def midium(inputs):
cnn = Conv1D(36, 3, padding="same")(inputs)
cnn = layers.PReLU()(cnn)
cnn = Conv1D(72, 3, padding="same")(cnn)
cnn = layers.PReLU()(cnn)
return cnn
def __init__(self, n_channels, name="decoder", **kwargs):
super(Decoder, self).__init__(name=name, **kwargs)
self.data_format = "channels_last"
num_filters = 256
self.sublayers = [
tfc.SignalConv2D(
num_filters,
(5, 5),
name="layer_out",
corr=False,
strides_up=1,
padding="same_zeros",
use_bias=True,
activation=tfc.GDN(name="igdn_out", inverse=True),
),
layers.PReLU(shared_axes=[1, 2]),
tfc.SignalConv2D(
num_filters,
(5, 5),
name="layer_0",
corr=False,
strides_up=1,
padding="same_zeros",
use_bias=True,
activation=tfc.GDN(name="igdn_0", inverse=True),
),
layers.PReLU(shared_axes=[1, 2]),
tfc.SignalConv2D(
num_filters,
(5, 5),
name="layer_1",
corr=False,
strides_up=1,
padding="same_zeros",
use_bias=True,
activation=tfc.GDN(name="igdn_1", inverse=True),
),
layers.PReLU(shared_axes=[1, 2]),
tfc.SignalConv2D(
num_filters,
(5, 5),
name="layer_2",
corr=False,
strides_up=2,
padding="same_zeros",
use_bias=True,
activation=tfc.GDN(name="igdn_2", inverse=True),
),
layers.PReLU(shared_axes=[1, 2]),
tfc.SignalConv2D(
n_channels,
(9, 9),
name="layer_3",
corr=False,
strides_up=2,
padding="same_zeros",
use_bias=True,
activation=tf.nn.sigmoid,
),
]
def __init__(self, conv_depth, name="encoder", **kwargs):
super(Encoder, self).__init__(name=name, **kwargs)
self.data_format = "channels_last"
num_filters = 256
self.sublayers = [
tfc.SignalConv2D(
num_filters,
(9, 9),
name="layer_0",
corr=True,
strides_down=2,
padding="same_zeros",
use_bias=True,
activation=tfc.GDN(name="gdn_0"),
),
layers.PReLU(shared_axes=[1, 2]),
tfc.SignalConv2D(
num_filters,
(5, 5),
name="layer_1",
corr=True,
strides_down=2,
padding="same_zeros",
use_bias=True,
activation=tfc.GDN(name="gdn_1"),
),
layers.PReLU(shared_axes=[1, 2]),
tfc.SignalConv2D(
num_filters,
(5, 5),
name="layer_2",
corr=True,
strides_down=1,
padding="same_zeros",
use_bias=True,
activation=tfc.GDN(name="gdn_2"),
),
layers.PReLU(shared_axes=[1, 2]),
tfc.SignalConv2D(
num_filters,
(5, 5),
name="layer_3",
corr=True,
strides_down=1,
padding="same_zeros",
use_bias=True,
activation=tfc.GDN(name="gdn_3"),
),
layers.PReLU(shared_axes=[1, 2]),
tfc.SignalConv2D(
conv_depth,
(5, 5),
name="layer_out",
corr=True,
strides_down=1,
padding="same_zeros",
use_bias=True,
activation=None,
),
]
def Decoder(x, feedback_bits):
B = 4
#decoder_input = DeuantizationLayer(B)(x)
decoder_input = tf.keras.layers.Reshape((-1, int(feedback_bits / B)))(x)
x = layers.Dense(1024, activation='linear')(decoder_input)
x = layers.Reshape((16, 32, 2))(x)
x = layers.Conv2D(conv_feature_size, 5, padding='SAME')(x)
x = layers.BatchNormalization()(x)
x_ini = layers.PReLU()(x)
x_tmp = layers.PReLU()(x)
for i in range(3):
x = layers.Conv2D(conv_feature_size, 5, padding='SAME')(x_ini)
x = layers.BatchNormalization()(x)
x = layers.PReLU()(x)
x = layers.Conv2D(conv_feature_size, 5, padding='SAME')(x)
x = layers.BatchNormalization()(x)
x = layers.PReLU()(x)
x_ini = keras.layers.Add()([x_ini, x])
x = keras.layers.Add()([x_ini, x_tmp])
#x = layers.UpSampling2D(size=(2,2))(x)
x = layers.Conv2D(conv_feature_size, 5, padding='SAME')(x)
x = layers.BatchNormalization()(x)
x = layers.PReLU()(x)
#x = layers.UpSampling2D(size=(4,4))(x)
decoder_output = layers.Conv2D(2, 7, padding='SAME',
activation='sigmoid')(x)
return decoder_output
def build_model():
model = Sequential([
layers.Dense(
50,
input_shape=(test_dataset.shape[-1], ),
),
layers.PReLU(alpha_initializer=tf.initializers.constant(0.25)),
layers.LayerNormalization(),
layers.Dropout(0.5),
layers.Dense(50),
layers.PReLU(alpha_initializer=tf.initializers.constant(0.25)),
layers.LayerNormalization(),
layers.Dropout(0.5),
layers.Dense(1, activation="sigmoid"),
])
metrics = [
keras.metrics.FalseNegatives(name="fn"),
keras.metrics.FalsePositives(name="fp"),
keras.metrics.TrueNegatives(name="tn"),
keras.metrics.TruePositives(name="tp"),
