def DenseBlock(stack,
n_layers,
growth_rate,
dropout_p,
attention_module=None,
scope=None):
"""
DenseBlock for DenseNet and FC-DenseNet
Arguments:
stack: input 4D tensor
n_layers: number of internal layers
growth_rate: number of feature maps per internal layer
Returns:
stack: current stack of feature maps (4D tensor)
new_features: 4D tensor containing only the new feature maps generated
in this block
"""
with tf.name_scope(scope) as sc:
new_features = []
for j in range(n_layers):
# Compute new feature maps
layer = preact_conv(stack, growth_rate, dropout_p=dropout_p)
new_features.append(layer)
# Stack new layer
stack = tf.concat([stack, layer], axis=-1)
new_features = tf.concat(new_features, axis=-1)
#add attention_module
attach_attention_module(stack, attention_module, scope)
return stack, new_features
def TransitionDown(inputs,
n_filters,
dropout_p=0.2,
attention_module=None,
scope=None):
"""
Transition Down (TD) for FC-DenseNet
Apply 1x1 BN + ReLU + conv then 2x2 max pooling
"""
with tf.name_scope(scope) as sc:
l = preact_conv(inputs,
n_filters,
kernel_size=[1, 1],
dropout_p=dropout_p)
l = slim.pool(l, [2, 2], stride=[2, 2], pooling_type='MAX')
# add attention_module
attach_attention_module(l, attention_module, scope)
return l
def conv_block2(input_tensor, kernel_size, filters, stage, block, strides=(2, 2)):
'''conv_block is the block that has a conv layer at shortcut
# Arguments
input_tensor: input tensor
kernel_size: defualt 3, the kernel size of middle conv layer at main path
filters: list of integers, the nb_filters of 3 conv layer at main path
stage: integer, current stage label, used for generating layer names
block: 'a','b'..., current block label, used for generating layer names
Note that from stage 3, the first conv layer at main path is with subsample=(2,2)
And the shortcut should have subsample=(2,2) as well
'''
nb_filter1, nb_filter2, nb_filter3 = filters
# if K.image_dim_ordering() == 'tf':
# bn_axis = 3
# else:
# bn_axis = 1
bn_axis = 3
conv_name_base = 'res' + str(stage) + block + '_branch'
bn_name_base = 'bn' + str(stage) + block + '_branch'
x = Conv2D(nb_filter1, 1, 1, subsample=strides,
name=conv_name_base + '2a')(input_tensor)
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x)
x = Activation('relu')(x)
x = Conv2D(nb_filter2, kernel_size, kernel_size, border_mode='same',
name=conv_name_base + '2b')(x)
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x)
x = Activation('relu')(x)
x = Conv2D(nb_filter3, 1, 1, name=conv_name_base + '2c')(x)
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2c')(x)
shortcut = Conv2D(nb_filter3, 1, 1, subsample=strides,
name=conv_name_base + '1')(input_tensor)
# shortcut = BatchNormalization(axis=bn_axis, name=bn_name_base + '1')(shortcut)
x = attach_attention_module(x, 'cbam_block')
x = add([x, shortcut])
x = Activation('relu')(x)
return x
def _wasnet_block(x, filter, strides=(2, 2), nl='RE'):
residual = layers.Conv2D(filter,
kernel_size=(1, 1),
strides=strides,
padding='same')(x)
residual = layers.BatchNormalization(axis=bn_axis)(residual)
cbam = attach_attention_module(residual, attention_module='cbam_block')
x = layers.SeparableConv2D(filter, (3, 3), padding='same')(x)
x = layers.BatchNormalization(axis=bn_axis)(x)
x = return_activation(x, nl)
x = layers.SeparableConv2D(filter, (3, 3), padding='same')(x)
x = layers.BatchNormalization(axis=bn_axis)(x)
x = layers.MaxPooling2D((3, 3), strides=strides, padding='same')(x)
x = layers.add([x, residual, cbam])
return x
def identity_block2(input_tensor, kernel_size, filters, stage, block):
'''The identity_block is the block that has no conv layer at shortcut
# Arguments
input_tensor: input tensor
kernel_size: defualt 3, the kernel size of middle conv layer at main path
filters: list of integers, the nb_filters of 3 conv layer at main path
stage: integer, current stage label, used for generating layer names
block: 'a','b'..., current block label, used for generating layer names
'''
nb_filter1, nb_filter2, nb_filter3 = filters
bn_axis = 3
# if K.image_dim_ordering() == 'tf':
# bn_axis = 3
# else:
# bn_axis = 1
conv_name_base = 'res' + str(stage) + block + '_branch'
bn_name_base = 'bn' + str(stage) + block + '_branch'
x = Conv2D(nb_filter1, 1, 1, name=conv_name_base + '2a')(input_tensor)
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2a')(x)
x = Activation('relu')(x)
x = Conv2D(nb_filter2, kernel_size, kernel_size,
border_mode='same', name=conv_name_base + '2b')(x)
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2b')(x)
x = Activation('relu')(x)
x = Conv2D(nb_filter3, 1, 1, name=conv_name_base + '2c')(x)
x = BatchNormalization(axis=bn_axis, name=bn_name_base + '2c')(x)
x = attach_attention_module(x, 'cbam_block')
x = add([x, input_tensor])
x = Activation('relu')(x)
return x
def CGS(data):
# 编码部分使用VGG166，不包含全连接层，输入图片大小定为224*224，最后一层是block5_pool(MaxPooling2D)大小为7*7*512
# base = VGG16(weights='imagenet', include_top=False, input_shape=(224, 224, 13), input_tensor=data)
# # 设置VGG模型部分为可训练
# base.trainable = True
x0 = layers.Conv2D(64, (3, 3),
activation='relu',
padding='same',
name='block1_conv')(data)
