def __init__(self, input_tensor, encoder, data_type, is_training, reuse):
self.input_tensor = input_tensor
self.input_depth = self.input_tensor.shape[3]
with tf.variable_scope('decoder'):
with tf.variable_scope('layer-1'):
net = mf.relu(self.input_tensor)
net = mf.deconv(net,
filters=128,
kernel_size=(5, 5, self.input_depth),
stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
net = mf.dropout(net, .5)
with tf.variable_scope('layer-2'):
net = mf.relu(mf.concat(net, encoder.l5))
net = mf.deconv(net,
filters=64,
kernel_size=(5, 5, self.input_depth),
stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
net = mf.dropout(net, .5)
with tf.variable_scope('layer-3'):
net = mf.relu(mf.concat(net, encoder.l4))
net = mf.deconv(net,
filters=32,
kernel_size=(5, 5, self.input_depth),
stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
net = mf.dropout(net, .5)
with tf.variable_scope('layer-4'):
net = mf.relu(mf.concat(net, encoder.l3))
net = mf.deconv(net,
filters=16,
kernel_size=(5, 5, self.input_depth),
stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
with tf.variable_scope('layer-5'):
net = mf.relu(mf.concat(net, encoder.l2))
net = mf.deconv(net,
filters=8,
kernel_size=(5, 5, self.input_depth),
stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
with tf.variable_scope('layer-6'):
if data_type == 'mag_phase_real_imag':
self.out_depth = 2
else:
self.out_depth = encoder.input_tensor.shape[3]
net = mf.relu(mf.concat(net, encoder.l1))
net = mf.deconv(net,
filters=1,
kernel_size=(5, 5, self.out_depth),
stride=(2, 2))
self.output = net
def __init__(self, input_tensor, encoder, data_type, is_training, reuse):
net = input_tensor
with tf.variable_scope('decoder'):
with tf.variable_scope('layer-1'):
net = mf.relu(net)
net = mf.deconv(net, filters=256, kernel_size=5, stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
net = mf.dropout(net, .5)
with tf.variable_scope('layer-2'):
net = mf.relu(mf.concat(net, encoder.l5))
net = mf.deconv(net, filters=128, kernel_size=5, stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
net = mf.dropout(net, .5)
with tf.variable_scope('layer-3'):
net = mf.relu(mf.concat(net, encoder.l4))
net = mf.deconv(net, filters=64, kernel_size=5, stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
net = mf.dropout(net, .5)
with tf.variable_scope('layer-4'):
net = mf.relu(mf.concat(net, encoder.l3))
net = mf.deconv(net, filters=32, kernel_size=5, stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
with tf.variable_scope('layer-5'):
net = mf.relu(mf.concat(net, encoder.l2))
net = mf.deconv(net, filters=16, kernel_size=5, stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
with tf.variable_scope('layer-6'):
if data_type == 'mag_phase_real_imag':
out_shape = 4
else:
out_shape = 2
net = mf.relu(mf.concat(net, encoder.l1))
net = mf.deconv(net,
filters=out_shape,
kernel_size=5,
stride=(2, 2))
self.output = net
def __init__(self, input_tensor, encoder, is_training, reuse):
net = input_tensor
with tf.variable_scope('decoder'):
with tf.variable_scope('layer-1'):
net = mf.relu(net)
net = mf.deconv(net, filters=256, kernel_size=5, stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
net = mf.dropout(net, .5)
with tf.variable_scope('layer-2'):
net = mf.relu(mf.concat(net, encoder.l5))
net = mf.deconv(net, filters=128, kernel_size=5, stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
net = mf.dropout(net, .5)
with tf.variable_scope('layer-3'):
net = mf.relu(mf.concat(net, encoder.l4))
net = mf.deconv(net, filters=64, kernel_size=5, stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
net = mf.dropout(net, .5)
with tf.variable_scope('layer-4'):
net = mf.relu(mf.concat(net, encoder.l3))
net = mf.deconv(net, filters=32, kernel_size=5, stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
with tf.variable_scope('layer-5'):
net = mf.relu(mf.concat(net, encoder.l2))
net = mf.deconv(net, filters=16, kernel_size=5, stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
with tf.variable_scope('layer-6'):
