def wrapper(x):
if(not only_resize):
# segmentation head
x = Conv2dBn(
channels * 8,
kernel_size=3,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
strides = 1,
use_batchnorm=True,
use_bias=False,
name=name + 'seghead_con3_1',
kernel_regularizer=l2(wd)
)(x)
x = Conv2dBn(
classes,
kernel_size=1,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
strides = 1,
use_batchnorm=False,
use_bias=False,
name=name + 'seghead_conv_end',
kernel_regularizer=l2(wd)
)(x)
# if(not self.scale == 1):
# x = layers.UpSampling2D(
# size=(self.scale, self.scale), interpolation='bilinear'
# )(x)
x = resize_bilinear(size=(h, w),name=name + 'resize_bil_end')(x)
return x
def wrapper(input_tensor):
x = backend.mean(input_tensor, axis=[1,2], keepdims=True) #layers.GlobalAveragePooling2D(axis=-1)(input_tensor)
# print(1, np.shape(x))
x = Conv2dBn(
out_channels,
kernel_size=1,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
strides = 1,
use_batchnorm=True,
name=name+"ce_block_cnn1_1",
use_bias=False,
kernel_regularizer=l2(wd)
)(x)
# print(2, np.shape(x))
x = layers.Add()([x, input_tensor])
# print(3, np.shape(x))
x= Conv2dBn(
out_channels,
kernel_size=3,
activation=None,
kernel_initializer='he_uniform',
padding='same',
strides = 1,
use_batchnorm=False,
use_bias=False,
name=name+"ge_block_cnn3_1",
kernel_regularizer=l2(wd)
)(x)
# print(4, np.shape(x))
return x
def wrapper(input_tensor):
x = Conv2dBn(out_channels,
kernel_size=3,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
strides=1,
use_batchnorm=True,
use_bias=False,
name=name + "ge_block_cnn3_1",
kernel_regularizer=l2(wd))(input_tensor)
x = layers.DepthwiseConv2D(3,
strides=(stride, stride),
padding='same',
depth_multiplier=e,
use_bias=False,
name=name + "geblock_dwconv3_1")(x)
if (stride == 2):
x = layers.DepthwiseConv2D(3,
strides=(1, 1),
padding='same',
use_bias=False,
name=name + "geblock_dwconv3_2",
kernel_regularizer=l2(wd))(x)
x = Conv2dBn(out_channels,
kernel_size=1,
activation=None,
kernel_initializer='he_uniform',
padding='same',
strides=1,
use_batchnorm=True,
use_bias=False,
name=name + "ge_block_cnn1_1",
kernel_regularizer=l2(wd))(x)
if (stride == 2):
x1 = layers.DepthwiseConv2D(
3,
strides=(stride, stride),
padding='same',
use_bias=False,
name=name + "geblock_dwconv3_3_shortcut")(input_tensor)
x1 = Conv2dBn(out_channels,
kernel_size=1,
activation=None,
kernel_initializer='he_uniform',
padding='same',
strides=1,
use_batchnorm=True,
name=name + "ge_block_cnn1_2_shortcut",
use_bias=False,
kernel_regularizer=l2(wd))(x1)
x = layers.Add()([x, x1])
else:
x = layers.Add()([x, input_tensor])
return x
def wrapper(input_tensor):
x = Conv2dBn(out_channels,
kernel_size=3,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
strides=2,
use_batchnorm=True,
use_bias=False,
name="stem_block_cnn",
kernel_regularizer=l2(wd))(input_tensor)
#----------------------------
x1 = Conv2dBn(out_channels / 2,
kernel_size=1,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
strides=1,
use_batchnorm=True,
use_bias=False,
name="stem_block_cnndown_cnn1",
kernel_regularizer=l2(wd))(x)
x1 = Conv2dBn(out_channels,
kernel_size=3,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
strides=2,
use_batchnorm=True,
use_bias=False,
name="stem_block_cnndown_cnn3",
kernel_regularizer=l2(wd))(x1)
#---------------------------------
#--------maxpooling--------------
x2 = layers.MaxPool2D(pool_size=(3, 3),
strides=2,
padding='same',
name="stem_block_maxpool")(x)
# x2, pooling_indices = tf.nn.max_pool_with_argmax(x1,
# ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")
# concatenate
x = layers.Concatenate(axis=-1)([x1, x2])
x = Conv2dBn(out_channels,
kernel_size=1,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
strides=1,
use_batchnorm=True,
use_bias=False,
name="stem_block_cnn_out",
kernel_regularizer=l2(wd))(x)
return x
def wrapper(input_tensor):
return Conv2dBn(
filters,
kernel_size=1,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
use_batchnorm=use_batchnorm,
name=name,
)(input_tensor)
def wrapper(detailed_in, semantic_in):
