def head(endpoints, embedding_dim, is_training):
endpoints['emb1'] = GlobalAveragePooling2D(data_format='channels_last')(endpoints['Conv2d_1_pointwise'])
endpoints['fc_layer1_1'] = Dense(1024, activation='relu', kernel_initializer='Orthogonal')(endpoints['emb1'])
endpoints['bn_1_1'] = BN()(endpoints['fc_layer1_1'])
endpoints['feature1'] = Dense(embedding_dim, activation=None, kernel_initializer='Orthogonal')(endpoints['bn_1_1'])
endpoints['emb2'] = GlobalAveragePooling2D(data_format='channels_last')(endpoints['Conv2d_3_pointwise'])
endpoints['fc_layer2_1'] = Dense(1024, activation='relu', kernel_initializer='Orthogonal')(endpoints['emb2'])
endpoints['bn_2_1'] = BN()(endpoints['fc_layer2_1'])
endpoints['feature2'] = Dense(embedding_dim, activation=None, kernel_initializer='Orthogonal')(endpoints['bn_2_1'])
endpoints['emb3'] = GlobalAveragePooling2D(data_format='channels_last')(endpoints['Conv2d_5_pointwise'])
endpoints['fc_layer3_1'] = Dense(1024, activation='relu', kernel_initializer='Orthogonal')(endpoints['emb3'])
endpoints['bn_3_1'] = BN()(endpoints['fc_layer3_1'])
endpoints['feature3'] = Dense(embedding_dim, activation=None, kernel_initializer='Orthogonal')(endpoints['bn_3_1'])
endpoints['emb4'] = GlobalAveragePooling2D(data_format='channels_last')(endpoints['Conv2d_11_pointwise'])
endpoints['fc_layer4_1'] = Dense(1024, activation='relu', kernel_initializer='Orthogonal')(endpoints['emb4'])
endpoints['bn_4_1'] = BN()(endpoints['fc_layer4_1'])
endpoints['feature4'] = Dense(embedding_dim, activation=None, kernel_initializer='Orthogonal')(endpoints['bn_4_1'])
endpoints['emb5'] = GlobalAveragePooling2D(data_format='channels_last')(endpoints['Conv2d_13_pointwise'])
endpoints['fc_layer5_1'] = Dense(1024, activation='relu', kernel_initializer='Orthogonal')(endpoints['emb5'])
endpoints['bn_5_1'] = BN()(endpoints['fc_layer5_1'])
endpoints['feature5'] = Dense(embedding_dim, activation=None, kernel_initializer='Orthogonal')(endpoints['bn_5_1'])
endpoints['fusion_layer'] = FusionLayer()([endpoints['feature1'], endpoints['feature2'], endpoints['feature3'],
endpoints['feature4'],endpoints['feature5']])
endpoints['emb'] = endpoints['emb_raw'] = endpoints['fusion_layer']
return endpoints
def Conv2D_Acti_BN(input_tensor, activation, *args):
output_tensor = Conv2D(*args,
activation=activation,
padding='same',
kernel_initializer='he_normal')(input_tensor)
output_tensor = BN()(output_tensor)
return output_tensor
def Conv2D_BN_Leaky(input_tensor, *args):
output_tensor = Conv2D(*args,
padding='same',
kernel_initializer='he_normal')(input_tensor)
output_tensor = BN()(output_tensor)
output_tensor = LeakyReLU(alpha=0.1)(output_tensor)
return output_tensor
def model_wj(ts_in:Input):
ts_z = ts_in
'''-----Embadding'''
### 256->128
ts_z = Conv2D( filters=16, kernel_size=3, strides=2, padding='same', use_bias=False )(ts_z)
ts_z = BN()(ts_z)
ts_z = GELU()(ts_z)
ts_z = Conv2D( filters=32, kernel_size=3, strides=1, padding='same', use_bias=False )(ts_z)
'''----Body'''
def BRC(x,filters,down=False):
s = 2 if down else 1
Cin = x.shape[3]
if Cin==filters and down==False:
P = x
else:
P = Conv2D( filters=filters, kernel_size=1, strides=s, padding='same', use_bias=False )(x)
P = BN()(P)
if down :
x = AveragePooling2D()(x)
x = BN()(x)
x = GELU()(x)
x = Conv2D( filters=filters, kernel_size=5, strides=1, padding='same', use_bias=False )(x)
# x = BN()(x)
# x = GELU()(x)
# x = Conv2D( filters=filters, kernel_size=3, padding='same', use_bias=False )(x)
x = P+x
return x
### 128->64
ts_z = BRC(ts_z,filters=64,down=True)
### 64->32
ts_z = BRC(ts_z,filters=128,down=True)
### 32->16
ts_z = BRC(ts_z,filters=256,down=True)
'''----Exit'''
# ts_z = BRC(ts_z,filters=128)
### 16->8
ts_z = BRC(ts_z,filters=512,down=True)
'''----GAP'''
