def model_v10(basemodel_name, model_image_size=(384, 512)):
""" use conv output 7 column
:param basemodel_name: string, one of 'resnet50', 'xception', 'inception_v3' or 'inception_resnet_v2'
:param model_image_size: original size is 540(height)*720(width)
:return: keras model
"""
# First, define the vision modules
# img input
input_img = Input((*model_image_size, 3))
# build basemodel
if basemodel_name == 'resnet50':
basemodel = ResNet50(input_tensor=input_img,
weights='imagenet',
include_top=False)
elif basemodel_name == 'xception':
basemodel = Xception(input_tensor=input_img,
weights='imagenet',
include_top=False)
elif basemodel_name == 'inception_v3':
basemodel = InceptionV3(input_tensor=input_img,
weights='imagenet',
include_top=False)
elif basemodel_name == 'inception_resnet_v2':
basemodel = InceptionResNetV2(input_tensor=input_img,
weights='imagenet',
include_top=False)
else:
print('basemodel_name not defined！')
raise NameError
# use serveral layers as output
out_2_35 = basemodel.get_layer(name='mixed2').output #block1_conv1
out_5_17 = basemodel.get_layer(name='mixed4').output
vision_model = Model(input_img, [out_2_35, out_5_17], name='vision_model')
# get output for a and b
input_a = Input((*model_image_size, 3), name='input_a')
input_b = Input((*model_image_size, 3), name='input_b')
a_out_2_35, a_out_5_17 = vision_model(input_a)
b_out_2_35, b_out_5_17 = vision_model(input_b)
# define abs layers
#my_abs = Lambda(lambda z: K.abs(z))
# sub
#p_low = concatenate([a_out_2_35, b_out_2_35])
#sub_1 = MaxPooling2D((3, 3), strides=(2, 2), padding='valid')(Subtract(name='sub2')([a_out_2_35, b_out_2_35]))
sub_4 = (Subtract(name='sub4')([a_out_5_17, b_out_5_17]))
x1 = sub_4
x = conv2d_bn(x1, 10, 1, 1, name='after_conv_2')
# flatten
y = Flatten(name='after_flatten')(x)
# add some dense layers
y = Dense(128, activation='relu', name='after_dense_1')(y)
out = Dense(1, activation='sigmoid')(y)
return Model([input_a, input_b], out)
def inception_v3_mixed_module(x):
if image_data_format() == 'channels_first':
channel_axis = 1
else:
channel_axis = 3
branch1x1 = conv2d_bn(x, 64, 1, 1)
branch5x5 = conv2d_bn(x, 48, 1, 1)
branch5x5 = conv2d_bn(branch5x5, 64, 5, 5)
branch3x3dbl = conv2d_bn(x, 64, 1, 1)
branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3)
branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3)
branch_pool = layers.AveragePooling2D((3, 3),
strides=(1, 1),
padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 32, 1, 1)
return layers.concatenate(
[branch1x1, branch5x5, branch3x3dbl, branch_pool], axis=channel_axis)
def localize_classify_deep(fine_tune=False):
model = InceptionV3(include_top=False, input_shape=(299, 299, 3))
if fine_tune:
for layer in model.layers:
layer.trainable = False
output = model.get_layer(index=-1).output
# 8x8x1024
branch1x1 = conv2d_bn(output, 160, 1, 1)
branch3x3 = conv2d_bn(output, 192, 1, 1)
branch3x3_1 = conv2d_bn(branch3x3, 192, 1, 3)
branch3x3_2 = conv2d_bn(branch3x3, 192, 3, 1)
branch3x3 = merge([branch3x3_1, branch3x3_2],
mode='concat',
concat_axis=-1)
branch3x3dbl = conv2d_bn(output, 224, 1, 1)
branch3x3dbl = conv2d_bn(output, 192, 3, 3)
branch3x3dbl_1 = conv2d_bn(branch3x3dbl, 192, 1, 3)
branch3x3dbl_2 = conv2d_bn(branch3x3dbl, 192, 3, 1)
branch3x3dbl = merge([branch3x3dbl_1, branch3x3dbl_2],