keras.metrics.Precision(name="precision"),
keras.metrics.Recall(name="recall"),
keras.metrics.AUC(name='auc'),
]
model.compile(optimizer=keras.optimizers.Adam(0.001),
loss="binary_crossentropy",
metrics=metrics)
return model
def Decoder(x, feedback_bits):
B = 4
#decoder_input = DeuantizationLayer(B)(x)
decoder_input = layers.Reshape((-1, int(feedback_bits / B)))(x)
x = layers.Reshape((8, 16, 2))(x)
x = layers.UpSampling2D(size=(2, 2),
data_format=None,
interpolation='nearest')(x)
tmp = layers.Conv2D(conv_feature_size, 7, padding='SAME')(x)
x = layers.PReLU()(tmp)
x = layers.UpSampling2D(size=(2, 2),
data_format=None,
interpolation='nearest')(x)
x = layers.Conv2D(conv_feature_size / 2, 3, padding='SAME')(x)
x = layers.BatchNormalization()(x)
x = layers.PReLU()(x)
x = layers.Conv2D(conv_feature_size / 2, 3, padding='SAME')(x)
x = layers.BatchNormalization()(x)
x = layers.PReLU()(x)
x = layers.Conv2D(conv_feature_size / 2, 3, padding='SAME')(x)
x = layers.BatchNormalization()(x)
x = layers.PReLU()(x)
x = layers.Conv2D(conv_feature_size / 4, 3, padding='SAME')(x)
x = layers.Add()([tmp, x])
decoder_output = layers.Conv2D(2, 7, padding='SAME',
activation='sigmoid')(x)
return decoder_output
def con2d_net_block(self, input_layer, block=1, repeat_times=1):
for i in range(repeat_times):
con2d_layer = layers.Conv2D(
int(math.pow(2, block - 1) * self.start_nodes), 1,
1)(input_layer)
bn_layer = layers.BatchNormalization()(con2d_layer)
act_layer = layers.PReLU()(bn_layer)
con2d_layer = layers.Conv2D(
int(math.pow(2, block - 1) * self.start_nodes), 3, 1,
'same')(act_layer)
bn_layer = layers.BatchNormalization()(con2d_layer)
act_layer = layers.PReLU()(bn_layer)
increase_channels_con2d = layers.Conv2D(
int(math.pow(2, block - 1) * self.start_nodes) * 4, 1,
1)(act_layer)
increase_channels_bn_layer = layers.BatchNormalization()(
increase_channels_con2d)
residual_con2d_layer = layers.Conv2D(
int(math.pow(2, block - 1) * self.start_nodes) * 4, 1,
1)(input_layer)
residual_bn_layer = layers.BatchNormalization()(
residual_con2d_layer)
add_layer = layers.add(
[increase_channels_bn_layer, residual_bn_layer])
#為方便做迴圈
input_layer = layers.PReLU()(add_layer)
return input_layer
def get_model():
inputs = keras.Input(shape=(50, 100, 1))
x = layers.Conv2D(16, 3)(inputs)
x = layers.Conv2D(16, 3, padding="same")(x)
x = layers.PReLU()(x)
x = layers.Conv2D(32, 3)(x)
x = layers.Conv2D(32, 3, padding="same")(x)
x = layers.PReLU()(x)
x = layers.MaxPooling2D(3)(x)
x = layers.Dropout(0.25)(x)
x = layers.Conv2D(64, 3)(x)
x = layers.Conv2D(64, 3, padding="same")(x)
x = layers.ReLU()(x)
x = layers.Conv2D(128, 3)(x)
x = layers.Conv2D(128, 3, padding="same")(x)
x = layers.ReLU()(x)
x = layers.MaxPooling2D(3)(x)
x = layers.Flatten()(x)
x = layers.Dropout(0.5)(x)
x = layers.Dense(CODE_LEN * DICT_LEN)(x)
x = layers.Reshape([CODE_LEN, DICT_LEN])(x)
outputs = layers.Softmax()(x)
model = keras.Model(inputs=inputs, outputs=outputs, name="captcha_model")
model.compile(
loss="categorical_crossentropy",
optimizer="Adam",
metrics=["accuracy"],
)
return model
def make_rnet(train=True):
input = layers.Input(shape=(24, 24, 3))
x = layers.Conv2D(28, kernel_size=(3, 3))(input)
x = layers.PReLU(shared_axes=[1, 2])(x)
x = layers.MaxPooling2D(pool_size=3, strides=2)(x)
x = layers.Conv2D(48, kernel_size=(3, 3))(x)
x = layers.PReLU(shared_axes=[1, 2])(x)
x = layers.MaxPooling2D(pool_size=3, strides=2)(x)
x = layers.Conv2D(64, kernel_size=(2, 2))(x)
x = layers.PReLU(shared_axes=[1, 2])(x)
x = layers.Flatten()(x)
x = layers.Dense(128)(x)
x = layers.PReLU()(x)
classifier = layers.Dense(3, activation='softmax', name='face_cls')(x)
bbox_regress = layers.Dense(5, name='bbox_reg')(x)
landmark_regress = layers.Dense(11, name='ldmk_reg')(x)
if train is False:
classifier = SampleLabelFilter(classifier)
classifier = layers.Activation('softmax')(classifier)
bbox_regress = SampleLabelFilter(bbox_regress)
landmark_regress = SampleLabelFilter(landmark_regress)
# outputs = layers.Concatenate()(
# [classifier, bbox_regress, landmark_regress])
outputs = [classifier, bbox_regress, landmark_regress]
