def __init__(self,
input_dim,
kernel_size=32,
dropout=0.0,
name="conv_module",
**kwargs):
super(ConvModule, self).__init__(name=name, **kwargs)
self.ln = tf.keras.layers.LayerNormalization()
self.pw_conv_1 = tf.keras.layers.Conv1D(filters=2 * input_dim,
kernel_size=1,
strides=1,
padding="same",
name="pw_conv_1")
self.glu = GLU()
self.dw_conv = tf.keras.layers.SeparableConv1D(filters=2 * input_dim,
kernel_size=kernel_size,
strides=1,
padding="same",
depth_multiplier=1,
name="dw_conv")
self.bn = SwitchNormalization()
self.swish = tf.keras.layers.Activation(tf.keras.activations.swish,
name="swish_activation")
self.pw_conv_2 = tf.keras.layers.Conv1D(filters=input_dim,
kernel_size=1,
strides=1,
padding="same",
name="pw_conv_2")
self.do = tf.keras.layers.Dropout(dropout)
self.res_add = tf.keras.layers.Add()
class ConvModule(tf.keras.layers.Layer):
def __init__(self,
input_dim,
kernel_size=32,
dropout=0.0,
name="conv_module",
**kwargs):
super(ConvModule, self).__init__(name=name, **kwargs)
self.ln = tf.keras.layers.LayerNormalization()
self.pw_conv_1 = tf.keras.layers.Conv1D(filters=2 * input_dim,
kernel_size=1,
strides=1,
padding="same",
name="pw_conv_1")
self.glu = GLU()
self.dw_conv = tf.keras.layers.SeparableConv1D(filters=2 * input_dim,
kernel_size=kernel_size,
strides=1,
padding="same",
depth_multiplier=1,
name="dw_conv")
self.bn = SwitchNormalization()
self.swish = tf.keras.layers.Activation(tf.keras.activations.swish,
name="swish_activation")
self.pw_conv_2 = tf.keras.layers.Conv1D(filters=input_dim,
kernel_size=1,
strides=1,
padding="same",
name="pw_conv_2")
self.do = tf.keras.layers.Dropout(dropout)
self.res_add = tf.keras.layers.Add()
# @tf.function(experimental_relax_shapes=True)
def call(self, inputs, training=False, **kwargs):
outputs = self.ln(inputs, training=training)
outputs = self.pw_conv_1(outputs, training=training)
outputs = self.glu(outputs)
outputs = self.dw_conv(outputs, training=training)
outputs = self.bn(outputs, training=training)
outputs = self.swish(outputs)
outputs = self.pw_conv_2(outputs, training=training)
outputs = self.do(outputs, training=training)
outputs = self.res_add([inputs, outputs])
return outputs
def get_config(self):
conf = super(ConvModule, self).get_config()
conf.update(self.ln.get_config())
conf.update(self.pw_conv_1.get_config())
conf.update(self.glu.get_config())
conf.update(self.dw_conv.get_config())
conf.update(self.bn.get_config())
conf.update(self.swish.get_config())
conf.update(self.pw_conv_2.get_config())
conf.update(self.do.get_config())
conf.update(self.res_add.get_config())
return conf
def __init__(self,filter_size,kernel_size):
super(ESP_alhpa,self).__init__()
filter_size//=4
self.chanle_projecter=tf.keras.layers.Conv2D(filter_size,kernel_size,padding='same')
self.dilated_conv1 = tf.keras.layers.Conv2D(filter_size,kernel_size,padding='same',
kernel_initializer=tf.random_uniform_initializer(seed=42),dilation_rate=(1, 1))
self.dilated_conv2 = tf.keras.layers.Conv2D(filter_size,kernel_size,padding='same',
kernel_initializer=tf.random_uniform_initializer(seed=42), dilation_rate=(2, 2))
self.dilated_conv4 = tf.keras.layers.Conv2D(filter_size,kernel_size,padding='same',
kernel_initializer=tf.random_uniform_initializer(seed=42), dilation_rate=(4, 4))
self.dilated_conv8 = tf.keras.layers.Conv2D(filter_size,kernel_size,padding='same',
kernel_initializer=tf.random_uniform_initializer(seed=42), dilation_rate=(8, 8))
self.dilated_conv16 = tf.keras.layers.Conv2D(filter_size,kernel_size,padding='same',
kernel_initializer=tf.random_uniform_initializer(seed=42), dilation_rate=(16, 16))
self.sw=SwitchNormalization()
def __init__(self):
super(BN_PRelu,self).__init__()
self.sw=SwitchNormalization()