y = attach_attention_module(x0, 'cbam_block')
fusion = layers.add([x0, y])
x0 = layers.Activation('relu')(fusion)
x = layers.Conv2D(64, (3, 3),
activation='relu',
padding='same',
name='block1_conv2')(x0)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x1 = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block1_pool')(x)
x = layers.Conv2D(128, (3, 3),
activation='relu',
padding='same',
name='block2_conv1')(x1)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(128, (3, 3),
activation='relu',
padding='same',
name='block2_conv2')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x2 = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block2_pool')(x)
x = layers.Conv2D(256, (3, 3),
activation='relu',
padding='same',
name='block3_conv1')(x2)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(256, (3, 3),
activation='relu',
padding='same',
name='block3_conv2')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(256, (3, 3),
activation='relu',
padding='same',
name='block3_conv3')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x3 = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block3_pool')(x)
x = layers.Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='block4_conv1')(x3)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='block4_conv2')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='block4_conv3')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x4 = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block4_pool')(x)
x = layers.Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='block5_conv1')(x4)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='block5_conv2')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='block5_conv3')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x5 = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block5_pool')(x)
# ##自己写的
# # 14*14*512
x = layers.Deconv2D(512, (4, 4),
strides=(2, 2),
padding='same',
activation='relu',
name='deconv1')(x5)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Concatenate()([x, x4])
x = layers.Conv2D(512, (3, 3),
padding='same',
activation='relu',
name='deconv1_conv1')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(512, (3, 3),
padding='same',
activation='relu',
name='deconv1_conv2')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(512, (3, 3),
padding='same',
activation='relu',
name='deconv1_conv3')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
xx1 = layers.Activation('relu')(fusion)
# x = layers.Conv2D(512, (1, 1), padding='same', activation='relu', name='deconv1_conv4')(x)
# y = attach_attention_module(x, 'cbam_block')
# x = layers.add([x, y])
# x = layers.Activation('relu')(x)
# 28*28*256
x = layers.Deconv2D(256, (4, 4),
strides=(2, 2),
padding='same',
activation='relu',
name='deconv2')(xx1)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Concatenate()([x, x3])
x = layers.Conv2D(256, (3, 3),
padding='same',
activation='relu',
name='deconv2_conv1')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(256, (3, 3),
padding='same',
activation='relu',
name='deconv2_conv2')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(256, (3, 3),
padding='same',
activation='relu',
name='deconv2_conv3')(x)
# x = layers.Conv2D(256, (1, 1), padding='same', activation='relu', name='deconv2_conv4')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
xx2 = layers.Activation('relu')(fusion)
# 56*56*128
x = layers.Deconv2D(128, (4, 4),
strides=(2, 2),
padding='same',
activation='relu',
name='deconv3')(xx2)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Concatenate()([x, x2])
x = layers.Conv2D(128, (3, 3),
padding='same',
activation='relu',
name='deconv3_conv1')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(128, (3, 3),
padding='same',
activation='relu',
name='deconv3_conv2')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(128, (3, 3),
padding='same',
activation='relu',
name='deconv3_conv3')(x)
# x = layers.Conv2D(128, (1, 1), padding='same', activation='relu', name='deconv3_conv4')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
xx3 = layers.Activation('relu')(fusion)
# 112*112*64
x = layers.Deconv2D(64, (4, 4),
strides=(2, 2),
padding='same',
activation='relu',
name='deconv4')(xx3)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Concatenate()([x, x1])
x = layers.Conv2D(64, (3, 3),
padding='same',
activation='relu',
name='deconv4_conv1')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(64, (3, 3),
padding='same',
activation='relu',
name='deconv4_conv2')(x)
# x = layers.Conv2D(64, (1, 1), padding='same', activation='relu', name='deconv4_conv4')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
xx4 = layers.Activation('relu')(fusion)
# 224*224*32
x = layers.Deconv2D(32, (4, 4),
strides=(2, 2),
padding='same',
activation='relu',
name='deconv5')(xx4)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Concatenate()([x, x0])
x = layers.Conv2D(32, (3, 3),
padding='same',
activation='relu',
name='deconv5_conv1')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(32, (3, 3),
padding='same',
activation='relu',
name='deconv5_conv2')(x)