model = models.Model(input, outputs)
return model
def reduce_image_size(self, input_layer, block=1):
#未來考慮加入名稱增強架構辨識度
block_name = "block_" + str(block)
reduce_size_con2d_layer = layers.Conv2D(
int(math.pow(2, block - 1) * self.start_nodes),
1,
2,
name=block_name + "_reduce_size_con2d_layer")(input_layer)
reduce_size_bn_layer = layers.BatchNormalization()(
reduce_size_con2d_layer)
reduce_size_act_layer = layers.PReLU()(reduce_size_bn_layer)
con2d_layer = layers.Conv2D(
int(math.pow(2, block - 1) * self.start_nodes), 3, 1,
"same")(reduce_size_act_layer)
bn_layer = layers.BatchNormalization()(con2d_layer)
act_layer = layers.PReLU()(bn_layer)
increase_channels_con2d = layers.Conv2D(
int(math.pow(2, block - 1) * self.start_nodes) * 4, 1,
1)(act_layer)
increase_channels_bn_layer = layers.BatchNormalization()(
increase_channels_con2d)
residual_con2d_layer = layers.Conv2D(
int(math.pow(2, block - 1) * self.start_nodes) * 4, 1,
2)(input_layer)
residual_bn_layer = layers.BatchNormalization()(residual_con2d_layer)
add_layer = layers.add([increase_channels_bn_layer, residual_bn_layer])
output_act_layer = layers.PReLU()(add_layer)
return output_act_layer
def classification_model(window_size, layer_name_list):
input_layer = Input(shape=(
window_size,
window_size,
1,
))
con2d_layer = layers.Conv2D(8, 2, name=layer_name_list[0])(input_layer)
act_layer = layers.PReLU(name=layer_name_list[1])(con2d_layer)
bn_layer = layers.BatchNormalization(
name=layer_name_list[2])(act_layer)
dropout_layer = layers.Dropout(0.5, name=layer_name_list[3])(bn_layer)
con2d_layer = layers.Conv2D(16, 2,
name=layer_name_list[4])(dropout_layer)
act_layer = layers.PReLU(name=layer_name_list[5])(con2d_layer)
bn_layer = layers.BatchNormalization(
name=layer_name_list[6])(act_layer)
dropout_layer = layers.Dropout(0.5, name=layer_name_list[7])(bn_layer)
con2d_layer = layers.Conv2D(32, 2,
name=layer_name_list[8])(dropout_layer)
act_layer = layers.PReLU(name=layer_name_list[9])(con2d_layer)
bn_layer = layers.BatchNormalization(
name=layer_name_list[10])(act_layer)
dropout_layer = layers.Dropout(0.5, name=layer_name_list[11])(bn_layer)
flatten_layer = layers.Flatten()(dropout_layer)
hidden_layer = layers.Dense(20)(flatten_layer)
act_layer = layers.PReLU()(hidden_layer)
dropout_layer = layers.Dropout(0.5)(act_layer)
hidden_layer = layers.Dense(40)(dropout_layer)
act_layer = layers.PReLU()(hidden_layer)
dropout_layer = layers.Dropout(0.5)(act_layer)
output_layer = layers.Dense(10, activation='softmax')(dropout_layer)
return Model(input_layer, output_layer)
def __init__(self, embedding_count_dict, embedding_dim_dict, embedding_features_list, user_behavior_features,
activation='PReLU'):
super(DIN, self).__init__()
# Init Embedding Layer
self.embedding_count_dict = embedding_count_dict
self.embedding_dim_dict = embedding_dim_dict
self.embedding_layers = dict()
for feature in embedding_features_list:
self.embedding_layers[feature] = layers.Embedding(input_dim=embedding_count_dict[feature],
output_dim=embedding_dim_dict[feature],
embeddings_initializer='random_uniform')
# DIN Attention + Sum Pooling
self.hist_at = attention(alibaba_utils.get_input_dim(embedding_dim_dict, user_behavior_features))
# Init Fully Connection Layer
self.fc = tf.keras.Sequential()
self.fc.add(layers.BatchNormalization())
self.fc.add(layers.Dense(200, activation="relu"))
if activation == "Dice":
self.fc.add(Dice())
elif activation == "dice":
self.fc.add(dice(200))
elif activation == "PReLU":
self.fc.add(layers.PReLU(alpha_initializer='zeros', weights=None))
self.fc.add(layers.Dense(80, activation="relu"))
if activation == "Dice":
self.fc.add(Dice())
elif activation == "dice":
self.fc.add(dice(80))
elif activation == "PReLU":
self.fc.add(layers.PReLU(alpha_initializer='zeros', weights=None))
self.fc.add(layers.Dense(2, activation=None))
def define_architecture(self,
model_input,
final_activation="softmax",
out_name="m"):
"""Build the V-Net model."""
