, num_of_classes for multi-label classification)
Returns:
loss (float) : loss
"""
with tf.variable_scope('seg_loss_layer'):
#flatten logits and labels
logits_flat = tf.reshape(logits, [-1, depth])
labels_flat = tf.reshape(labels, [-1, depth])
cross_entropy_loss = calc_cross_ent_loss(logits_flat, labels_flat, depth)
iou_loss = calc_iou_loss(logits_flat, labels_flat, depth)
loss = iou_loss + cross_entropy_loss
return loss
def calc_iou_loss(logits_flat, labels_flat, depth):
"""
calculate intersection over union loss
unscaled scores need to be converted to probability distribution using sigmoid or softmax depending on
number of classes. If it a binary classification, use sigmoid. Use softmax if it is multi-class classification
Args:
logits (matrix [float]) : flattened version of the unscaled output generated by the network
labels (matrix [float]) : flattened verison of the ground-truth
depth (int) : depth of the label layer (1 for binary classification, num_of_classes for multi-label classification)
Returns:
loss (float) : scalar loss
"""
#convert unscaled output generated by the network to probs
if depth == 1:
probs_flat = tf.nn.sigmoid(logits_flat)
# #probs and labels for both foreground and background classes
# # probs_flat = tf.concat([probs_flat, tf.subtract(tf.constant(1.0), probs_flat)], axis = 1)
# # labels_flat = tf.concat([labels_flat, tf.subtract(tf.constant(1.0), labels_flat)], axis = 1)
else:
probs_flat = tf.nn.softmax(logits_flat)
#calculate intersection over union loss
with tf.variable_scope('iou_loss'):
#calculate intersection of probs_flat and labels_flat (pixelwise multiplication)
inter = tf.multiply(probs_flat, labels_flat)
#calculate union of probs_flat and labels_flat
union = tf.subtract(tf.add(probs_flat, labels_flat), inter)
#sum each column of inter and union
inter_sum = tf.reduce_sum(inter)
union_sum = tf.reduce_sum(union)
inter_sum += 1e-16
union_sum += 1e-16
loss = tf.multiply(tf.constant(-1.0), tf.log(tf.divide(inter_sum, union_sum)))
return loss
def iou_single(preds,labels,depth):
inter = tf.reduce_sum(tf.multiply(labels,preds))
union = tf.subtract(tf.reduce_sum(tf.add(labels, preds)), inter)
return tf.divide(inter, union + 1e-16)
def iou(preds, labels ,batch_size,depth):
labels = tf.expand_dims(tf.argmax(labels, axis = -1), -1)
iou,_ = tf.metrics.mean_iou(
labels,
preds,
depth)
return iou
# return tf.reduce_mean([iou_single(logits[i,:,:,:],labels[i,:,:,:],depth) for i in range(batch_size)])
def calc_cross_ent_loss(logits_flat, labels_flat, depth):
"""
if it is a binary classification, calculate sigmoid cross entropy loss given the logits and the ground-truth
otherwise, calculate softmax cross entropy loss given the logits and the ground-truth
Args:
logits (matrix [float]) : flattened version of the unscaled output generated by the network
labels (matrix [float]) : flattened verison of the ground-truth
depth (int) : depth of the label layer (1 for binary classification
Returns:
loss (float) : scalar loss
"""
with tf.variable_scope('cross_ent_loss'):
if depth == 1:
loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels = labels_flat, logits = logits_flat))
else:
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels = labels_flat, logits = logits_flat))
return loss
def conv_block(inputs, filters, kernel_size, strides, training, scope_name,bn=True):
"""
2d convolution block
Args:
inputs (4d tensor [float]) : input 4d tensor
filters (int) : number of output filters
kernel_size (int) : size of the kernel for the convolution
strides (int) : strides for the convolution
training (bool) : True = training, False = test
scope_name (str) : name of the block
Returns:
output (4d tensor [float]) : output
"""
with tf.variable_scope(scope_name):
logits = conv2d(inputs = inputs, filters = filters, kernel_size = kernel_size, strides = strides, conv_name = 'conv2d')
if bn:
logits_bn = tf.layers.batch_normalization(logits, fused = True, axis = 1, training = training)
output = tf.nn.relu(logits_bn)
# output = tf.nn.elu(logits_bn)
else:
output = tf.nn.relu(logits)
# output = tf.nn.elu(logits)
return output
def conv2d(inputs, filters, kernel_size, strides, conv_name):
"""
2d convolution
Args:
inputs (4d tensor [float]) : input 4d tensor
filters (int) : number of output filters
kernel_size (int) : soze of the kernel for the convolution
strides (int) : strides for the convolution
conv_name (str) : name of the convolution operation
"""
logits = tf.layers.conv2d(inputs = inputs, filters = filters, kernel_size = kernel_size, strides = strides,
data_format = 'channels_first', padding = 'same', name = conv_name)
return logits
def upsample_concat(inputs1, inputs2, num_of_channels_reduce_factor, training, scope_name):
"""
double height and width, reduce number of channels
concatenate inputs1 and upsampled version of inputs2
Args:
inputs1 (4d tensor [float]) : input that would be concatenated with inputs2