# x = layers.Conv2D(32, (1, 1), padding='same', activation='relu', name='deconv5_conv4')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
# 224*224*1 结果1
x0 = layers.Conv2D(1, (1, 1),
padding='same',
activation='relu',
name='conv6_end')(x)
# concat
d2 = layers.Lambda(lambda x: x[:, :, :, 3])(data)
y7 = layers.Lambda(lambda x: K.expand_dims(x, axis=3))(d2)
f = layers.Concatenate()([y7, x0])
# optimization
fusion1 = layers.Conv2D(64, (3, 3),
padding='same',
activation='relu',
name='fusion1_conv1')(f)
y = attach_attention_module(fusion1, 'cbam_block')
fusion1 = layers.add([fusion1, y])
fusion1 = layers.Activation('relu')(fusion1)
fusion1 = layers.Conv2D(64, (3, 3),
padding='same',
activation='relu',
name='fusion1_conv2')(fusion1)
y = attach_attention_module(fusion1, 'cbam_block')
fusion1 = layers.add([fusion1, y])
fusion1 = layers.Activation('relu')(fusion1)
fusion1 = layers.MaxPooling2D((2, 2),
strides=(2, 2),
padding='same',
name='fusion1_maxpooling')(fusion1)
fusion2 = layers.Conv2D(128, (3, 3),
padding='same',
activation='relu',
name='fusion2_conv1')(fusion1)
y = attach_attention_module(fusion2, 'cbam_block')
fusion2 = layers.add([fusion2, y])
fusion2 = layers.Activation('relu')(fusion2)
fusion2 = layers.Conv2D(128, (3, 3),
padding='same',
activation='relu',
name='fusion2_conv2')(fusion2)
y = attach_attention_module(fusion2, 'cbam_block')
fusion2 = layers.add([fusion2, y])
fusion2 = layers.Activation('relu')(fusion2)
fusion2 = layers.MaxPooling2D((2, 2),
strides=(2, 2),
padding='same',
name='fusion2_maxpooling')(fusion2)
fusion3 = layers.Conv2D(256, (3, 3),
padding='same',
activation='relu',
name='fusion3_conv1')(fusion2)
y = attach_attention_module(fusion3, 'cbam_block')
fusion3 = layers.add([fusion3, y])
fusion3 = layers.Activation('relu')(fusion3)
fusion3 = layers.Conv2D(256, (3, 3),
padding='same',
activation='relu',
name='fusion3_conv2')(fusion3)
y = attach_attention_module(fusion3, 'cbam_block')
fusion3 = layers.add([fusion3, y])
fusion3 = layers.Activation('relu')(fusion3)
fusion3 = layers.Conv2D(256, (3, 3),
padding='same',
activation='relu',
name='fusion3_conv3')(fusion3)
y = attach_attention_module(fusion3, 'cbam_block')
fusion3 = layers.add([fusion3, y])
fusion3 = layers.Activation('relu')(fusion3)
fusion3 = layers.MaxPooling2D((2, 2),
strides=(2, 2),
padding='same',
name='fusion3_maxpooling')(fusion3)
fusion4 = layers.Deconv2D(128, (4, 4),
strides=(2, 2),
padding='same',
activation='relu',
name='fusion4_deconv')(fusion3)
y = attach_attention_module(fusion4, 'cbam_block')
fusion4 = layers.add([fusion4, y])
fusion4 = layers.Activation('relu')(fusion4)
fusion4 = layers.Conv2D(128, (3, 3),
padding='same',
activation='relu',
name='fusion4_conv1')(fusion4)
y = attach_attention_module(fusion4, 'cbam_block')
fusion4 = layers.add([fusion4, y])
fusion4 = layers.Activation('relu')(fusion4)
fusion4 = layers.Conv2D(128, (3, 3),
padding='same',
activation='relu',
name='fusion4_conv2')(fusion4)
y = attach_attention_module(fusion4, 'cbam_block')
fusion4 = layers.add([fusion4, y])
fusion4 = layers.Activation('relu')(fusion4)
fusion4 = layers.Conv2D(128, (3, 3),
padding='same',
activation='relu',
name='fusion4_conv3')(fusion4)
y = attach_attention_module(fusion4, 'cbam_block')
fusion4 = layers.add([fusion4, y])
fusion4 = layers.Activation('relu')(fusion4)
fusion5 = layers.Deconv2D(64, (4, 4),
strides=(2, 2),
padding='same',
activation='relu',
name='fusion5_deconv')(fusion4)
y = attach_attention_module(fusion5, 'cbam_block')
fusion5 = layers.add([fusion5, y])
fusion5 = layers.Activation('relu')(fusion5)
fusion5 = layers.Conv2D(64, (3, 3),
padding='same',
activation='relu',
name='fusion5_conv1')(fusion5)
y = attach_attention_module(fusion5, 'cbam_block')
fusion5 = layers.add([fusion5, y])
fusion5 = layers.Activation('relu')(fusion5)
fusion5 = layers.Conv2D(64, (3, 3),
padding='same',
activation='relu',
name='fusion5_conv2')(fusion5)
y = attach_attention_module(fusion5, 'cbam_block')
fusion5 = layers.add([fusion5, y])
fusion5 = layers.Activation('relu')(fusion5)
fusion6 = layers.Deconv2D(32, (4, 4),
strides=(2, 2),
padding='same',
activation='relu',
name='fusion6_deconv')(fusion5)
y = attach_attention_module(fusion6, 'cbam_block')
fusion6 = layers.add([fusion6, y])
fusion6 = layers.Activation('relu')(fusion6)
fusion6 = layers.Conv2D(32, (1, 1),
padding='same',
activation='relu',
name='fusion6_conv1')(fusion6)
y = attach_attention_module(fusion6, 'cbam_block')
fusion6 = layers.add([fusion6, y])
fusion6 = layers.Activation('relu')(fusion6)
fusion6 = layers.Conv2D(1, (1, 1),
padding='same',
activation='relu',
name='fusion6_conv2')(fusion6)
# y = attach_attention_module(fusion6, 'cbam_block')
# fusion6 = layers.add([fusion6, y])
# fusion6 = layers.Activation('relu')(fusion6)
fusion = layers.Conv2D(1, (1, 1),
padding='same',
activation='sigmoid',
name='fusion')(fusion6)
# # train
# return [fusion, fusion]
# # return [x, fusion]
# test
m = Model(data, fusion, name='CGS')
return m
class_weight = class_weight.compute_class_weight(
'balanced', np.unique(train_generator.classes), train_generator.classes)
print(class_weight)
print("Downloading Base Model.....")