# Downsampling / encoding portion
conv0 = self.side_conv_block(model_input, 16, length=1)
down0 = layers.Conv3D(32, (2, 2, 2), strides=(2, 2, 1),
padding="same")(conv0)
down0 = layers.PReLU()(down0)
conv1 = self.side_conv_block(down0, 32, length=2)
down1 = layers.Conv3D(64, (2, 2, 2), strides=(2, 2, 1),
padding="same")(conv1)
down1 = layers.PReLU()(down1)
conv2 = self.side_conv_block(down1, 64, length=3)
down2 = layers.Conv3D(128, (2, 2, 2),
strides=(2, 2, 1),
padding="same")(conv2)
down2 = layers.PReLU()(down2)
conv3 = self.side_conv_block(down2, 128, length=3)
down3 = layers.Conv3D(256, (2, 2, 2),
strides=(2, 2, 1),
padding="same")(conv3)
down3 = layers.PReLU()(down3)
# Middle of network
conv4 = self.side_conv_block(down3, 256, length=3)
# Upsampling / decoding portion
uconv3 = self.up_conv_block(conv4,
conv3,
256,
length=3,
strides=(2, 2, 1))
uconv2 = self.up_conv_block(uconv3,
conv2,
128,
length=3,
strides=(2, 2, 1))
uconv1 = self.up_conv_block(uconv2,
conv1,
64,
length=2,
strides=(2, 2, 1))
uconv0 = self.up_conv_block(uconv1,
conv0,
32,
length=1,
strides=(2, 2, 1))
out = layers.Conv3D(self.output_length, (1, 1, 1),
padding="same",
activation=None)(uconv0)
out = layers.Activation(final_activation, name=out_name)(out)
return out
def cnn_model(input_shape):
img_input = layers.Input(input_shape)
## First CNN block:
conv2D_1 = layers.Conv2D(64, (3, 3),
strides=2,
padding='valid',
name='Conv_1',
input_shape=input_shape,
kernel_initializer='he_normal')(img_input)
conv2D_1 = layers.BatchNormalization(axis=3)(conv2D_1)
conv2D_1 = layers.PReLU()(conv2D_1)
conv2D_2 = layers.Conv2D(64, (3, 3),
strides=1,
padding='same',
name='Conv_2',
kernel_initializer='he_normal')(conv2D_1)
conv2D_2 = layers.BatchNormalization(axis=3)(conv2D_2)
conv2D_2 = layers.PReLU()(conv2D_2)
maxPooling_2 = layers.MaxPooling2D((3, 3),
strides=(2, 2),
name='MaxPooling_1')(conv2D_2)
## Second CNN block:
conv2D_3 = layers.Conv2D(128, (3, 3),
strides=1,
padding='same',
name='Conv_3',
kernel_initializer='he_normal')(maxPooling_2)
conv2D_4 = layers.BatchNormalization(axis=3)(conv2D_3)
conv2D_3 = layers.PReLU()(conv2D_3)
conv2D_4 = layers.Conv2D(128, (3, 3),
strides=1,
padding='same',
name='Conv_4',
kernel_initializer='he_normal')(conv2D_3)
conv2D_4 = layers.BatchNormalization(axis=3)(conv2D_4)
conv2D_4 = layers.PReLU()(conv2D_4)
averagePooling_4 = layers.AveragePooling2D(
(3, 3), strides=(2, 2), name='AveragePooling_1')(conv2D_4)
flatten = layers.Flatten()(averagePooling_4)
fc = layers.Dropout(0.5)(flatten)
fc = layers.Dense(256,
kernel_regularizer=regularizers.l2(0.01),
kernel_initializer='he_normal',
name='FC1')(fc)
fc = layers.Activation('relu')(fc)
fc = layers.Dropout(0.3)(fc)
fc = layers.Dense(128,
kernel_regularizer=regularizers.l2(0.01),
kernel_initializer='he_normal',
name='FC2')(fc)
fc = layers.Activation('relu')(fc)
fc = layers.Dropout(0.3)(fc)
output = layers.Dense(1)(fc)
cnn_model = models.Model(inputs=[img_input], outputs=[output])
return cnn_model
def fine(inputs):
cnn = Conv1D(36, 4, padding="same")(inputs)
cnn = layers.PReLU()(cnn)
cnn = Conv1D(36, 4, padding="same")(cnn)
cnn = layers.PReLU()(cnn)
cnn = Conv1D(72, 4, padding="same")(cnn)
cnn = layers.PReLU()(cnn)
return cnn
def __init__(self,
embedding_count_dict,
embedding_dim_dict,
embedding_features_list,
user_behavior_features,
activation="PReLU"):
super(DIEN, self).__init__(embedding_count_dict, embedding_dim_dict,
embedding_features_list, activation)
"""DIEN初始化model函数
该函数在调用DIEN时进行DIEN的Embedding层,GRU层,AUGRU层,全连接层的初始化操作
Args:
embedding_count_dict:string->int格式,该变量记录需要embedding各个特征的词典个数,即最大整数索引+ 1的大小;
embedding_dim_dict:string->int格式,该变量记录需要embedding各个特征的输出维数,即密集嵌入的尺寸;
embedding_features_list:list(string)格式,该变量记录DIEN中user_profile部分所有需要embedding的feature名称;
user_behavior_features:list(string)格式,该变量记录DIEN中user_behavior与target_item部分所有需要embedding的feature名称
activation:string格式,默认值"PReLU",该变量空值全连接层激活函数,”PReLU“->PReLU,"Dice"->Dice
"""
#Init Embedding Layer
self.embedding_dim_dict = embedding_dim_dict
self.embedding_count_dict = embedding_count_dict
self.embedding_layers = dict()
for feature in embedding_features_list:
self.embedding_layers[feature] = layers.Embedding(
embedding_count_dict[feature], embedding_dim_dict[feature])
#Init GRU Layer
self.user_behavior_gru = layers.GRU(self.get_GRU_input_dim(
embedding_dim_dict, user_behavior_features),