#--------------preparing detailed_in----------------------------------
_, w, h, _ = np.shape(detailed_in)
x_d = layers.DepthwiseConv2D(
3, strides=(1, 1), padding='same',depth_multiplier=1, use_bias=False, name=name+"agglayer_dwconv1_1"
)(detailed_in)
x_d = Conv2dBn(
out_channels,
kernel_size=1,
activation=None,
kernel_initializer='he_uniform',
padding='same',
strides = 1,
use_batchnorm=False,
name=name+"agglayer_cnn1_1",
use_bias=False,
kernel_regularizer=l2(wd)
)(x_d)
# now prepare detailed in to be the mixed with semantic
x_d2s= Conv2dBn(
out_channels,
kernel_size=3,
activation=None,
kernel_initializer='he_uniform',
padding='same',
strides = 2,
use_batchnorm=True,
use_bias=False,
name=name+"agglayer_cnn3_1",
kernel_regularizer=l2(wd)
)(detailed_in)
x_d2s = layers.AveragePooling2D(
pool_size=(3, 3), strides=(2,2), padding='same'
)(x_d2s)
#------------------------prepare semantic in--------------------------
x_s = layers.DepthwiseConv2D(
3, strides=(1, 1), padding='same',depth_multiplier=1, use_bias=False, name=name+"agglayer_dwconv1_2"
)(semantic_in)
# print(5, np.shape(x_s))
x_s = Conv2dBn(
out_channels,
kernel_size=1,
activation=None,
kernel_initializer='he_uniform',
padding='same',
strides = 1,
use_batchnorm=False,
name=name+"agglayer_cnn1_2",
use_bias=False,
kernel_regularizer=l2(wd)
)(x_s)
#now prepare semantic in to be mixed with detailed
x_s2d = Conv2dBn(
out_channels,
kernel_size=3,
activation=None,
kernel_initializer='he_uniform',
padding='same',
strides = 1,
use_batchnorm=True,
use_bias=False,
name=name+"agglayer_cnn3_2",
kernel_regularizer=l2(wd)
)(semantic_in)
x_s2d = resize_bilinear(size=(w,h),name=name+'resize_bil_1')(x_s2d)
# x_s2d = layers.UpSampling2D(
# size=(4, 4), interpolation='bilinear'
# )(x_s2d)
# mixing-----------------------------------------------
x_d = layers.Multiply()([x_d , x_s2d])
x_s = layers.Multiply()([x_s , x_d2s])
# x_s = layers.UpSampling2D(
# size=(4, 4), interpolation='bilinear'
# )(x_s)
x_s = resize_bilinear(size=(w,h),name=name+'resize_bil_1')(x_s)
x = layers.Add()([x_s, x_d])
x = Conv2dBn(
out_channels,
kernel_size=3,
activation=None,
kernel_initializer='he_uniform',
padding='same',
strides = 1,
use_batchnorm=True,
use_bias=False,
name=name+"agglayer_output",
kernel_regularizer=l2(wd)
)(x)
return x
def wrapper(input_tensor):
x= Conv2dBn(
64,
kernel_size=3,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
strides = 2,
use_batchnorm=True,
use_bias=False,
name="detail_branch_cnn_1_1",
kernel_regularizer=l2(wd),
)(input_tensor)
x= Conv2dBn(
64,
kernel_size=3,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
strides = 1,
use_batchnorm=True,
use_bias=False,
name="detail_branch_cnn_1_2",
kernel_regularizer=l2(wd)
)(x)
#---------------------------------------------------------------
x= Conv2dBn(
64,
kernel_size=3,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
strides = 2,
use_batchnorm=True,
use_bias=False,
name="detail_branch_cnn_2_1",
kernel_regularizer=l2(wd)
)(x)
x= Conv2dBn(
64,
kernel_size=3,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
strides = 1,
use_batchnorm=True,
use_bias=False,
name="detail_branch_cnn_2_2",
kernel_regularizer=l2(wd)
)(x)
x= Conv2dBn(
64,
kernel_size=3,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
strides = 1,
use_batchnorm=True,
use_bias=False,
name="detail_branch_cnn_2_3",
kernel_regularizer=l2(wd)
)(x)
#---------------------------------------------------------------
x= Conv2dBn(
128,
kernel_size=3,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
strides = 2,
use_batchnorm=True,
name="detail_branch_cnn_3_1",
)(x)
x= Conv2dBn(
128,
kernel_size=3,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
strides = 1,
use_batchnorm=True,
use_bias=False,
name="detail_branch_cnn_3_2",
kernel_regularizer=l2(wd)
)(x)
x= Conv2dBn(
128,
kernel_size=3,
activation='relu',
kernel_initializer='he_uniform',
padding='same',
strides = 1,
use_batchnorm=True,
use_bias=False,
name="detail_branch_cnn_3_3",
kernel_regularizer=l2(wd)
)(x)
# x = layers.Dropout(0.3)(x)
return x