base_model = Xception(include_top=False, weights='imagenet')
for layer in base_model.layers:
layer.trainable = False
# get layers and add average pooling layer
## set model architechture
x = base_model.output
x = attach_attention_module(x, attention_module)
x = GlobalAveragePooling2D()(x)
x = Dropout(dropout)(x)
x = Dense(1024, activation='relu')(x)
predictions = Dense(no_of_classes, activation='softmax')(x)
lr_scheduler = LearningRateScheduler(lr_schedule)
lr_reducer = ReduceLROnPlateau(factor=np.sqrt(0.1),
cooldown=0,
patience=2,
min_lr=0.5e-6)
model = Model(input=base_model.input, output=predictions)
model.compile(optimizer=Adam(lr_schedule(0)),
loss='categorical_crossentropy',
metrics=['categorical_accuracy', 'accuracy'])
def res_Net50(input, classes=51, attention_module=None):
#global backend, layers, models, keras_utils
#backend, layers, models, keras_utils = get_submodules_from_kwargs(kwargs)
#x = layers.Lambda(finite_difference)(input)
#print(x.get_shape())
#exit()
if attention_module is not None:
x = attach_attention_module(input, 'fcbam_block')
x = layers.ZeroPadding2D(padding=(3, 3), name='conv1_pad')(input)
x = layers.Conv2D(128, (7, 7),
strides=(2, 2),
padding='valid',
kernel_initializer='he_normal',
name='conv1_he_normal')(x)
x = layers.BatchNormalization(name='bn_conv1_he_normal')(x)
x = layers.Activation('relu')(x)
x = layers.ZeroPadding2D(padding=(1, 1), name='pool1_pad_he_normal')(x)
x = layers.MaxPooling2D((3, 3), strides=(2, 2))(x)
if attention_module is not None:
x = attach_attention_module(x, attention_module)
x = conv_block(x,
3, [64, 64, 256],
stage=2,
block='a_he_normal',
strides=(1, 1))
x = identity_block(x, 3, [64, 64, 256], stage=2, block='b_he_normal')
x = identity_block(x, 3, [64, 64, 256], stage=2, block='c_he_normal')
if attention_module is not None:
x = attach_attention_module(x, attention_module)
x = conv_block(x, 3, [128, 128, 512], stage=3, block='a')
x = identity_block(x, 3, [128, 128, 512], stage=3, block='b')
x = identity_block(x, 3, [128, 128, 512], stage=3, block='c')
x = identity_block(x, 3, [128, 128, 512], stage=3, block='d')
if attention_module is not None:
x = attach_attention_module(x, attention_module)
x = conv_block(x, 3, [256, 256, 1024], stage=4, block='a')
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='b')
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='c')
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='d')
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='e')
x = identity_block(x, 3, [256, 256, 1024], stage=4, block='f')
if attention_module is not None:
x = attach_attention_module(x, attention_module)
x = conv_block(x, 3, [512, 512, 2048], stage=5, block='a')
x = identity_block(x, 3, [512, 512, 2048], stage=5, block='b')
x = identity_block(x, 3, [512, 512, 2048], stage=5, block='c')
x = layers.GlobalAveragePooling2D(name='avg_pool')(x)
# linear = layers.Dense(units=512,activation='sigmoid',name='dense_layer_1')(x)
# linear = layers.Dropout(rate=0.75)(linear)
linear = layers.Dense(units=classes,
activation='softmax',
name='dense_layer')(x)
model = Model(inputs=input, outputs=linear)
weights_path = utils.get_file(
'resnet50_weights_tf_dim_ordering_tf_kernels_notop.h5',
WEIGHTS_PATH_NO_TOP,
cache_subdir='models',
md5_hash='a268eb855778b3df3c7506639542a6af')
model.load_weights(weights_path, by_name=True)
return model
# for layer in model.layers[:172]:
# layer.trainable = False
# for layer in model.layers[172:]:
# layer.trainable = True
# change this code for every attribute - set the layers to true for training
for layer in base_model.layers:
layer.trainable = False
# for i, layer in enumerate(base_model.layers):
# print(i, layer.name)
# color attribute layer
color_attribute = base_model.output
color_attribute = attach_attention_module(x, attention_module)
color_attribute = GlobalAveragePooling2D()(color_attribute)
# let's add a fully-connected layer
color_attribute = Dropout(dropout)(color_attribute)
color_attribute = Dense(1024, activation='relu',
name="attribute_color")(color_attribute)
predictions_color = Dense(no_of_classes,
activation='softmax',
name="predictions_color")(color_attribute)
model = Model(inputs=base_model.input, outputs=predictions_color)
# model.load_weights("../models/label_color/label_color_inceptionv3_41_0.37.h5")
# print ("Checkpoint loaded.")
# this is the model we will train
def _depthwise_conv_block_v2(inputs, pointwise_conv_filters, alpha, expansion_factor,
depth_multiplier=1, strides=(1, 1), bn_epsilon=1e-3,
bn_momentum=0.99, weight_decay=0.0, block_id=1):
"""Adds a depthwise convolution block V2.
A depthwise convolution V2 block consists of a depthwise conv,
batch normalization, relu6, pointwise convolution,
batch normalization and relu6 activation.
# Arguments
inputs: Input tensor of shape `(rows, cols, channels)`
(with `channels_last` data format) or
(channels, rows, cols) (with `channels_first` data format).
pointwise_conv_filters: Integer, the dimensionality of the output space
(i.e. the number output of filters in the pointwise convolution).
alpha: controls the width of the network.
- If `alpha` < 1.0, proportionally decreases the number
of filters in each layer.
- If `alpha` > 1.0, proportionally increases the number
of filters in each layer.
- If `alpha` = 1, default number of filters from the paper
are used at each layer.
expansion_factor: controls the expansion of the internal bottleneck
blocks. Should be a positive integer >= 1
depth_multiplier: The number of depthwise convolution output channels
for each input channel.
The total number of depthwise convolution output
channels will be equal to `filters_in * depth_multiplier`.
strides: An integer or tuple/list of 2 integers,
specifying the strides of the convolution along the width and height.
Can be a single integer to specify the same value for
all spatial dimensions.
Specifying any stride value != 1 is incompatible with specifying
any `dilation_rate` value != 1.
bn_epsilon: Epsilon value for BatchNormalization
bn_momentum: Momentum value for BatchNormalization
block_id: Integer, a unique identification designating the block number.
# Input shape
4D tensor with shape:
`(batch, channels, rows, cols)` if data_format='channels_first'
or 4D tensor with shape:
`(batch, rows, cols, channels)` if data_format='channels_last'.
# Output shape
4D tensor with shape:
`(batch, filters, new_rows, new_cols)` if data_format='channels_first'
or 4D tensor with shape:
`(batch, new_rows, new_cols, filters)` if data_format='channels_last'.
`rows` and `cols` values might have changed due to stride.
# Returns
Output tensor of block.