mode='concat',
concat_axis=-1)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1),
border_mode='same')(output)
branch_pool = conv2d_bn(branch_pool, 96, 1, 1)
output = merge([branch1x1, branch3x3, branch3x3dbl, branch_pool],
mode='concat',
concat_axis=-1)
# 8x8x512
branch1x1 = conv2d_bn(output, 80, 1, 1)
branch3x3 = conv2d_bn(output, 96, 1, 1)
branch3x3_1 = conv2d_bn(branch3x3, 96, 1, 3)
branch3x3_2 = conv2d_bn(branch3x3, 96, 3, 1)
branch3x3 = merge([branch3x3_1, branch3x3_2],
mode='concat',
concat_axis=-1)
branch3x3dbl = conv2d_bn(output, 112, 1, 1)
branch3x3dbl = conv2d_bn(output, 96, 3, 3)
branch3x3dbl_1 = conv2d_bn(branch3x3dbl, 96, 1, 3)
branch3x3dbl_2 = conv2d_bn(branch3x3dbl, 96, 3, 1)
branch3x3dbl = merge([branch3x3dbl_1, branch3x3dbl_2],
mode='concat',
concat_axis=-1)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1),
border_mode='same')(output)
branch_pool = conv2d_bn(branch_pool, 48, 1, 1)
output = merge([branch1x1, branch3x3, branch3x3dbl, branch_pool],
mode='concat',
concat_axis=-1)
localize = Convolution2D(4, 1, 1)(output)
localize = GlobalAveragePooling2D(name='localize')(localize)
classify = Convolution2D(8, 1, 1)(output)
classify = GlobalAveragePooling2D()(classify)
classify = Activation('softmax', name=OUTPUT_NAME)(classify)
model = Model(model.input, [localize, classify])
return model
def localize(fine_tune=False):
from keras.models import Model
model = InceptionV3(include_top=False, input_shape=(299, 299, 3))
if fine_tune:
for layer in model.layers:
layer.trainable = False
output = model.get_layer(index=-1).output
# 8x8x1024
branch1x1 = conv2d_bn(output, 160, 1, 1)
branch3x3 = conv2d_bn(output, 192, 1, 1)
branch3x3_1 = conv2d_bn(branch3x3, 192, 1, 3)
branch3x3_2 = conv2d_bn(branch3x3, 192, 3, 1)
branch3x3 = merge([branch3x3_1, branch3x3_2],
mode='concat',
concat_axis=-1)
branch3x3dbl = conv2d_bn(output, 224, 1, 1)
branch3x3dbl = conv2d_bn(output, 192, 3, 3)
branch3x3dbl_1 = conv2d_bn(branch3x3dbl, 192, 1, 3)
branch3x3dbl_2 = conv2d_bn(branch3x3dbl, 192, 3, 1)
branch3x3dbl = merge([branch3x3dbl_1, branch3x3dbl_2],
mode='concat',
concat_axis=-1)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1),
border_mode='same')(output)
branch_pool = conv2d_bn(branch_pool, 96, 1, 1)
output = merge([branch1x1, branch3x3, branch3x3dbl, branch_pool],
mode='concat',
concat_axis=-1)
# 8x8x512
branch1x1 = conv2d_bn(output, 80, 1, 1)
branch3x3 = conv2d_bn(output, 96, 1, 1)
branch3x3_1 = conv2d_bn(branch3x3, 96, 1, 3)
branch3x3_2 = conv2d_bn(branch3x3, 96, 3, 1)
branch3x3 = merge([branch3x3_1, branch3x3_2],
mode='concat',
concat_axis=-1)
branch3x3dbl = conv2d_bn(output, 112, 1, 1)
branch3x3dbl = conv2d_bn(output, 96, 3, 3)
branch3x3dbl_1 = conv2d_bn(branch3x3dbl, 96, 1, 3)
branch3x3dbl_2 = conv2d_bn(branch3x3dbl, 96, 3, 1)
branch3x3dbl = merge([branch3x3dbl_1, branch3x3dbl_2],
mode='concat',
concat_axis=-1)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1),
border_mode='same')(output)
branch_pool = conv2d_bn(branch_pool, 48, 1, 1)
output = merge([branch1x1, branch3x3, branch3x3dbl, branch_pool],
mode='concat',
concat_axis=-1)
localize = Convolution2D(4, 1, 1)(output)
localize = GlobalAveragePooling2D(name='localize')(localize)
model = Model(model.input, localize)
return model
def InceptionV3_deform_model(nb_classes):
