forked from kevinddchen/Keras-FCN
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodels.py
217 lines (183 loc) · 10.9 KB
/
models.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
'''
This file contains the FCN models.
'''
import numpy as np
import tensorflow as tf
import tensorflow.keras as keras
def vgg16(l2=0, dropout=0):
'''Convolutionized VGG16 network.
Args:
l2 (float): L2 regularization strength
dropout (float): Dropout rate
Returns:
(keras Model)
'''
## Input
input_layer = keras.Input(shape=(None, None, 3), name='input')
## Preprocessing
x = keras.layers.Lambda(tf.keras.applications.vgg16.preprocess_input, name='preprocessing')(input_layer)
## Block 1
x = keras.layers.Conv2D(filters=64, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu',
kernel_regularizer=keras.regularizers.L2(l2=l2), name='block1_conv1')(x)
x = keras.layers.Conv2D(filters=64, kernel_size=3, strides=(1,1), padding='same', activation='relu',
kernel_regularizer=keras.regularizers.L2(l2=l2), name='block1_conv2')(x)
x = keras.layers.MaxPool2D(pool_size=(2,2), strides=(2,2), padding='valid', name='block1_pool')(x)
## Block 2
x = keras.layers.Conv2D(filters=128, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu',
kernel_regularizer=keras.regularizers.L2(l2=l2), name='block2_conv1')(x)
x = keras.layers.Conv2D(filters=128, kernel_size=3, strides=(1,1), padding='same', activation='relu',
kernel_regularizer=keras.regularizers.L2(l2=l2), name='block2_conv2')(x)
x = keras.layers.MaxPool2D(pool_size=(2,2), strides=(2,2), padding='valid', name='block2_pool')(x)
## Block 3
x = keras.layers.Conv2D(filters=256, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu',
kernel_regularizer=keras.regularizers.L2(l2=l2), name='block3_conv1')(x)
x = keras.layers.Conv2D(filters=256, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu',
kernel_regularizer=keras.regularizers.L2(l2=l2), name='block3_conv2')(x)
x = keras.layers.Conv2D(filters=256, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu',
kernel_regularizer=keras.regularizers.L2(l2=l2), name='block3_conv3')(x)
x = keras.layers.MaxPool2D(pool_size=(2,2), strides=(2,2), padding='valid', name='block3_pool')(x)
## Block 4
x = keras.layers.Conv2D(filters=512, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu',
kernel_regularizer=keras.regularizers.L2(l2=l2), name='block4_conv1')(x)
x = keras.layers.Conv2D(filters=512, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu',
kernel_regularizer=keras.regularizers.L2(l2=l2), name='block4_conv2')(x)
x = keras.layers.Conv2D(filters=512, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu',
kernel_regularizer=keras.regularizers.L2(l2=l2), name='block4_conv3')(x)
x = keras.layers.MaxPool2D(pool_size=(2,2), strides=(2,2), padding='valid', name='block4_pool')(x)
## Block 5
x = keras.layers.Conv2D(filters=512, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu',
kernel_regularizer=keras.regularizers.L2(l2=l2), name='block5_conv1')(x)
x = keras.layers.Conv2D(filters=512, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu',
kernel_regularizer=keras.regularizers.L2(l2=l2), name='block5_conv2')(x)
x = keras.layers.Conv2D(filters=512, kernel_size=(3,3), strides=(1,1), padding='same', activation='relu',
kernel_regularizer=keras.regularizers.L2(l2=l2), name='block5_conv3')(x)
x = keras.layers.MaxPool2D(pool_size=(2,2), strides=(2,2), padding='valid', name='block5_pool')(x)
## Convolutionized fully-connected layers
x = keras.layers.Conv2D(filters=4096, kernel_size=(7,7), strides=(1,1), padding='same', activation='relu',
kernel_regularizer=keras.regularizers.L2(l2=l2), name='conv6')(x)
x = keras.layers.Dropout(rate=dropout, name='drop6')(x)
x = keras.layers.Conv2D(filters=4096, kernel_size=(1,1), strides=(1,1), padding='same', activation='relu',
kernel_regularizer=keras.regularizers.L2(l2=l2), name='conv7')(x)
x = keras.layers.Dropout(rate=dropout, name='drop7')(x)
## Inference layer
x = keras.layers.Conv2D(filters=1000, kernel_size=(1,1), strides=(1,1), padding='same', activation='softmax',
name='pred')(x)
return keras.Model(input_layer, x)
def fcn32(vgg16, l2=0):
'''32x upsampled FCN.
Args:
vgg16 (keras Model): VGG16 model to build upon
l2 (float): L2 regularization strength
Returns:
(keras Model)
'''
x = keras.layers.Conv2D(filters=21, kernel_size=(1,1), strides=(1,1), padding='same', activation='linear',
kernel_regularizer=keras.regularizers.L2(l2=l2),
name='score7')(vgg16.get_layer('drop7').output)
x = keras.layers.Conv2DTranspose(filters=21, kernel_size=(64,64), strides=(32,32),
padding='same', use_bias=False, activation='softmax',
kernel_initializer=BilinearInitializer(),
kernel_regularizer=keras.regularizers.L2(l2=l2),
name='fcn32')(x)
return keras.Model(vgg16.input, x)
def fcn16(vgg16, fcn32, l2=0):
'''16x upsampled FCN.