"""
channel_axis = 1 if K.image_data_format() == 'channels_first' else -1
input_shape = K.int_shape(inputs)
depthwise_conv_filters = _make_divisible(
input_shape[channel_axis] * expansion_factor)
pointwise_conv_filters = _make_divisible(pointwise_conv_filters * alpha)
if depthwise_conv_filters > input_shape[channel_axis]:
x = Conv2D(depthwise_conv_filters, (1, 1),
padding='same',
use_bias=False,
strides=(1, 1),
kernel_initializer=initializers.he_normal(),
kernel_regularizer=regularizers.l2(weight_decay),
name='conv_expand_%d' % block_id)(inputs)
x = BatchNormalization(axis=channel_axis, momentum=bn_momentum, epsilon=bn_epsilon,
name='conv_expand_%d_bn' % block_id)(x)
x = Activation(relu6, name='conv_expand_%d_relu' % block_id)(x)
else:
x = inputs
x = DepthwiseConv2D((3, 3),
padding='same',
depth_multiplier=depth_multiplier,
strides=strides,
use_bias=False,
depthwise_initializer=initializers.he_normal(),
depthwise_regularizer=regularizers.l2(weight_decay),
name='conv_dw_%d' % block_id)(x)
x = BatchNormalization(axis=channel_axis, momentum=bn_momentum, epsilon=bn_epsilon,
name='conv_dw_%d_bn' % block_id)(x)
x = Activation(relu6, name='conv_dw_%d_relu' % block_id)(x)
x = Conv2D(pointwise_conv_filters, (1, 1),
padding='same',
use_bias=False,
strides=(1, 1),
kernel_initializer=initializers.he_normal(),
kernel_regularizer=regularizers.l2(weight_decay),
name='conv_pw_%d' % block_id)(x)
x = BatchNormalization(axis=channel_axis, momentum=bn_momentum, epsilon=bn_epsilon,
name='conv_pw_%d_bn' % block_id)(x)
if strides == (2, 2):
return x
else:
if input_shape[channel_axis] == pointwise_conv_filters:
x = add([inputs, x])
x = attach_attention_module(x, 'cbam_block')
return x
def ResCBAM(in_image=(height, width, 1)):
img_in = Input(shape=in_image, name='image_in')
img_in_b = img_in
x0 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv0_0')(img_in_b)
y0 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv0_1')(img_in_b)
y0 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv0_2')(y0)
y0 = attach_attention_module(y0, attention_module='cbam_block')
conv0 = add([x0, y0])
pool1 = MaxPooling2D(pool_size=(2, 2), name='down0')(conv0)
x1 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv1_0')(pool1)
#x1 = BatchNormalization(x1)
y1 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv1_1')(pool1)
#y1 = BatchNormalization(y1)
y1 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv1_2')(y1)
#y1 = BatchNormalization(y1)
y1 = attach_attention_module(y1, attention_module='cbam_block')
conv1 = add([x1, y1])
pool2 = MaxPooling2D(pool_size=(2, 2), name='down1')(conv1)
x2 = Conv2D(256,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv2_0')(pool2)
# x2 = BatchNormalization(x2)
y2 = Conv2D(256,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv2_1')(pool2)
#y2 = BatchNormalization(y2)
y2 = Conv2D(256,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv2_2')(y2)
#y2 = BatchNormalization(y2)
y2 = attach_attention_module(y2, attention_module='cbam_block')
conv2 = add([x2, y2])
pool3 = MaxPooling2D(pool_size=(2, 2), name='down2')(conv2)
x3 = Conv2D(512,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv3_0')(pool3)
#x3 = BatchNormalization(x3)
y3 = Conv2D(512,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv3_1')(pool3)
#y3 = BatchNormalization(y2)
y3 = Conv2D(512,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv3_2')(y3)
# y3 = BatchNormalization(y2)
y3 = attach_attention_module(y3, attention_module='cbam_block')
conv3 = add([x3, y3])
pool3 = AveragePooling2D(pool_size=2, name='down3')(conv3)
down_4_f = Flatten(name='down_2_flat')(pool3)
classification = Dense(3, activation='sigmoid',
name='classification')(down_4_f)
conv5 = Conv2D(256,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv5_1')(pool3)
conv5 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv5_2')(conv5)
drop5 = Dropout(0.2)(conv5)
x6 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv6_0')(drop5)
y6 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv6_1')(drop5)
y6 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv6_2')(y6)
y6 = attach_attention_module(y6, attention_module='cbam_block')
xy6 = add([x6, y6])
up6 = UpSampling2D(size=(2, 2), name='up6')(xy6)
merge6 = concatenate([conv3, up6], axis=3)
x7 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv7_0')(merge6)
y7 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv7_1')(merge6)
y7 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv7_2')(y7)
y7 = attach_attention_module(y7, attention_module='cbam_block')
up7 = add([x7, y7])
up7 = UpSampling2D(size=(2, 2), name='up7')(up7)
merge7 = concatenate([conv2, up7], axis=3)
x8 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv8_0')(up7)
y8 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv8_1')(up7)
y8 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv8_2')(y8)
y8 = attach_attention_module(y8, attention_module='cbam_block')
up8 = add([x8, y8])
up8 = UpSampling2D(size=(2, 2), name='up8')(up8)
merge8 = concatenate([conv1, up8], axis=3)
x9 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv9_0')(merge8)
y9 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv9_1')(merge8)
y9 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv9_2')(y9)
y9 = attach_attention_module(y9, attention_module='cbam_block')
XY9 = add([x9, y9])
up9 = UpSampling2D(size=(2, 2), name='Up9')(XY9)
merge9 = concatenate([conv0, up9], axis=3)
conv9 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv9_4')(merge9)
conv9 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv9_5')(conv9)
segmentation = Conv2D(1, 1, activation='sigmoid',
name='segmentation')(merge9)
model = Model(inputs=img_in, outputs=[segmentation, classification])
model.summary()
plot_model(model, to_file='model.png')
return model
# for layer in model.layers[:172]:
# layer.trainable = False
# for layer in model.layers[172:]:
# layer.trainable = True
# change this code for every attribute - set the layers to true for training
for layer in base_model.layers:
layer.trainable = False
# for i, layer in enumerate(base_model.layers):
# print(i, layer.name)
# pattern attribute layer
pattern_attribute = base_model.output
pattern_attribute = attach_attention_module(pattern_attribute,
attention_module)
pattern_attribute = GlobalAveragePooling2D()(pattern_attribute)
# let's add a fully-connected layer
pattern_attribute = Dropout(dropout)(pattern_attribute)
pattern_attribute_layer = Dense(1024,
activation='relu',
name="attribute_pattern")(pattern_attribute)
pattern_attribute = Dropout(dropout)(pattern_attribute)
predictions_pattern = Dense(
no_of_classes, activation='softmax',
name="predictions_pattern")(pattern_attribute_layer)
model = Model(inputs=base_model.input, outputs=predictions_pattern)