ts_feature = GAP2D()(ts_z)
ts_z = ts_feature
'''----FC'''
ts_z = Dense( units=26, activation=None, use_bias=False)(ts_z)
ts_out = Activation('linear', dtype=tf.float32)(ts_z)
model = keras.Model(inputs=ts_in,outputs=ts_out)
return model
def BRC(x,filters,down=False):
s = 2 if down else 1
Cin = x.shape[3]
if Cin==filters and down==False:
P = x
else:
P = Conv2D( filters=filters, kernel_size=1, strides=s, padding='same', use_bias=False )(x)
P = BN()(P)
if down :
x = AveragePooling2D()(x)
x = BN()(x)
x = GELU()(x)
x = Conv2D( filters=filters, kernel_size=5, strides=1, padding='same', use_bias=False )(x)
# x = BN()(x)
# x = GELU()(x)
# x = Conv2D( filters=filters, kernel_size=3, padding='same', use_bias=False )(x)
x = P+x
return x
def build(self, classes=CLASSES):
INPUT_SHAPE = (self.height, self.width, self.dept)
channel_dimension = -1
if k == "channels_first":
INPUT_SHAPE = (self.dept, self.height, self.width)
channel_dimension = 1
if classes > 2:
FINAL_ACTIVATION = "softmax"
model = Sequential()
# first layer
model.add(Conv2D(32, kernel_size=KERNEL_SIZE, padding=PADDING, activation=ACTIVATION, input_shape=INPUT_SHAPE))
model.add(BN(axis=channel_dimension))
# second layer
model.add(Conv2D(32, kernel_size=KERNEL_SIZE, padding=PADDING, activation=ACTIVATION))
model.add(BN(axis=channel_dimension))
model.add(MaxPool2D(pool_size=POOL_SIZE, padding=PADDING))
model.add(Dropout(rate=DROPOUT1))
# third layer
model.add(Conv2D(64, kernel_size=KERNEL_SIZE, padding=PADDING, activation=ACTIVATION))
model.add(BN(axis=channel_dimension))
# fourth layer
model.add(Conv2D(64, kernel_size=KERNEL_SIZE, padding=PADDING, activation=ACTIVATION))
model.add(BN(axis=channel_dimension))
model.add(MaxPool2D(pool_size=POOL_SIZE))
model.add(Dropout(rate=DROPOUT1))
# last layer
model.add(Flatten())
model.add(Dense(units=500, activation=ACTIVATION))
model.add(BN(axis=channel_dimension))
model.add(Dropout(rate=DROPOUT2))
model.add(Dense(units=classes, activation=FINAL_ACTIVATION))
return model
for i in range(K - 1):
fltr = fltr.dot(fltr)
fltr.sort_indices()
# Model definition
#GCN input
area_in = Input(shape=(F, )) #area feature
fltr_in = Input((N, ), sparse=True) #adjacent matrix
bn0 = BN(
axis=-1,
momentum=0.99,
epsilon=0.001,
center=True,
scale=True,
beta_initializer='zeros',
gamma_initializer='ones',
moving_mean_initializer='zeros',
moving_variance_initializer='ones',
renorm_momentum=0.99,
)(area_in) #batch norm on the area data
if not include_batch_norm_layers:
bn0 = area_in
if not include_batch_norm_layers:
g1 = GraphConv(
embedding_vecor_length1, #first graph conv layer
activation='relu',
kernel_regularizer=l2(l2_reg),
use_bias=True)([bn0, fltr_in])
def head(endpoints, embedding_dim, backbone_model, is_training):
if backbone_model == 'mobilenet_v1_1_224':
endpoints['emb1'] = GlobalAveragePooling2D(data_format='channels_last')(endpoints['Conv2d_1_pointwise'])
endpoints['fc_layer1_1'] = Dense(1024, activation='relu', kernel_initializer='Orthogonal')(endpoints['emb1'])
endpoints['bn_1_1'] = BN()(endpoints['fc_layer1_1'])
endpoints['feature1'] = Dense(embedding_dim, activation=None, kernel_initializer='Orthogonal')(
endpoints['bn_1_1'])
endpoints['emb2'] = GlobalAveragePooling2D(data_format='channels_last')(endpoints['Conv2d_3_pointwise'])
endpoints['fc_layer2_1'] = Dense(1024, activation='relu', kernel_initializer='Orthogonal')(endpoints['emb2'])
endpoints['bn_2_1'] = BN()(endpoints['fc_layer2_1'])
endpoints['feature2'] = Dense(embedding_dim, activation=None, kernel_initializer='Orthogonal')(
endpoints['bn_2_1'])
endpoints['emb3'] = GlobalAveragePooling2D(data_format='channels_last')(endpoints['Conv2d_5_pointwise'])
endpoints['fc_layer3_1'] = Dense(1024, activation='relu', kernel_initializer='Orthogonal')(endpoints['emb3'])