return_sequences=True)
#Init Attention Layer
self.attention_layer = layers.Softmax()
#Init Auxiliary Layer
self.AuxNet = AuxLayer()
#Init AUGRU Layer
self.user_behavior_augru = AUGRU(
self.get_GRU_input_dim(embedding_dim_dict, user_behavior_features))
#Init Fully Connection Layer
self.fc = tf.keras.Sequential()
self.fc.add(layers.BatchNormalization())
self.fc.add(layers.Dense(200, activation="relu"))
if activation == "Dice":
self.fc.add(Dice())
elif activation == "dice":
self.fc.add(dice(200))
elif activation == "PReLU":
self.fc.add(layers.PReLU(alpha_initializer='zeros', weights=None))
self.fc.add(layers.Dense(80, activation="relu"))
if activation == "Dice":
self.fc.add(Dice())
elif activation == "dice":
self.fc.add(dice(80))
elif activation == "PReLU":
self.fc.add(layers.PReLU(alpha_initializer='zeros', weights=None))
self.fc.add(layers.Dense(2, activation=None))
def regression_subnet(shape):
input_img = layers.Input(shape)
# downsample a bit
net = layers.Conv2D(filters=16,
kernel_size=3,
strides=2,
padding='same')(input_img)
net = layers.PReLU(alpha_initializer='zeros', shared_axes=[1,
2])(net)
# (48 48 16)
net = layers.Conv2D(filters=32,
kernel_size=3,
strides=2,
padding='same')(net)
net = layers.PReLU(alpha_initializer='zeros', shared_axes=[1,
2])(net)
# (24 24 32)
net = layers.Conv2D(filters=64,
kernel_size=3,
strides=2,
padding='same')(net)
net = layers.PReLU(alpha_initializer='zeros', shared_axes=[1,
2])(net)
# (12 12 64)
net = layers.Conv2D(filters=128,
kernel_size=3,
strides=2,
padding='same')(net)
net = layers.PReLU(alpha_initializer='zeros', shared_axes=[1,
2])(net)
# (6 6 128)
net = layers.Conv2D(filters=256,
kernel_size=3,
strides=2,
padding='same')(net)
net = layers.PReLU(alpha_initializer='zeros', shared_axes=[1,
2])(net)
# (3 3 256)
net = layers.Conv2D(filters=self.config.num_landmarks * 2,
kernel_size=1,
strides=1,
padding='valid')(net)
out = tf.reshape(net, [-1, self.config.num_landmarks, 2],
name='landmarks')
model = tf.keras.Model(inputs=input_img,
outputs=out,
name='regression_subnet')
return model
def __init__(self, n_channels, name="decoder", **kwargs):
super(Decoder, self).__init__(name=name, **kwargs)
self.data_format = "channels_last"
self.sublayers = [
tfc.SignalConv2D(
32,
(5, 5),
name="conv_1",
corr=False,
strides_up=1,
padding="same_zeros",
use_bias=True,
),
layers.PReLU(shared_axes=[1, 2]),
tfc.SignalConv2D(
32,
(5, 5),
name="conv_2",
corr=False,
strides_up=1,
padding="same_zeros",
use_bias=True,
),
layers.PReLU(shared_axes=[1, 2]),
tfc.SignalConv2D(
32,
(5, 5),
name="conv_3",
corr=False,
strides_up=1,
padding="same_zeros",
use_bias=True,
),
layers.PReLU(shared_axes=[1, 2]),
tfc.SignalConv2D(
16,
(5, 5),
name="conv_4",
corr=False,
strides_up=2,
padding="same_zeros",
use_bias=True,
),
layers.PReLU(shared_axes=[1, 2]),
tfc.SignalConv2D(
n_channels,
(5, 5),
name="conv_5",
corr=False,
strides_up=2,
padding="same_zeros",
use_bias=True,
activation=tf.nn.sigmoid,
),
]
def __init__(self, conv_depth, name="encoder", **kwargs):
super(Encoder, self).__init__(name=name, **kwargs)
self.data_format = "channels_last"
self.sublayers = [
tfc.SignalConv2D(
16,
(5, 5),
name="conv_1",
corr=True,
strides_down=2,
padding="same_zeros",
use_bias=True,
),
layers.PReLU(shared_axes=[1, 2]),
tfc.SignalConv2D(
32,
(5, 5),
name="conv_2",
corr=True,
strides_down=2,
padding="same_zeros",
use_bias=True,
),
layers.PReLU(shared_axes=[1, 2]),
tfc.SignalConv2D(
32,
(5, 5),
name="conv_3",
corr=True,
strides_down=1,
padding="same_zeros",
use_bias=True,
),
layers.PReLU(shared_axes=[1, 2]),
tfc.SignalConv2D(
32,
(5, 5),
name="conv_4",
corr=True,
strides_down=1,
padding="same_zeros",
use_bias=True,
),
layers.PReLU(shared_axes=[1, 2]),
tfc.SignalConv2D(
conv_depth,
(5, 5),
name="conv_5",
corr=True,
strides_down=1,
padding="same_zeros",
use_bias=True,
activation=None,
),
]
def call(self, inputs, training):
x, x_highway = inputs
merged = tfkl.concatenate([x, x_highway])
in1 = merged
for _ in range(self.num_conv):
in_1 = tfkl.PReLU()(self.conv(in_1))
add_1 = tfkl.add([in1, merged])
output = self.up_conv(add_1)
output = tfkl.PReLU()(output)
return output
def __bottleneck_block(input,
filters=96,
cardinality=16,
strides=1,
weight_decay=0):
''' Adds a bottleneck block
Args:
input: input tensor
filters: number of output filters
cardinality: cardinality factor described number of