# model.load_weights("../models/label_pattern/label_pattern_inceptionv3_41_0.37.h5")
# print ("Checkpoint loaded.")
def Vgg16CBAM(in_image=(height, width, 1)):
img_in = Input(shape=in_image, name='image_in')
img_in_b = BatchNormalization(name='in_BN')(img_in)
#c0 = Conv2D(8, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_normal',name='c0')(img_in_b)
c1 = Erosion2D(8, (3, 3), padding="same", strides=(1, 1))(img_in_b)
c1 = Dilation2D(8, (3, 3), padding="same", strides=(1, 1))(c1)
c1 = BatchNormalization()(c1)
c1 = attach_attention_module(c1, attention_module='cbam_block')
#c10 = add([c0,c1])
cc0 = BatchNormalization(name='Morph_BN')(c1)
cc00 = Conv2D(8,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='cc0')(cc0)
cc1 = Erosion2D(8, (3, 3), padding="same", strides=(1, 1))(cc0)
cc1 = Dilation2D(8, (3, 3), padding="same", strides=(1, 1))(cc1)
cc1 = BatchNormalization()(cc1)
cc1 = attach_attention_module(cc1, attention_module='cbam_block')
#cc10 = add([cc00,cc1])
conv0 = Conv2D(32,
6,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv0_5')(cc1)
conv0 = Conv2D(32,
6,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv0_6')(conv0)
conv0 = BatchNormalization()(conv0)
#conv0 = attach_attention_module(conv0, attention_module='cbam_block')
pool0 = MaxPooling2D(pool_size=(2, 2), name='down0')(conv0)
conv1 = Conv2D(64,
5,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv1_1')(pool0)
conv1 = Conv2D(64,
5,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv1_2')(conv1)
conv1 = BatchNormalization()(conv1)
#conv1 = attach_attention_module(conv1, attention_module='cbam_block')
pool1 = MaxPooling2D(pool_size=(2, 2), name='down1')(conv1)
conv2 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv2_1')(pool1)
conv2 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv2_2')(conv2)
conv2 = BatchNormalization()(conv2)
# conv2 = attach_attention_module(conv2, attention_module='cbam_block')
pool2 = MaxPooling2D(pool_size=(2, 2), name='down2')(conv2)
conv3 = Conv2D(256,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv3_1')(pool2)
conv3 = Conv2D(256,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv3_2')(conv3)
conv3 = BatchNormalization()(conv3)
#conv3 = attach_attention_module(conv3, attention_module='cbam_block')
pool3 = MaxPooling2D(pool_size=(2, 2), name='down3')(conv3)
conv4 = Conv2D(512,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv4_1')(pool3)
conv4 = Conv2D(512,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv4_2')(conv4)
conv4 = BatchNormalization()(conv4)
#conv4 = attach_attention_module(conv4, attention_module='cbam_block')
pool4 = MaxPooling2D(pool_size=(2, 2), name='down4')(conv4)
down_4_f = Flatten(name='down_2_flat')(pool4)
down_classsify = Dense(512, activation='relu', name='classify_1')(down_4_f)
down_classsify = Dropout(0.6)(down_classsify)
down_classsify = Dense(128, activation='relu',
name='classify_2')(down_classsify)
down_classsify = Dropout(0.65)(down_classsify)
classification = Dense(2, activation='sigmoid',
name='classification')(down_classsify)
conv5 = Conv2D(512,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv5_1')(pool4)
conv5 = Conv2D(512,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv5_2')(conv5)
conv5 = BatchNormalization()(conv5)
up6 = Conv2D(512,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv6_1')(conv5)
up6 = UpSampling2D(size=(2, 2), name='up_1')(up6)
merge6 = concatenate([conv4, up6], axis=3)
conv6 = Conv2D(512,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv6_2')(merge6)
conv6 = Conv2D(512,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv6_3')(conv6)
conv6 = BatchNormalization()(conv6)
up7 = Conv2D(256,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv7_1')(conv6)
up7 = UpSampling2D(size=(2, 2), name='up2')(up7)
merge7 = concatenate([conv3, up7], axis=3)
conv7 = Conv2D(256,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv7_2')(merge7)
conv7 = Conv2D(256,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv7_3')(conv7)
conv7 = BatchNormalization()(conv7)
up8 = Conv2D(128,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv8_1')(conv7)
up8 = UpSampling2D(size=(2, 2), name='up3')(up8)
merge8 = concatenate([conv2, up8], axis=3)
conv8 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv8_2')(merge8)
conv8 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv8_3')(conv8)
conv8 = BatchNormalization()(conv8)
up9 = Conv2D(64,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv9_1')(conv8)
up9 = UpSampling2D(size=(2, 2), name='up4')(up9)
merge9 = concatenate([conv1, up9], axis=3)
conv9 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv9_2')(merge9)
conv9 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv9_3')(conv9)
conv9 = BatchNormalization()(conv9)
up10 = Conv2D(32,
2,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv10_1')(conv9)
up10 = UpSampling2D(size=(2, 2), name='up5')(up10)
merge10 = concatenate([conv0, up10], axis=3)
conv10 = Conv2D(32,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv10_2')(merge10)
conv10 = Conv2D(32,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv10_3')(conv10)
conv10 = BatchNormalization()(conv10)
conv10 = Conv2D(2,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv10_4')(conv10)
segmentation = Conv2D(1, 1, activation='sigmoid',
name='segmentation')(conv10)
model = Model(inputs=img_in, outputs=[segmentation, classification])
model.summary()
plot_model(model, to_file='model.png')
return model
def ResCBAM1(in_image=(height, width, 1)):
img_in = Input(shape=in_image, name='image_in')
img_in_b = BatchNormalization(name='in_BN')(img_in)
x0 = Conv2D(32,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv0_0')(img_in_b)
y0 = Conv2D(32,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv0_1')(img_in_b)
y0 = Conv2D(32,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv0_2')(y0)
y0 = attach_attention_module(y0, attention_module='cbam_block')
conv0 = add([x0, y0]) #32
