def conv_block8(feat_maps_out, prev):
prev = BatchNormalization(axis=CHANNEL_AXIS, name='BN8_1', freeze=False)(prev) # Specifying the axis and mode allows for later merging
prev = layers.LeakyReLU(name='Activation8_1')(prev)
prev = layers.Conv2D(feat_maps_out, (3,3), padding = 'same', kernel_initializer = 'he_normal', name='Conv8_1')(prev)
prev = BatchNormalization(axis=CHANNEL_AXIS, name='BN8_2', freeze=False)(prev) # Specifying the axis and mode allows for later merging
prev = layers.LeakyReLU(name='Activation8_2')(prev)
prev = layers.Conv2D(feat_maps_out, (3,3), padding = 'same', kernel_initializer = 'he_normal', name='Conv8_2')(prev)
return prev
def f(input):
y = Conv2D(filters, 3, strides=stride, padding="same", kernel_initializer="he_normal", kernel_regularizer=l2(1e-4))(input)
y = BatchNormalization(axis=3, freeze=False)(y)
y = LeakyReLU()(y)
y = Conv2D(filters, 3, strides=1, padding="same", kernel_initializer="he_normal", kernel_regularizer=l2(1e-4))(y)
y = BatchNormalization(axis=3, freeze=False)(y)
if block == 0:
shortcut = Conv2D(filters, 1, strides=stride, padding="same", kernel_initializer="he_normal", kernel_regularizer=l2(1e-4))(input)
shortcut = BatchNormalization(axis=3, freeze=False)(shortcut)
else:
shortcut = input
y = Add()([y, shortcut])
y = LeakyReLU()(y)
return y
def f(input):
x = Conv2D(48, 3, padding='same', use_bias=False, kernel_initializer='he_normal',
kernel_regularizer=l2(weight_decay))(input)
x = BatchNormalization(axis=3, freeze=False)(x)
x = LeakyReLU()(x)
return x
def initial_conv_block1(input, weight_decay=5e-4):
x = layers.Conv2D(32, (3, 3),
padding='same',
use_bias=False,
kernel_initializer='he_normal',
kernel_regularizer=l2(weight_decay),
name='conv1_1')(input)
x = BatchNormalization(axis=CHANNEL_AXIS, name='BN1_1', freeze=True)(x)
x = layers.LeakyReLU(name='Activation1_1')(x)
return x
def f(input):
b = Conv1D(filters=filters, kernel_size=1,
strides=1, padding="same",
#dilation_rate=dilationrate1D,
kernel_initializer="he_normal",
kernel_regularizer=l2(1e-4))(input)
b = BatchNormalization(axis=3, freeze=False)(b)
b = LeakyReLU()(b)
b_res = b
b = Conv1D(filters=filters, kernel_size=1,
strides=1, padding="same",
kernel_initializer="he_normal",
kernel_regularizer=l2(1e-4))(b)
b = BatchNormalization(axis=3, freeze=False)(b)
b_attention = Activation("softmax")(b)
b = Conv2D(filters=filters, kernel_size=3,
strides=(1,1), padding="same",
#dilation_rate=dilationrate2D,
kernel_initializer="he_normal",
kernel_regularizer=l2(1e-4))(b)
b = BatchNormalization(axis=3, freeze=False)(b)
b = Multiply()([b, b_attention])
b = LeakyReLU()(b)
b = Conv1D(filters=filters, kernel_size=1,
strides=1, padding="same",
kernel_initializer="he_normal",
kernel_regularizer=l2(1e-4))(b)
b = BatchNormalization(axis=3, freeze=False)(b)
b = Add()([b, b_res])
b = LeakyReLU()(b)
return b
def _bn_relu(input):
"""Helper to build a BN -> relu block
"""
norm = BatchNormalization(axis=CHANNEL_AXIS, freeze=False)(input)
return Activation("relu")(norm)