grouped convolutions
strides: performs strided convolution for downsampling if > 1
weight_decay: weight decay factor
Returns: a keras tensor
'''
init = input
grouped_channels = int(filters / cardinality)
channel_axis = -1
if init.shape[-1] != 2 * filters:
init = Conv2D(filters * 2, (1, 1),
padding='same',
strides=(strides, strides),
use_bias=False,
kernel_initializer='he_normal',
kernel_regularizer=l2(weight_decay))(init)
init = BatchNormalization(axis=channel_axis)(init)
x = Conv2D(filters, (1, 1),
padding='same',
use_bias=False,
kernel_initializer='he_normal',
kernel_regularizer=l2(weight_decay))(input)
x = BatchNormalization(axis=channel_axis)(x)
x = layers.PReLU()(x)
# x = layers.Activation(tf.nn.swish)(x)
x = __grouped_convolution_block(x, grouped_channels, cardinality, strides,
weight_decay)
x = Conv2D(filters * 2, (1, 1),
padding='same',
use_bias=False,
kernel_initializer='he_normal',
kernel_regularizer=l2(weight_decay))(x)
x = BatchNormalization(axis=channel_axis)(x)
x = SqueezeExciteLayer(16)(x)
x = add([init, x])
x = layers.PReLU()(x)
return x
def creat_clustering_model_pre(number_of_class, maxlen, input_dim, output_dim, layers_name):
input_layer = Input(shape = (maxlen,), name = layers_name[0])
#input_length 是每筆輸入的長度
#如果要接flatten的話是必須參數
embedding_layer = layers.Embedding(input_dim, output_dim, input_length = maxlen, name = layers_name[1])(input_layer)
flatten_layer = layers.Flatten(name = layers_name[2])(embedding_layer)
hidden_layer = layers.Dense(64, name = layers_name[3])(flatten_layer)
activation_layer = layers.PReLU(name = layers_name[4])(hidden_layer)
hidden_layer = layers.Dense(128, name = layers_name[5])(activation_layer)
activation_layer = layers.PReLU(name = layers_name[6])(hidden_layer)
hidden_layer = layers.Dense(256, name = layers_name[7])(activation_layer)
activation_layer = layers.PReLU(name = layers_name[8])(hidden_layer)
hidden_layer = layers.Dense(512, name = layers_name[9])(activation_layer)
activation_layer = layers.PReLU(name = layers_name[10])(hidden_layer)
label_layer = layers.Dense(number_of_class, activation = 'softmax', name = layers_name[11])(activation_layer)
hidden_layer = layers.Dense(512)(label_layer)
activation_layer = layers.PReLU()(hidden_layer)
hidden_layer = layers.Dense(256)(activation_layer)
activation_layer = layers.PReLU()(hidden_layer)
hidden_layer = layers.Dense(128)(activation_layer)
activation_layer = layers.PReLU()(hidden_layer)
hidden_layer = layers.Dense(64)(activation_layer)
activation_layer = layers.PReLU()(hidden_layer)
output_layer = layers.Dense(maxlen * output_dim)(activation_layer)
#output_layer = layers.PReLU()(hidden_layer)
return Model(input_layer, [output_layer, flatten_layer, label_layer])
def __grouped_convolution_block(input,
grouped_channels,
cardinality,
strides,
weight_decay=0):
''' Adds a grouped convolution block. It is an equivalent block from the paper
Args:
input: input tensor
grouped_channels: grouped number of filters
cardinality: cardinality factor describing the number of groups
strides: performs strided convolution for downscaling if > 1
weight_decay: weight decay term
Returns: a keras tensor
'''
init = input
channel_axis = -1
group_list = []
if cardinality == 1:
# with cardinality 1, it is a standard convolution
x = Conv2D(grouped_channels, (3, 3),
padding='same',
use_bias=False,
strides=(strides, strides),
kernel_initializer='he_normal',
kernel_regularizer=l2(weight_decay))(init)
x = BatchNormalization(axis=channel_axis)(x)
x = layers.PReLU()(x)
return x
for c in range(cardinality):
x = Lambda(lambda z: z[:, :, :, c * grouped_channels:
(c + 1) * grouped_channels])(input)
x = Conv2D(grouped_channels, (7, 7),
padding='same',
use_bias=False,
strides=(strides, strides),
kernel_initializer='he_normal',
kernel_regularizer=l2(weight_decay))(x)
group_list.append(x)
group_merge = concatenate(group_list, axis=channel_axis)
x = BatchNormalization(axis=channel_axis)(group_merge)
x = layers.PReLU()(x)
# x = layers.Activation(tf.nn.swish)(x)
return x
def build(num_classes: int):
'''
Build the Seq12 architecture given width, height and depth
as dimensions of the input tensor and the corresponding
number of classes of the data.
parameters
----------
num_classes: output size
returns
-------
model: the Seq12 model compatible with given inputs
as a keras sequential model.