x1 = conv0
y1 = Conv2D(32,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv1_1')(conv0)
y1 = Conv2D(32,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv1_2')(y1)
y1 = attach_attention_module(y1, attention_module='cbam_block')
conv1 = add([x1, y1]) #32
x2 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv2_0')(conv1)
y2 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv2_1')(conv1)
y2 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv2_2')(y2)
y2 = attach_attention_module(y2, attention_module='cbam_block')
conv2 = add([x2, y2]) #64
x3 = conv2
y3 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv3_1')(conv2)
y3 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv3_2')(y3)
y3 = attach_attention_module(y3, attention_module='cbam_block')
conv3 = add([x3, y3]) #64
x4 = conv3
y4 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv4_1')(conv3)
y4 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv4_2')(y4)
y4 = attach_attention_module(y4, attention_module='cbam_block')
conv4 = add([x4, y4])
x5 = conv4
y5 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv5_1')(conv4)
y5 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv5_2')(y5)
y5 = attach_attention_module(y5, attention_module='cbam_block')
conv5 = add([x5, y5])
x6 = conv5
y6 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv6_1')(conv5)
y6 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv6_2')(y6)
y6 = attach_attention_module(y6, attention_module='cbam_block')
conv6 = add([x6, y6])
x7 = conv6
y7 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv7_1')(conv6)
y7 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal',
name='conv7_2')(y6)
y7 = attach_attention_module(y7, attention_module='cbam_block')
conv7 = add([x7, y7])
segmentation = Conv2D(1, 1, activation='sigmoid',
name='segmentation')(conv7)
pool1 = AveragePooling2D(pool_size=8, name='AvePool1')(conv7)
down_1 = Flatten(name='down_1')(pool1)
classification = Dense(2, activation='sigmoid',
name='classification')(down_1)
model = Model(inputs=img_in, outputs=[segmentation, classification])
model.summary()
plot_model(model, to_file='model.png')
return model
def CGScom(data):
# 编码部分使用VGG166，不包含全连接层，输入图片大小定为224*224，最后一层是block5_pool(MaxPooling2D)大小为7*7*512
# base = VGG16(weights='imagenet', include_top=False, input_shape=(224, 224, 13), input_tensor=data)
# # 设置VGG模型部分为可训练
# base.trainable = True
x0 = layers.Conv2D(64, (3, 3),
activation='relu',
padding='same',
name='block1_conv')(data)
y = attach_attention_module(x0, 'cbam_block')
fusion = layers.add([x0, y])
x0 = layers.Activation('relu')(fusion)
x = layers.Conv2D(64, (3, 3),
activation='relu',
padding='same',
name='block1_conv2')(x0)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x1 = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block1_pool')(x)
x = layers.Conv2D(128, (3, 3),
activation='relu',
padding='same',
name='block2_conv1')(x1)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(128, (3, 3),
activation='relu',
padding='same',
name='block2_conv2')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x2 = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block2_pool')(x)
x = layers.Conv2D(256, (3, 3),
activation='relu',
padding='same',
name='block3_conv1')(x2)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(256, (3, 3),
activation='relu',
padding='same',
name='block3_conv2')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(256, (3, 3),
activation='relu',
padding='same',
name='block3_conv3')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x3 = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block3_pool')(x)
x = layers.Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='block4_conv1')(x3)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='block4_conv2')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='block4_conv3')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x4 = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block4_pool')(x)
x = layers.Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='block5_conv1')(x4)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='block5_conv2')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(512, (3, 3),
activation='relu',
padding='same',
name='block5_conv3')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x5 = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block5_pool')(x)
# ##自己写的
# # 14*14*512
x = layers.Deconv2D(512, (4, 4),
strides=(2, 2),
padding='same',
activation='relu',
name='deconv1')(x5)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Concatenate()([x, x4])
x = layers.Conv2D(512, (3, 3),
padding='same',
activation='relu',
name='deconv1_conv1')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(512, (3, 3),
padding='same',
activation='relu',
name='deconv1_conv2')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(512, (3, 3),
padding='same',
activation='relu',
name='deconv1_conv3')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
xx1 = layers.Activation('relu')(fusion)
# x = layers.Conv2D(512, (1, 1), padding='same', activation='relu', name='deconv1_conv4')(x)
# y = attach_attention_module(x, 'cbam_block')
# x = layers.add([x, y])
# x = layers.Activation('relu')(x)
# 28*28*256
x = layers.Deconv2D(256, (4, 4),
strides=(2, 2),
padding='same',
activation='relu',
name='deconv2')(xx1)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Concatenate()([x, x3])
x = layers.Conv2D(256, (3, 3),
padding='same',
activation='relu',
name='deconv2_conv1')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(256, (3, 3),
padding='same',
activation='relu',
name='deconv2_conv2')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(256, (3, 3),
padding='same',
activation='relu',
name='deconv2_conv3')(x)
# x = layers.Conv2D(256, (1, 1), padding='same', activation='relu', name='deconv2_conv4')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
xx2 = layers.Activation('relu')(fusion)
