# coding=utf-8
from keras import backend as K
from keras.layers import Activation, Lambda
from keras.layers.convolutional import Conv2D
from keras.layers.merge import Add
from ..layers.core import interp, Conv2D_BN as ConvBN
def conv_block_side(x):
x = ConvBN(
filters=32,
kernel_size=3,
strides=2,
padding='same',
activation='relu',
name='conv1_sub1')(x)
x = ConvBN(
filters=32,
kernel_size=3,
strides=2,
padding='same',
activation='relu',
name='conv2_sub1')(x)
x = ConvBN(
filters=64,
kernel_size=3,
strides=2,
padding='same',
activation='relu',
name='conv3_sub1')(x)
x = ConvBN(
filters=128,
kernel_size=1,
padding='same',
name='conv3_sub1_proj')(x)
return x
def build(inp, encoder, nc, valid_shapes):
side = conv_block_side(inp)
x = Lambda(
interp,
arguments={'shape': valid_shapes[3]},
name='sub24_sum_interp')(encoder)
main = ConvBN(
filters=128,
kernel_size=3,
dilation_rate=2,
padding='same',
name='conv_sub2')(x)
x = Add(name='sub12_sum')([main, side])
x = Activation('relu')(x)
x = Lambda(
interp,
arguments={'shape': valid_shapes[2]},
name='sub12_sum_interp')(x)
x = Conv2D(
filters=nc,
kernel_size=1,
name='conv6_cls')(x)
out = Lambda(
interp,
arguments={'shape': valid_shapes[0]},
name='conv6_interp')(x)
return out