Args:
vgg16 (keras Model): VGG16 model to build upon
fcn32 (keras Model): FCN32 model to build upon
l2 (float): L2 regularization strength
Returns:
(keras Model)
'''
x = keras.layers.Conv2DTranspose(filters=21, kernel_size=(4,4), strides=(2,2),
padding='same', use_bias=False, activation='linear',
kernel_initializer=BilinearInitializer(),
kernel_regularizer=keras.regularizers.L2(l2=l2),
name='score7_upsample')(fcn32.get_layer('score7').output)
y = keras.layers.Conv2D(filters=21, kernel_size=(1,1), strides=(1,1), padding='same', activation='linear',
kernel_initializer=keras.initializers.Zeros(),
kernel_regularizer=keras.regularizers.L2(l2=l2),
name='score4')(vgg16.get_layer('block4_pool').output)
x = keras.layers.Add(name='skip4')([x, y])
x = keras.layers.Conv2DTranspose(filters=21, kernel_size=(32,32), strides=(16, 16),
padding='same', use_bias=False, activation='softmax',
kernel_initializer=BilinearInitializer(),
kernel_regularizer=keras.regularizers.L2(l2=l2),
name='fcn16')(x)
return keras.Model(fcn32.input, x)
def fcn8(vgg16, fcn16, l2=0):
'''8x upsampled FCN.
Args:
vgg16 (keras Model): VGG16 model to build upon
fcn16 (keras Model): FCN16 model to build upon
l2 (float): L2 regularization strength
Returns:
(keras Model)
'''
x = keras.layers.Conv2DTranspose(filters=21, kernel_size=(4,4), strides=(2,2),
padding='same', use_bias=False, activation='linear',
kernel_initializer=BilinearInitializer(),
kernel_regularizer=keras.regularizers.L2(l2=l2),
name='skip4_upsample')(fcn16.get_layer('skip4').output)
y = keras.layers.Conv2D(filters=21, kernel_size=(1,1), strides=(1,1), padding='same', activation='linear',
kernel_initializer=keras.initializers.Zeros(),
kernel_regularizer=keras.regularizers.L2(l2=l2),
name='score3')(vgg16.get_layer('block3_pool').output)
x = keras.layers.Add(name='skip3')([x, y])
x = keras.layers.Conv2DTranspose(filters=21, kernel_size=(16,16), strides=(8,8),
padding='same', use_bias=False, activation='softmax',
kernel_initializer=BilinearInitializer(),
kernel_regularizer=keras.regularizers.L2(l2=l2),
name='fcn8')(x)
return keras.Model(fcn16.input, x)
## ================
## Misc functions for training
## ================
class BilinearInitializer(keras.initializers.Initializer):
'''Initializer for Conv2DTranspose to perform bilinear interpolation on each channel.'''
def __call__(self, shape, dtype=None, **kwargs):
kernel_size, _, filters, _ = shape
arr = np.zeros((kernel_size, kernel_size, filters, filters))
## make filter that performs bilinear interpolation through Conv2DTranspose
upscale_factor = (kernel_size+1)//2
if kernel_size % 2 == 1:
center = upscale_factor - 1
else:
center = upscale_factor - 0.5
og = np.ogrid[:kernel_size, :kernel_size]
kernel = (1-np.abs(og[0]-center)/upscale_factor) * \
(1-np.abs(og[1]-center)/upscale_factor) # kernel shape is (kernel_size, kernel_size)
for i in range(filters):
arr[..., i, i] = kernel
return tf.convert_to_tensor(arr, dtype=dtype)
def crossentropy(y_true, y_pred_onehot):
'''Custom cross-entropy to handle borders (class = -1).'''
n_valid = tf.math.reduce_sum(tf.cast(y_true != 255, tf.float32))
y_true_onehot = tf.cast(np.arange(21) == y_true, tf.float32)
return tf.reduce_sum(-y_true_onehot * tf.math.log(y_pred_onehot + 1e-7)) / n_valid
def pixelacc(y_true, y_pred_onehot):
'''Custom pixel accuracy to handle borders (class = -1).'''
n_valid = tf.math.reduce_sum(tf.cast(y_true != 255, tf.float32))
y_true = tf.cast(y_true, tf.int32)[..., 0]
y_pred = tf.argmax(y_pred_onehot, axis=-1, output_type=tf.int32)
return tf.reduce_sum(tf.cast(y_true == y_pred, tf.float32)) / n_valid
class MyMeanIoU(keras.metrics.MeanIoU):
'''Custom meanIoU to handle borders (class = -1).'''
def update_state(self, y_true, y_pred_onehot, sample_weight=None):
y_pred = tf.argmax(y_pred_onehot, axis=-1)
## add 1 so boundary class=0
y_true = tf.cast(y_true+1, self._dtype)
y_pred = tf.cast(y_pred+1, self._dtype)
## Flatten the input if its rank > 1.
if y_pred.shape.ndims > 1:
y_pred = tf.reshape(y_pred, [-1])
if y_true.shape.ndims > 1:
y_true = tf.reshape(y_true, [-1])
## calculate confusion matrix with one extra class
current_cm = tf.math.confusion_matrix(
y_true,
y_pred,
self.num_classes+1,
weights=sample_weight,
dtype=self._dtype)
return self.total_cm.assign_add(current_cm[1:, 1:]) # remove boundary