# 56*56*128
x = layers.Deconv2D(128, (4, 4),
strides=(2, 2),
padding='same',
activation='relu',
name='deconv3')(xx2)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Concatenate()([x, x2])
x = layers.Conv2D(128, (3, 3),
padding='same',
activation='relu',
name='deconv3_conv1')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(128, (3, 3),
padding='same',
activation='relu',
name='deconv3_conv2')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(128, (3, 3),
padding='same',
activation='relu',
name='deconv3_conv3')(x)
# x = layers.Conv2D(128, (1, 1), padding='same', activation='relu', name='deconv3_conv4')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
xx3 = layers.Activation('relu')(fusion)
# 112*112*64
x = layers.Deconv2D(64, (4, 4),
strides=(2, 2),
padding='same',
activation='relu',
name='deconv4')(xx3)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Concatenate()([x, x1])
x = layers.Conv2D(64, (3, 3),
padding='same',
activation='relu',
name='deconv4_conv1')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(64, (3, 3),
padding='same',
activation='relu',
name='deconv4_conv2')(x)
# x = layers.Conv2D(64, (1, 1), padding='same', activation='relu', name='deconv4_conv4')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
xx4 = layers.Activation('relu')(fusion)
# 224*224*32
x = layers.Deconv2D(32, (4, 4),
strides=(2, 2),
padding='same',
activation='relu',
name='deconv5')(xx4)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Concatenate()([x, x0])
x = layers.Conv2D(32, (3, 3),
padding='same',
activation='relu',
name='deconv5_conv1')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
x = layers.Conv2D(32, (3, 3),
padding='same',
activation='relu',
name='deconv5_conv2')(x)
# x = layers.Conv2D(32, (1, 1), padding='same', activation='relu', name='deconv5_conv4')(x)
y = attach_attention_module(x, 'cbam_block')
fusion = layers.add([x, y])
x = layers.Activation('relu')(fusion)
# 224*224*1 结果1
x0 = layers.Conv2D(1, (1, 1),
padding='same',
activation='sigmoid',
name='conv6_end')(x)
# # train
# return [x0, x0]
# # return [x, fusion]
# test
m = Model(data, x0, name='CGScom')
return m
# for layer in model.layers[:172]:
# layer.trainable = False
# for layer in model.layers[172:]:
# layer.trainable = True
# change this code for every attribute - set the layers to true for training
for layer in base_model.layers:
layer.trainable = False
# for i, layer in enumerate(base_model.layers):
# print(i, layer.name)
# gender attribute layer
gender_attribute = base_model.output
gender_attribute = attach_attention_module(gender_attribute, attention_module)
gender_attribute = GlobalAveragePooling2D()(gender_attribute)
# let's add a fully-connected layer
gender_attribute = Dropout(dropout)(gender_attribute)
gender_attribute_layer = Dense(1024,
activation='relu',
name="attribute_gender")(gender_attribute)
predictions_gender = Dense(no_of_classes,
activation='softmax',
name="predictions_gender")(gender_attribute_layer)
model = Model(inputs=base_model.input, outputs=predictions_gender)
lr_scheduler = LearningRateScheduler(lr_schedule)
lr_reducer = ReduceLROnPlateau(factor=np.sqrt(0.1),
cooldown=0,
def _depthwise_conv_block(inputs, pointwise_conv_filters, alpha,
depth_multiplier=1, strides=(1, 1), block_id=1, attention_module=None):
"""Adds a depthwise convolution block.
A depthwise convolution block consists of a depthwise conv,
batch normalization, relu6, pointwise convolution,
batch normalization and relu6 activation.
# Arguments
inputs: Input tensor of shape `(rows, cols, channels)`
(with `channels_last` data format) or
(channels, rows, cols) (with `channels_first` data format).
pointwise_conv_filters: Integer, the dimensionality of the output space
(i.e. the number output of filters in the pointwise convolution).
alpha: controls the width of the network.
- If `alpha` < 1.0, proportionally decreases the number
of filters in each layer.
- If `alpha` > 1.0, proportionally increases the number
of filters in each layer.
- If `alpha` = 1, default number of filters from the paper
are used at each layer.
depth_multiplier: The number of depthwise convolution output channels
for each input channel.
The total number of depthwise convolution output
channels will be equal to `filters_in * depth_multiplier`.
strides: An integer or tuple/list of 2 integers,
specifying the strides of the convolution along the width and height.
Can be a single integer to specify the same value for
all spatial dimensions.
Specifying any stride value != 1 is incompatible with specifying
any `dilation_rate` value != 1.
block_id: Integer, a unique identification designating the block number.
# Input shape
4D tensor with shape:
`(batch, channels, rows, cols)` if data_format='channels_first'
or 4D tensor with shape:
`(batch, rows, cols, channels)` if data_format='channels_last'.
# Output shape
4D tensor with shape:
`(batch, filters, new_rows, new_cols)` if data_format='channels_first'
or 4D tensor with shape:
`(batch, new_rows, new_cols, filters)` if data_format='channels_last'.
`rows` and `cols` values might have changed due to stride.
# Returns
Output tensor of block.
"""
channel_axis = 1 if K.image_data_format() == 'channels_first' else -1
pointwise_conv_filters = int(pointwise_conv_filters * alpha)
x = DepthwiseConv2D((3, 3),
padding='same',
depth_multiplier=depth_multiplier,
strides=strides,
use_bias=False,
name='conv_dw_%d' % block_id)(inputs)
x = BatchNormalization(axis=channel_axis, name='conv_dw_%d_bn' % block_id)(x)
x = Activation(relu6, name='conv_dw_%d_relu' % block_id)(x)
x = Conv2D(pointwise_conv_filters, (1, 1),
padding='same',
use_bias=False,
strides=(1, 1),
name='conv_pw_%d' % block_id)(x)
x = BatchNormalization(axis=channel_axis, name='conv_pw_%d_bn' % block_id)(x)
x = Activation(relu6, name='conv_pw_%d_relu' % block_id)(x)
# attention_module
if attention_module is not None:
x = attach_attention_module(x, attention_module)
return x