if K.image_data_format() == 'channels_first':
channel_axis = 1
else:
channel_axis = 3
img_input = InceptionV3().input
x = conv2d_bn(img_input, 32, 3, 3, strides=(2, 2), padding='valid')
x = ConvOffset2D(32, name='conv1_offset')(x)
x = conv2d_bn(x, 32, 3, 3, padding='valid')
x = ConvOffset2D(32, name='conv2_offset')(x)
x = conv2d_bn(x, 64, 3, 3)
x = MaxPooling2D((3, 3), strides=(2, 2))(x)
x = conv2d_bn(x, 80, 1, 1, padding='valid')
x = ConvOffset2D(80, name='conv3_offset')(x)
x = conv2d_bn(x, 192, 3, 3, padding='valid')
x = MaxPooling2D((3, 3), strides=(2, 2))(x)
# mixed 0, 1, 2: 35 x 35 x 256
branch1x1 = conv2d_bn(x, 64, 1, 1)
branch5x5 = conv2d_bn(x, 48, 1, 1)
branch5x5 = conv2d_bn(branch5x5, 64, 5, 5)
branch3x3dbl = conv2d_bn(x, 64, 1, 1)
branch3x3dbl = ConvOffset2D(64, name='conv4_offset')(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3)
branch3x3dbl = ConvOffset2D(96, name='conv5_offset')(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1), padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 32, 1, 1)
x = concatenate([branch1x1, branch5x5, branch3x3dbl, branch_pool],
axis=channel_axis,
name='mixed0')
# mixed 1: 35 x 35 x 256
branch1x1 = conv2d_bn(x, 64, 1, 1)
branch5x5 = conv2d_bn(x, 48, 1, 1)
branch5x5 = conv2d_bn(branch5x5, 64, 5, 5)
branch3x3dbl = conv2d_bn(x, 64, 1, 1)
branch3x3dbl = ConvOffset2D(64, name='conv6_offset')(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3)
branch3x3dbl = ConvOffset2D(96, name='conv7_offset')(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1), padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 64, 1, 1)
x = concatenate([branch1x1, branch5x5, branch3x3dbl, branch_pool],
axis=channel_axis,
name='mixed1')
# mixed 2: 35 x 35 x 256
branch1x1 = conv2d_bn(x, 64, 1, 1)
branch5x5 = conv2d_bn(x, 48, 1, 1)
branch5x5 = conv2d_bn(branch5x5, 64, 5, 5)
branch3x3dbl = conv2d_bn(x, 64, 1, 1)
branch3x3dbl = ConvOffset2D(64, name='conv8_offset')(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3)
branch3x3dbl = ConvOffset2D(96, name='conv9_offset')(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1), padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 64, 1, 1)
x = concatenate([branch1x1, branch5x5, branch3x3dbl, branch_pool],
axis=channel_axis,
name='mixed2')
# mixed 3: 17 x 17 x 768
x = ConvOffset2D(288, name='conv10_offset')(x)
branch3x3 = conv2d_bn(x, 384, 3, 3, strides=(2, 2), padding='valid')
branch3x3dbl = conv2d_bn(x, 64, 1, 1)
branch3x3dbl = ConvOffset2D(64, name='conv11_offset')(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3)
branch3x3dbl = ConvOffset2D(96, name='conv12_offset')(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl,
96,
3,
3,
strides=(2, 2),
padding='valid')
branch_pool = MaxPooling2D((3, 3), strides=(2, 2))(x)
x = concatenate([branch3x3, branch3x3dbl, branch_pool],
axis=channel_axis,
name='mixed3')
# mixed 4: 17 x 17 x 768
branch1x1 = conv2d_bn(x, 192, 1, 1)
branch7x7 = conv2d_bn(x, 128, 1, 1)
branch7x7 = conv2d_bn(branch7x7, 128, 1, 7)
branch7x7 = conv2d_bn(branch7x7, 192, 7, 1)
branch7x7dbl = conv2d_bn(x, 128, 1, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 128, 7, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 128, 1, 7)