'''
# initialize model
print("[INFO] preparing model...")
# augment images
data_augmentation = tf.keras.Sequential([
layers.experimental.preprocessing.RandomFlip(
"horizontal",
input_shape=(config.img_height, config.img_width,
config.depth)),
layers.experimental.preprocessing.RandomRotation(0.1),
layers.experimental.preprocessing.RandomZoom(0.1),
layers.experimental.preprocessing.Rescaling(1. / 255)
])
model = Sequential(name='SeqS')
model.add(data_augmentation)
model.add(layers.Conv2D(16, 3, padding='same', activation='relu'))
model.add(layers.BatchNormalization())
model.add(layers.PReLU())
model.add(layers.MaxPooling2D())
model.add(layers.Flatten())
model.add(layers.Dense(128, activation='relu'))
model.add(layers.BatchNormalization())
model.add(layers.PReLU())
model.add(layers.Dropout(0.5))
model.add(layers.Dense(num_classes))
model.add(layers.Activation('softmax'))
#model.summary()
# return the constructed network architecture
return model
def ACR_ENC2(x):
res = x
input_ch = x.shape[-1]
x = layers.Conv2D(input_ch,
kernel_size=(1, 3),
strides=1,
padding='same',
data_format='channels_last',
use_bias=False)(x)
x = add_common_layers(x)
x = layers.Conv2D(input_ch,
kernel_size=(3, 1),
strides=1,
padding='same',
data_format='channels_last',
use_bias=False)(x)
x = layers.BatchNormalization()(x)
x = SqueezeExciteLayer(16)(x)
x = layers.PReLU()(x + res)
return x
def creat_model(self):
#測試使用類似ResNet架構
input_layer = Input(shape = [400, 400, self.channels, ], name = "Input_layer")
#第一區多用較大的stride降維
block_0_con2d_layer = layers.Conv2D(self.start_nodes, 3, 2, name = "block_0_con2d_layer")(input_layer)
block_0_bn_layer = layers.BatchNormalization(name = "block_0_bn_layer")(block_0_con2d_layer)
block_0_act_layer = layers.PReLU(name = "block_0_act_layer")(block_0_bn_layer)
block_0_maxpool_layer = layers.MaxPool2D(name = "block_0_maxpool_layer")(block_0_act_layer)
#第二區由數個block組成
#通常會在每個block開頭先降解析度並增加維度
block_1_reduce_image_size = self.reduce_image_size(block_0_maxpool_layer, 1)
block_1_main_con2d_net = self.con2d_net_block(block_1_reduce_image_size, 1, 2)
block_2_reduce_image_size = self.reduce_image_size(block_1_main_con2d_net, 2)
block_2_main_con2d_net = self.con2d_net_block(block_2_reduce_image_size, 2, 2)
block_3_reduce_image_size = self.reduce_image_size(block_2_main_con2d_net, 3)
block_3_main_con2d_net = self.con2d_net_block(block_3_reduce_image_size, 3, 3)
block_4_reduce_image_size = self.reduce_image_size(block_3_main_con2d_net, 4)
block_4_main_con2d_net = self.con2d_net_block(block_4_reduce_image_size, 4, 5)
block_5_reduce_image_size = self.reduce_image_size(block_4_main_con2d_net, 5)
block_5_main_con2d_net = self.con2d_net_block(block_5_reduce_image_size, 5, 2)
#最後以一個avg_pool降維並連接到輸出
avg_pool_layer = layers.AvgPool2D(4, 1)(block_5_main_con2d_net)
flatten_layer = layers.Flatten()(avg_pool_layer)
output_layer = layers.Dense(self.number_of_class, activation = 'softmax')(flatten_layer)
self.model = Model(input_layer, output_layer)
def gen_unary_operator_test(name, type, input_shape):
# Create model.
inp = layers.Input(name='input',
batch_size=input_shape[0],
shape=input_shape[1:])
if type == 'relu':
out = tf.nn.relu(inp)
elif type == 'relu_n1to1':
out = tf.clip_by_value(inp, -1.0, 1.0)
elif type == 'relu6':
out = tf.nn.relu6(inp)
elif type == 'sigmoid':
out = tf.nn.sigmoid(inp)
elif type == 'exp':
out = tf.exp(inp)
elif type == 'log':
out = tf.math.log(inp)
elif type == 'tanh':
out = tf.nn.tanh(inp)
elif type == 'leaky_relu':
out = tf.nn.leaky_relu(inp, alpha=0.1)
elif type == 'prelu':
out = layers.PReLU(alpha_initializer='random_uniform')(inp)
elif type == 'square':
out = tf.math.square(inp)
elif type == 'abs':
out = tf.math.abs(inp)
elif type == 'neg':
out = tf.math.negative(inp)
elif type == 'sqrt':
out = tf.math.sqrt(inp)
elif type == 'rsqrt':
out = tf.math.rsqrt(inp)
elif type == 'sin':
out = tf.math.sin(inp)
elif type == 'cos':
out = tf.math.cos(inp)
elif type == 'ceil':
out = tf.math.ceil(inp)
elif type == 'round':
out = tf.math.round(inp)
elif type == 'floor':
out = tf.math.floor(inp)
else:
print('Unary operator "%s" not supported!')
