from keras.initializers import RandomNormal, Ones
weight_init = RandomNormal(mean=0., stddev=0.02)
from keras.layers import Input, Activation, Concatenate, Flatten
from keras.layers import Conv2D, Conv2DTranspose, Dense, UpSampling2D
from keras.layers.advanced_activations import PReLU
from keras.layers.normalization import BatchNormalization
from keras.models import Model
from keras import backend as K
from keras.layers import Layer
class ScaleLayer(Layer):
def __init__(self, output_dim, **kwargs):
self.output_dim = output_dim
super(ScaleLayer, self).__init__(**kwargs)
def build(self, input_shape):
# Create a trainable weight variable # for this layer.
self.weights = self.add_weight(
name='weights',
shape=(input_shape[1], self.output_dim),
initializer=Ones(),
trainable=True)
super(ScaleLayer, self).build(input_shape)
# Be sure to call this at the end
def call(self, x):
return x * self.weights
def compute_output_shape(self, input_shape):
return (input_shape[0], self.output_dim)
def build_segan_generator(noisy_input_shape, z_input_shape,
n_filters=[64, 128, 256, 512, 1024],
kernel_size=(1, 31), use_upsampling=False):
noisy_input = Input(shape=noisy_input_shape)
z_input = Input(shape=z_input_shape)
# skip connections
skip_connections = []
# encode
x = noisy_input
for i in range(len(n_filters)):
x = Conv2D(filters=n_filters[i], kernel_size=kernel_size,
strides=(1, 4), padding='same', use_bias=True,
kernel_initializer=weight_init)(x)
x = PReLU()(x)
skip_connections.append(ScaleLayer(n_filters[i])(x))
# prepend single channel filter and remove the last filter size
n_filters = [1] + n_filters[:-1]
# update current x input
x = z_input
# decode
for i in range(len(n_filters)-1, -1, -1):
x = Concatenate(3)([x, skip_connections[i]])
if use_upsampling:
x = UpSampling2D(size=(1, 4))(x)
x = Conv2D(filters=n_filters[i], kernel_size=kernel_size,
strides=(1, 1), padding='same',
kernel_initializer=weight_init, use_bias=True)(x)
else:
x = Conv2DTranspose(filters=n_filters[i], kernel_size=kernel_size,
strides=(1, 4), padding='same',
kernel_initializer=weight_init)(x)
x = PReLU()(x) if i > 0 else Activation("tanh")(x)
# create model graph
model = Model(inputs=[noisy_input, z_input], outputs=x, name='Generator')
print("\nGenerator")
model.summary()
return model
"""
z_input_shape = (1, 8, 1024)
audio_input_shape = (1, 16384, 1)
g = build_segan_generator(z_input_shape, audio_input_shape)
"""
from keras.layers import Input, Reshape
from keras.layers.advanced_activations import PReLU
from keras.layers.convolutional import Conv2D
from keras.layers.core import Dense, Flatten
from keras.layers.normalization import BatchNormalization
from keras.models import Model
def build_segan_discriminator(noisy_input_shape, clean_input_shape,
n_filters=[64, 128, 256, 512, 1024],
kernel_size=(1, 31)):
clean_input = Input(shape=clean_input_shape)
noisy_input = Input(shape=noisy_input_shape)
x = Concatenate(-1)([clean_input, noisy_input])
# convolution layers
for i in range(len(n_filters)):
x = Conv2D(filters=n_filters[i], kernel_size=kernel_size,
strides=(1, 4), padding='same', use_bias=True,
kernel_initializer=weight_init)(x)
x = BatchNormalization(epsilon=1e-5, momentum=0.1)(x)
x = PReLU()(x)
x = Reshape((16384, ))(x)
# dense layers
x = Dense(256, activation=None, use_bias=True)(x)
x = PReLU()(x)
x = Dense(128, activation=None, use_bias=True)(x)
x = PReLU()(x)
x = Dense(1, activation=None, use_bias=True)(x)
# create model graph
model = Model(inputs=[noisy_input, clean_input], outputs=x, name='Discriminator')
print("\nDiscriminator")
model.summary()
return model
"""
clean_input_shape = (1, 16384, 1)
noisy_input_shape = (1, 16384, 1)
build_segan_discriminator(clean_input_shape, noisy_input_shape)
"""