branch7x7dbl = conv2d_bn(branch7x7dbl, 128, 7, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 1, 7)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1), padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 192, 1, 1)
x = concatenate([branch1x1, branch7x7, branch7x7dbl, branch_pool],
axis=channel_axis,
name='mixed4')
# mixed 5, 6: 17 x 17 x 768
for i in range(2):
branch1x1 = conv2d_bn(x, 192, 1, 1)
branch7x7 = conv2d_bn(x, 160, 1, 1)
branch7x7 = conv2d_bn(branch7x7, 160, 1, 7)
branch7x7 = conv2d_bn(branch7x7, 192, 7, 1)
branch7x7dbl = conv2d_bn(x, 160, 1, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 160, 7, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 160, 1, 7)
branch7x7dbl = conv2d_bn(branch7x7dbl, 160, 7, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 1, 7)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1),
padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 192, 1, 1)
x = concatenate([branch1x1, branch7x7, branch7x7dbl, branch_pool],
axis=channel_axis,
name='mixed' + str(5 + i))
# mixed 7: 17 x 17 x 768
branch1x1 = conv2d_bn(x, 192, 1, 1)
branch7x7 = conv2d_bn(x, 192, 1, 1)
branch7x7 = conv2d_bn(branch7x7, 192, 1, 7)
branch7x7 = conv2d_bn(branch7x7, 192, 7, 1)
branch7x7dbl = conv2d_bn(x, 192, 1, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 7, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 1, 7)
branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 7, 1)
branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 1, 7)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1), padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 192, 1, 1)
x = concatenate([branch1x1, branch7x7, branch7x7dbl, branch_pool],
axis=channel_axis,
name='mixed7')
# mixed 8: 8 x 8 x 1280
branch3x3 = conv2d_bn(x, 192, 1, 1)
branch3x3 = ConvOffset2D(192, name='conv13_offset')(branch3x3)
branch3x3 = conv2d_bn(branch3x3,
320,
3,
3,
strides=(2, 2),
padding='valid')
branch7x7x3 = conv2d_bn(x, 192, 1, 1)
branch7x7x3 = conv2d_bn(branch7x7x3, 192, 1, 7)
branch7x7x3 = conv2d_bn(branch7x7x3, 192, 7, 1)
branch7x7x3 = ConvOffset2D(192, name='conv14_offset')(branch7x7x3)
branch7x7x3 = conv2d_bn(branch7x7x3,
192,
3,
3,
strides=(2, 2),
padding='valid')
branch_pool = MaxPooling2D((3, 3), strides=(2, 2))(x)
x = concatenate([branch3x3, branch7x7x3, branch_pool],
axis=channel_axis,
name='mixed8')
# mixed 9: 8 x 8 x 2048
for i in range(2):
branch1x1 = conv2d_bn(x, 320, 1, 1)
branch3x3 = conv2d_bn(x, 384, 1, 1)
branch3x3_1 = conv2d_bn(branch3x3, 384, 1, 3)
branch3x3_2 = conv2d_bn(branch3x3, 384, 3, 1)
branch3x3 = concatenate([branch3x3_1, branch3x3_2],
axis=channel_axis,
name='mixed9_' + str(i))
branch3x3dbl = conv2d_bn(x, 448, 1, 1)
branch3x3dbl = ConvOffset2D(448, name='conv15_%d_offset' %
(i))(branch3x3dbl)
branch3x3dbl = conv2d_bn(branch3x3dbl, 384, 3, 3)
branch3x3dbl_1 = conv2d_bn(branch3x3dbl, 384, 1, 3)
branch3x3dbl_2 = conv2d_bn(branch3x3dbl, 384, 3, 1)
branch3x3dbl = concatenate([branch3x3dbl_1, branch3x3dbl_2],
axis=channel_axis)
branch_pool = AveragePooling2D((3, 3), strides=(1, 1),
padding='same')(x)
branch_pool = conv2d_bn(branch_pool, 192, 1, 1)
x = concatenate([branch1x1, branch3x3, branch3x3dbl, branch_pool],
axis=channel_axis,
name='mixed' + str(9 + i))
# Classification block
x = GlobalAveragePooling2D(name='avg_pool')(x)
x = Dense(nb_classes, activation='softmax', name='predictions')(x)