inputs2 (4d tensor [float]) : input that would be upsampled and concatenated with inputs1
training (1d tensor [bool]) : True = training, False = test
num_of_channels_reduce_factor (int) : 2 = # of channels is halved
4 = # of channels is divided by 4
scope_name (str) : name of the upsampling layer
Returns:
output (4d tensor [float]) : output
"""
num_of_filters2 = inputs2.get_shape().as_list()[1]
with tf.variable_scope(scope_name):
# logits = tf.layers.conv2d_transpose(inputs = inputs2, filters = num_of_filters2 // num_of_channels_reduce_factor,
# kernel_size = 2, strides = (2, 2),
# data_format = 'channels_first', padding = 'same', name = 'deconv')
inputs2 = tf.transpose(inputs2, [0, 2, 3, 1])
x_shape = inputs2.shape
new_size = (x_shape[1]*2, x_shape[2]*2)
upsampled = tf.image.resize_images(inputs2, new_size,method=tf.image.ResizeMethod.NEAREST_NEIGHBOR)
upsampled = tf.transpose(upsampled, [0, 3, 1, 2])
# if bn:
# logits_bn = tf.layers.batch_normalization(logits, fused = True, axis = 1, training = training)
# inputs2_upsampled = tf.nn.relu(logits_bn)
# else:
# inputs2_upsampled = tf.nn.relu(logits)
#concat along the first dimension
output = tf.concat([upsampled, inputs1], axis = 1)
return output
def deconv(inputs,filters,training, scope_name,bn=True):
with tf.variable_scope(scope_name):
logits = tf.layers.conv2d_transpose(inputs = inputs, filters = filters,
kernel_size = 2, strides = (2, 2),
data_format = 'channels_first', padding = 'same', name = 'deconv')
if bn:
logits_bn = tf.layers.batch_normalization(logits, fused = True, axis = 1, training = training)
upsampled = tf.nn.relu(logits_bn)
else:
upsampled = tf.nn.relu(logits)
return upsampled
def conv_block_sequence(inputs, filters, num_of_conv_blocks, training, scope_name ,kernel_size=3,bn=True):
"""
two consecutive convolutions in Unet model
Args:
inputs (4d tensor [float]) : input 4d tensor
filters (int) : number of output filters for the first and second convolutions
num_of_conv_blocks (int) : number of convolutional blocks in a row
training (bool) : True = training, False = test
scope_name (str) : name of the sequence
Returns:
layer2_output (4d tensor [float]) : output
"""
strides = (1, 1)
outputs = inputs
with tf.variable_scope(scope_name):
#apply convolution blocks in a row
for conv_block_no in range(1, num_of_conv_blocks + 1):
outputs = conv_block(outputs, filters, kernel_size, strides, training, 'conv_' + str(conv_block_no),bn=bn)
return outputs
def max_pool(inputs, scope_name):
"""
Pooling operation that reduces width and height of the input layer to half
Args:
inputs (4d tensor [float]) : input 4d tensor
scope_name (str) : name of the pooling layer
Returns:
output (4d tensor [float]) : output
"""
with tf.variable_scope(scope_name):
output = tf.layers.max_pooling2d(inputs = inputs, pool_size = (2, 2), strides = (2, 2), data_format='channels_first')
return output
def calc_loss_p(logits, labels, batch_size,depth):
with tf.variable_scope('pansharpen_loss_layer'):
probs = tf.nn.sigmoid(logits)
return (1.-Q(labels, probs,batch_size,depth)) + (1.- psnr(labels, probs)/ 50.)
# return tf.losses.mean_squared_error(labels,tf.nn.sigmoid(logits))
def __conv(input,filter):
return tf.nn.conv2d(
input,
filter,
[1,1,1,1],
"SAME",
use_cudnn_on_gpu=True,
data_format='NHWC',
dilations=[1, 1, 1, 1],
name=None)
def q_fs(img1,img2):
BLOCK_SIZE = 8
patch_h =patch_w =256
N = BLOCK_SIZE**2
sum2_filter = tf.ones((BLOCK_SIZE,BLOCK_SIZE,1,1))
img1_sq = img1*img1;
img2_sq = img2*img2;
img12 = img1*img2;
img1_sum = __conv(tf.reshape(img1,[1,patch_h,patch_w,1]), sum2_filter)
img2_sum = __conv(tf.reshape(img2,[1,patch_h,patch_w,1]), sum2_filter)
img1_sq_sum = __conv(tf.reshape(img1_sq,[1,patch_h,patch_w,1]), sum2_filter)
img2_sq_sum = __conv(tf.reshape(img2_sq,[1,patch_h,patch_w,1]), sum2_filter)
img12_sum = __conv(tf.reshape(img12,[1,patch_h,patch_w,1]), sum2_filter)
img12_sum_mul = img1_sum*img2_sum
img12_sq_sum_mul = img1_sum*img1_sum + img2_sum*img2_sum
numerator = 4*(N*img12_sum - img12_sum_mul)*img12_sum_mul
denominator1 = N*(img1_sq_sum + img2_sq_sum) - img12_sq_sum_mul
denominator = denominator1*img12_sq_sum_mul
dd = tf.where(tf.less(denominator, 1e-7), tf.ones_like(denominator), denominator)
dd = numerator/dd
# cln = tf.where(tf.is_nan(dd),tf.zeros_like(denominator), dd)
cln = tf.where(tf.less(denominator, 1e-7), tf.zeros_like(dd), dd)
return tf.reduce_mean(cln[:,4:-4,4:-4,:])
def Q(y_true,y_pred,batch_size,ms_channels):
return tf.reduce_mean([tf.reduce_mean([q_fs( y_true[i,:,:,b],y_pred[i,:,:,b]) for b in range(ms_channels)]) for i in range(batch_size)])
def log10(x):
numerator = tf.log(x)
denominator = tf.log(tf.constant(10, dtype=numerator.dtype))
return numerator / denominator
def mse(y_true, y_pred):
return tf.reduce_mean((y_true-y_pred)**2)
def psnr(y_true, y_pred,norm=1):
normalization = 1.0
msev = mse(y_true, y_pred)
value = 10.0 * log10(normalization / msev)
return value