Models:
Catherine Stones
cathy-stones/
DUCK-Net
0
Downloads: 0
[541a77]: / ModelArchitecture / DUCK_Net.py

Download this file

71 lines (50 with data), 2.7 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
import tensorflow as tf

from keras.layers import Conv2D, UpSampling2D

from keras.layers import add

from keras.models import Model



from CustomLayers.ConvBlock2D import conv_block_2D



kernel_initializer = 'he_uniform'

interpolation = "nearest"





def create_model(img_height, img_width, input_chanels, out_classes, starting_filters):

    input_layer = tf.keras.layers.Input((img_height, img_width, input_chanels))



    print('Starting DUCK-Net')



    p1 = Conv2D(starting_filters * 2, 2, strides=2, padding='same')(input_layer)

    p2 = Conv2D(starting_filters * 4, 2, strides=2, padding='same')(p1)

    p3 = Conv2D(starting_filters * 8, 2, strides=2, padding='same')(p2)

    p4 = Conv2D(starting_filters * 16, 2, strides=2, padding='same')(p3)

    p5 = Conv2D(starting_filters * 32, 2, strides=2, padding='same')(p4)



    t0 = conv_block_2D(input_layer, starting_filters, 'duckv2', repeat=1)



    l1i = Conv2D(starting_filters * 2, 2, strides=2, padding='same')(t0)

    s1 = add([l1i, p1])

    t1 = conv_block_2D(s1, starting_filters * 2, 'duckv2', repeat=1)



    l2i = Conv2D(starting_filters * 4, 2, strides=2, padding='same')(t1)

    s2 = add([l2i, p2])

    t2 = conv_block_2D(s2, starting_filters * 4, 'duckv2', repeat=1)



    l3i = Conv2D(starting_filters * 8, 2, strides=2, padding='same')(t2)

    s3 = add([l3i, p3])

    t3 = conv_block_2D(s3, starting_filters * 8, 'duckv2', repeat=1)



    l4i = Conv2D(starting_filters * 16, 2, strides=2, padding='same')(t3)

    s4 = add([l4i, p4])

    t4 = conv_block_2D(s4, starting_filters * 16, 'duckv2', repeat=1)



    l5i = Conv2D(starting_filters * 32, 2, strides=2, padding='same')(t4)

    s5 = add([l5i, p5])

    t51 = conv_block_2D(s5, starting_filters * 32, 'resnet', repeat=2)

    t53 = conv_block_2D(t51, starting_filters * 16, 'resnet', repeat=2)



    l5o = UpSampling2D((2, 2), interpolation=interpolation)(t53)

    c4 = add([l5o, t4])

    q4 = conv_block_2D(c4, starting_filters * 8, 'duckv2', repeat=1)



    l4o = UpSampling2D((2, 2), interpolation=interpolation)(q4)

    c3 = add([l4o, t3])

    q3 = conv_block_2D(c3, starting_filters * 4, 'duckv2', repeat=1)



    l3o = UpSampling2D((2, 2), interpolation=interpolation)(q3)

    c2 = add([l3o, t2])

    q6 = conv_block_2D(c2, starting_filters * 2, 'duckv2', repeat=1)



    l2o = UpSampling2D((2, 2), interpolation=interpolation)(q6)

    c1 = add([l2o, t1])

    q1 = conv_block_2D(c1, starting_filters, 'duckv2', repeat=1)



    l1o = UpSampling2D((2, 2), interpolation=interpolation)(q1)

    c0 = add([l1o, t0])

    z1 = conv_block_2D(c0, starting_filters, 'duckv2', repeat=1)



    output = Conv2D(out_classes, (1, 1), activation='sigmoid')(z1)



    model = Model(inputs=input_layer, outputs=output)



    return model