# Ensure that the model takes into account
# any potential predecessors of `input_tensor`.
inputs = img_input
# Create model.
model = Model(inputs, x, name='inception_v3_deform')
return model
def model_v3(basemodel_name, model_image_size=(360, 480)):
""" use conv output 7 column
:param basemodel_name: string, one of 'resnet50', 'xception', 'inception_v3' or 'inception_resnet_v2'
:param model_image_size: original size is 540(height)*720(width)
:return: keras model
"""
# First, define the vision modules
# img input
input_img = Input((*model_image_size, 3))
# build basemodel
if basemodel_name == 'resnet50':
basemodel = ResNet50(input_tensor=input_img,
weights='imagenet',
include_top=False)
elif basemodel_name == 'xception':
basemodel = Xception(input_tensor=input_img,
weights='imagenet',
include_top=False)
elif basemodel_name == 'inception_v3':
basemodel = InceptionV3(input_tensor=input_img,
weights='imagenet',
include_top=False)
elif basemodel_name == 'inception_resnet_v2':
basemodel = InceptionResNetV2(input_tensor=input_img,
weights='imagenet',
include_top=False)
else:
print('basemodel_name not defined！')
raise NameError
# get serveral output layers
out_0_35 = basemodel.get_layer(name='mixed0').output
out_1_35 = basemodel.get_layer(name='mixed1').output
out_2_35 = basemodel.get_layer(name='mixed2').output
out_3_17 = basemodel.get_layer(name='mixed3').output
out_4_17 = basemodel.get_layer(name='mixed4').output
vision_model = Model(input_img,
[out_0_35, out_1_35, out_2_35, out_3_17, out_4_17],
name='vision_model')
# Then define the tell-img-apart model
input_a = Input((*model_image_size, 3), name='input_a')
input_b = Input((*model_image_size, 3), name='input_b')
# The vision model will be shared, weights and all
a_out_0_35, a_out_1_35, a_out_2_35, a_out_3_17, a_out_4_17 = vision_model(
input_a)
b_out_0_35, b_out_1_35, b_out_2_35, b_out_3_17, b_out_4_17 = vision_model(
input_b)
sub_0_35 = Subtract(name='sub0')([a_out_0_35, b_out_0_35])
sub_1_35 = Subtract(name='sub1')([a_out_1_35, b_out_1_35])
sub_2_35 = Subtract(name='sub2')([a_out_2_35, b_out_2_35])
sub_3_17 = Subtract(name='sub3')([a_out_3_17, b_out_3_17])
sub_4_17 = Subtract(name='sub4')([a_out_4_17, b_out_4_17])
l0 = MaxPooling2D((3, 3), strides=(2, 2), padding='valid')(sub_0_35)
l1 = MaxPooling2D((3, 3), strides=(2, 2), padding='valid')(sub_1_35)
l2 = MaxPooling2D((3, 3), strides=(2, 2), padding='valid')(sub_2_35)
l_35 = MaxPooling2D((3, 3), strides=(2, 2),
padding='valid')(concatenate([a_out_2_35, b_out_2_35]))
l_17 = concatenate([a_out_4_17, b_out_4_17])
x1 = concatenate([l0, l1, l2, sub_3_17, sub_4_17])
x2 = concatenate([l_35, l_17])
# add some conv layers
x1 = conv2d_bn(x1, 128, 1, 1, name='after_conv_x1')
x2 = conv2d_bn(x2, 128, 1, 1, name='after_conv_x2')
x = concatenate([x1, x2])
x = conv2d_bn(x, 10, 1, 1, name='after_conv_2')
# flatten
y = Flatten(name='after_flatten')(x)
# add some dense layers
y = Dense(128, activation='relu', name='after_dense_1')(y)
out = Dense(1, activation='sigmoid')(y)
return Model([input_a, input_b], out)