exit(1)
model = Model(inputs=[inp], outputs=[out])
# Create data.
np.random.seed(0)
inp_tensor = np.random.randn(*input_shape).astype(np.float32)
if type in ['log', 'sqrt', 'rsqrt']:
inp_tensor = np.abs(inp_tensor) + 1
out_tensor = model.predict(inp_tensor)
# Save model.
save_model(model, name)
# Save data.
save_tensor(inp_tensor, name + '.inp0')
save_tensor(out_tensor, name + '.out0')
# Clear session.
keras_backend.clear_session()
def T_regressor(inputs, num_landmarks):
dropout_rate = 0.5
net = layers.Conv2D(filters=32, kernel_size=3, strides=2, padding='same')(inputs) # (48, 48, 32)
net = layers.PReLU(alpha_initializer='zeros', shared_axes=[1, 2])(net)
net = layers.Dropout(rate=dropout_rate)(net)
net = conv_block(net, conv_filters=32)
net = branch_block(net, depth_strides=2, conv_filters=64) # (24, 24, 32)
net = layers.Dropout(rate=dropout_rate)(net)
net = conv_block(net, conv_filters=64)
net = branch_block(net, depth_strides=2, conv_filters=64, pad=False) # (12, 12, 64)
net = layers.Dropout(rate=dropout_rate)(net)
net = conv_block(net, conv_filters=64)
net = branch_block(net, depth_strides=2, conv_filters=64, pad=False) # (6, 6, 64)
net = layers.Dropout(rate=dropout_rate)(net)
net = conv_block(net, conv_filters=64)
net = branch_block(net, depth_strides=2, conv_filters=128) # (3, 3, 64)
net = layers.Dropout(rate=dropout_rate)(net)
net = conv_block(net, conv_filters=128)
net = conv_block(net, conv_filters=128)
net = conv_block(net, conv_filters=128)
net = layers.DepthwiseConv2D(kernel_size=3, strides=1, padding='valid',
kernel_regularizer=tf.keras.regularizers.l2(0.01))(net)
net = layers.Conv2D(filters=num_landmarks * 2, kernel_size=1, padding='same',
kernel_regularizer=tf.keras.regularizers.l2(0.01))(net)
# (1, 1, num_landmarks)
# reshape result
return tf.reshape(net, [-1, num_landmarks, 2], name='landmark')
def conv_bn_relu(image_in, mask_in, filters, kernel_size,
downsampling=1, upsampling=1, act="relu",
concat_img=None, concat_mask=None, reps=1):
assert not (concat_img is None)^(concat_mask is None) # XORは常にFalse
# Upsamplingする場合
if upsampling > 1:
conv = layers.Lambda(upsampling2d_tpu, arguments={"scale":upsampling})(image_in)
mask = layers.Lambda(upsampling2d_tpu, arguments={"scale":upsampling})(mask_in)
else:
conv, mask = image_in, mask_in
if concat_img is not None and concat_mask is not None:
conv = layers.Concatenate()([conv, concat_img])
mask = layers.Concatenate()([mask, concat_mask])
# 計算量削減のために1x1 Convを入れる
conv, mask = PConv2D(filters=filters, kernel_size=1)([conv, mask])
conv = layers.BatchNormalization()(conv)
conv = layers.Activation("relu")(conv)
for i in range(reps):
stride = downsampling if i == 0 else 1
# strideでダウンサンプリング
conv, mask = PConv2D(filters=filters, kernel_size=kernel_size,
padding="same", strides=stride)([conv, mask])
# Image側だけBN->ReLUを入れる
conv = layers.BatchNormalization()(conv)
if act == "relu":
conv = layers.Activation("relu")(conv)
elif act == "prelu":
conv = layers.PReLU()(conv)
elif act == "custom_tanh":
# 元の画像の白を黒に変えるには、tanhの2倍のスケール[-2,2]が必要
conv = layers.Lambda(lambda x: 2*K.tanh(x), name="unmasked")(conv)
return conv, mask
def initial_block(input_tensor):
"""ENet initial block
:param input_tensor: input tensor
:return: initial block tensor
"""
if backend.image_data_format() == 'channels_last':
channel_axis = 3
else:
channel_axis = 1
branch_conv = layers.Conv2D(
filters=13,
kernel_size=(3, 3),
strides=(2, 2),
padding='same',
use_bias=False,
kernel_initializer='he_normal',
kernel_regularizer=regularizers.l2(L2_WEIGHT_DECAY),
name='initial_block_conv')(input_tensor)
branch_conv = layers.BatchNormalization(
axis=channel_axis, name='initial_block_bn')(branch_conv)
branch_pool = layers.MaxPool2D(pool_size=(2, 2),
strides=(2, 2),
padding='valid')(input_tensor)
x = layers.Concatenate(axis=channel_axis)([branch_conv, branch_pool])
x = layers.PReLU(alpha_initializer=PRELU_ALPHA)(x)
return x