endpoints['bn_3_1'] = BN()(endpoints['fc_layer3_1'])
endpoints['feature3'] = Dense(embedding_dim, activation=None, kernel_initializer='Orthogonal')(
endpoints['bn_3_1'])
endpoints['emb4'] = GlobalAveragePooling2D(data_format='channels_last')(endpoints['Conv2d_11_pointwise'])
endpoints['fc_layer4_1'] = Dense(1024, activation='relu', kernel_initializer='Orthogonal')(endpoints['emb4'])
endpoints['bn_4_1'] = BN()(endpoints['fc_layer4_1'])
endpoints['feature4'] = Dense(embedding_dim, activation=None, kernel_initializer='Orthogonal')(
endpoints['bn_4_1'])
endpoints['emb5'] = GlobalAveragePooling2D(data_format='channels_last')(endpoints['Conv2d_13_pointwise'])
endpoints['fc_layer5_1'] = Dense(1024, activation='relu', kernel_initializer='Orthogonal')(endpoints['emb5'])
endpoints['bn_5_1'] = BN()(endpoints['fc_layer5_1'])
endpoints['feature5'] = Dense(embedding_dim, activation=None, kernel_initializer='Orthogonal')(
endpoints['bn_5_1'])
endpoints['fusion_layer'] = FusionLayer_mob()([endpoints['feature1'], endpoints['feature2'],
endpoints['feature3'], endpoints['feature4'],
endpoints['feature5']])
endpoints['emb'] = endpoints['emb_raw'] = endpoints['fusion_layer']
return endpoints
elif backbone_model == 'resnet_v1_50':
endpoints['emb1'] = GlobalAveragePooling2D(data_format='channels_last')(endpoints['resnet_v1_50/block1'])
endpoints['fc_layer1_1'] = Dense(1024, activation='relu', kernel_initializer='Orthogonal')(endpoints['emb1'])
endpoints['bn_1_1'] = BN()(endpoints['fc_layer1_1'])
endpoints['feature1'] = Dense(embedding_dim, activation=None, kernel_initializer='Orthogonal')(
endpoints['bn_1_1'])
endpoints['emb2'] = GlobalAveragePooling2D(data_format='channels_last')(endpoints['resnet_v1_50/block2'])
endpoints['fc_layer2_1'] = Dense(1024, activation='relu', kernel_initializer='Orthogonal')(endpoints['emb2'])
endpoints['bn_2_1'] = BN()(endpoints['fc_layer2_1'])
endpoints['feature2'] = Dense(embedding_dim, activation=None, kernel_initializer='Orthogonal')(
endpoints['bn_2_1'])
endpoints['emb3'] = GlobalAveragePooling2D(data_format='channels_last')(endpoints['resnet_v1_50/block3'])
endpoints['fc_layer3_1'] = Dense(1024, activation='relu', kernel_initializer='Orthogonal')(endpoints['emb3'])
endpoints['bn_3_1'] = BN()(endpoints['fc_layer3_1'])
endpoints['feature3'] = Dense(embedding_dim, activation=None, kernel_initializer='Orthogonal')(
endpoints['bn_3_1'])
endpoints['emb4'] = GlobalAveragePooling2D(data_format='channels_last')(endpoints['resnet_v1_50/block4'])
endpoints['fc_layer4_1'] = Dense(1024, activation='relu', kernel_initializer='Orthogonal')(endpoints['emb4'])
endpoints['bn_4_1'] = BN()(endpoints['fc_layer4_1'])
endpoints['feature4'] = Dense(embedding_dim, activation=None, kernel_initializer='Orthogonal')(
endpoints['bn_4_1'])
endpoints['fusion_layer'] = FusionLayer_res()(
[endpoints['feature1'], endpoints['feature2'], endpoints['feature3'],
endpoints['feature4']])
endpoints['emb'] = endpoints['emb_raw'] = endpoints['fusion_layer']
return endpoints
else:
print('no such model, failure to build merged head layer')
exit(1)
x = keras.layers.maximum([c1, c2])
x = MaxPool2D(pool_size=(2, 2), strides=(2, 2), padding='same')(x)
x = Dropout(.2)(x)
c1 = Conv2D(filters=32,
kernel_size=1,
strides=(1, 1),
padding='same',
activation=None)(x)
c2 = Conv2D(filters=32,
kernel_size=1,
strides=(1, 1),
padding='same',
activation=None)(x)
x = keras.layers.maximum([c1, c2])
x = BN(axis=3, scale=False)(x)
x = Dropout(.2)(x)
c1 = Conv2D(filters=48,
kernel_size=3,
strides=(1, 1),
padding='same',
activation=None)(x)
c2 = Conv2D(filters=48,
kernel_size=3,
strides=(1, 1),
padding='same',
activation=None)(x)
x = keras.layers.maximum([c2, c1])
x = Dropout(.2)(x)
x = MaxPool2D(pool_size=(2, 2), strides=(2, 2), padding='same')(x)