def model_v4(basemodel_name, model_image_size=(360, 480)):
"""
:param basemodel_name: string, one of 'resnet50', 'xception', 'inception_v3' or 'inception_resnet_v2'
:param model_image_size: original size is 540(height)*720(width)
:return: keras model
"""
# First, define the vision modules
# img input
input_img = Input((*model_image_size, 3))
# build basemodel
if basemodel_name == 'resnet50':
basemodel = ResNet50(input_tensor=input_img,
weights='imagenet',
include_top=False)
elif basemodel_name == 'xception':
basemodel = Xception(input_tensor=input_img,
weights='imagenet',
include_top=False)
elif basemodel_name == 'inception_v3':
basemodel = InceptionV3(input_tensor=input_img,
weights='imagenet',
include_top=False)
elif basemodel_name == 'inception_resnet_v2':
basemodel = InceptionResNetV2(input_tensor=input_img,
weights='imagenet',
include_top=False)
else:
print('basemodel_name not defined！')
raise NameError
# use serveral layers as output
out_0_35 = basemodel.get_layer(name='mixed0').output
out_1_35 = basemodel.get_layer(name='mixed1').output
out_2_35 = basemodel.get_layer(name='mixed2').output
out_3_17 = basemodel.get_layer(name='mixed3').output
out_4_17 = basemodel.get_layer(name='mixed4').output
out_5_17 = basemodel.get_layer(name='mixed5').output
vision_model = Model(
input_img,
[out_0_35, out_1_35, out_2_35, out_3_17, out_4_17, out_5_17],
name='vision_model')
# get output for a and b
input_a = Input((*model_image_size, 3), name='input_a')
input_b = Input((*model_image_size, 3), name='input_b')
a_out_0_35, a_out_1_35, a_out_2_35, a_out_3_17, a_out_4_17, a_out_5_17 = vision_model(
input_a)
b_out_0_35, b_out_1_35, b_out_2_35, b_out_3_17, b_out_4_17, b_out_5_17 = vision_model(
input_b)
# Then, define the tell-img-apart model
# define abs layers
my_abs = Lambda(lambda z: K.abs(z))
# sub
sub_0_35 = my_abs(Subtract(name='sub0')([a_out_0_35, b_out_0_35]))
sub_1_35 = my_abs(Subtract(name='sub1')([a_out_1_35, b_out_1_35]))
sub_2_35 = my_abs(Subtract(name='sub2')([a_out_2_35, b_out_2_35]))
sub_3_17 = my_abs(Subtract(name='sub3')([a_out_3_17, b_out_3_17]))
sub_4_17 = my_abs(Subtract(name='sub4')([a_out_4_17, b_out_4_17]))
sub_5_17 = my_abs(Subtract(name='sub5')([a_out_5_17, b_out_5_17]))
# max poool
sub_35 = MaxPooling2D(
(3, 3), strides=(2, 2),
padding='valid')(concatenate([sub_0_35, sub_1_35, sub_2_35]))
x_sub = concatenate([sub_35, sub_3_17, sub_4_17, sub_5_17])
# original
ori_35 = MaxPooling2D(
(3, 3), strides=(2, 2),
padding='valid')(concatenate([a_out_2_35, b_out_2_35]))
x_ori = concatenate([ori_35, a_out_5_17, b_out_5_17])
# add some conv layers
x_sub = conv2d_bn(x_sub, 128, 1, 1, name='after_conv_sub')
x_ori = conv2d_bn(x_ori, 128, 1, 1, name='after_conv_ori')
x = concatenate([x_sub, x_ori])
x = conv2d_bn(x, 10, 1, 1, name='after_conv_1')
# flatten
y = Flatten(name='after_flatten')(x)
# add some dense layers
y = Dense(128, activation='relu', name='after_dense_1')(y)
out = Dense(1, activation='sigmoid', name='output_dense')(y)
return Model([input_a, input_b], out)