net = mf.relu(mf.concat(net, encoder.l1))
net = mf.deconv(net,
filters=encoder.input_tensor.shape[3],
kernel_size=5,
stride=(2, 2))
self.output = net
def __init__(self, mixed_input, voice_input, mixed_phase, mixed_audio,
voice_audio, background_audio, is_training, learning_rate,
data_type, phase_weight, phase_loss_masking,
phase_loss_approximation, name):
with tf.variable_scope(name):
self.mixed_input = mixed_input
self.voice_input = voice_input
self.mixed_phase = mixed_phase
self.mixed_audio = mixed_audio
self.voice_audio = voice_audio
self.background_audio = background_audio
self.is_training = is_training
# Initialise the selected model variant
if data_type == 'complex_to_mag_phase':
self.voice_mask_network = UNet(mixed_input[:, :, :, 0:2],
data_type,
is_training=is_training,
reuse=False,
name='voice-mask-unet')
else:
self.voice_mask_network = UNet(mixed_input,
data_type,
is_training=is_training,
reuse=False,
name='voice-mask-unet')
self.voice_mask = self.voice_mask_network.output
# Depending on the data_type, setup the loss functions and optimisation
if data_type == 'mag':
self.gen_voice = self.voice_mask * mixed_input
self.cost = mf.l1_loss(self.gen_voice, voice_input)
elif data_type == 'mag_phase':
self.gen_voice = self.voice_mask * mixed_input
self.mag_loss = mf.l1_loss(self.gen_voice[:, :, :, 0],
voice_input[:, :, :, 0])
self.phase_loss = mf.l1_phase_loss(
self.gen_voice[:, :, :, 1], voice_input[:, :, :, 1],
phase_loss_masking, phase_loss_approximation,
self.gen_voice[:, :, :, 0]) * phase_weight
#self.phase_loss = mf.l1_masked_phase_loss(self.gen_voice[:, :, :, 1], voice_input[:, :, :, 1], self.voice_input[:, :, :, 0]) * phase_weight
self.cost = (self.mag_loss + self.phase_loss) / 2
elif data_type == 'mag_phase_diff2':
self.gen_voice_mag = tf.expand_dims(
self.voice_mask[:, :, :, 0] * mixed_input[:, :, :, 0],
axis=3)
self.mag_loss = mf.l1_loss(self.gen_voice_mag[:, :, :, 0],
voice_input[:, :, :, 0])
self.phase_loss = mf.l1_phase_loss(
mf.phase_difference(
mixed_input[:, :, :, 1],
voice_input[:, :, :, 1]), self.voice_mask[:, :, :, 1],
phase_loss_masking, phase_loss_approximation,
self.gen_voice_mag) * phase_weight
self.cost = (self.mag_loss + self.phase_loss) / 2
self.gen_voice_phase = tf.expand_dims(
self.voice_mask[:, :, :, 1] + mixed_input[:, :, :, 1],
axis=3)
self.gen_voice = mf.concat(self.gen_voice_mag,
self.gen_voice_phase)
elif data_type == 'mag_phase_diff':
self.gen_voice_mag = tf.expand_dims(
self.voice_mask[:, :, :, 0] * mixed_input[:, :, :, 0],
axis=3)
self.gen_voice_phase = tf.expand_dims(
self.voice_mask[:, :, :, 1] + mixed_input[:, :, :, 1],
axis=3)
self.gen_voice = mf.concat(self.gen_voice_mag,
self.gen_voice_phase)
self.mag_loss = mf.l1_loss(self.gen_voice[:, :, :, 0],
voice_input[:, :, :, 0])
self.phase_loss = mf.l1_phase_loss(
self.gen_voice_phase, voice_input[:, :, :, 1],
phase_loss_masking, phase_loss_approximation,
self.gen_voice_mag) * phase_weight
self.cost = (self.mag_loss + self.phase_loss) / 2
elif data_type == 'real_imag':
self.gen_voice = self.voice_mask * mixed_input
self.real_loss = mf.l1_loss(self.gen_voice[:, :, :, 0],
voice_input[:, :, :, 0])
self.imag_loss = mf.l1_loss(self.gen_voice[:, :, :, 1],
voice_input[:, :, :, 1])
self.cost = (self.real_loss + self.imag_loss) / 2
elif data_type == 'mag_real_imag':
self.gen_voice = self.voice_mask * mixed_input
self.mag_loss = mf.l1_loss(self.gen_voice[:, :, :, 0],
voice_input[:, :, :, 0])
self.real_loss = mf.l1_loss(self.gen_voice[:, :, :, 1],
voice_input[:, :, :, 1])
self.imag_loss = mf.l1_loss(self.gen_voice[:, :, :, 2],
voice_input[:, :, :, 2])
self.cost = (self.mag_loss + self.real_loss +
self.imag_loss) / 3
elif data_type == 'mag_phase2':
self.mag_mask = self.voice_mask[:, :, :, 0]
self.phase_mask = tf.angle(
tf.complex(self.voice_mask[:, :, :, 1],
self.voice_mask[:, :, :, 2]))
self.voice_mask = mf.concat(
tf.expand_dims(self.mag_mask, axis=3),
tf.expand_dims(self.phase_mask, axis=3))
self.gen_voice_mag = self.mag_mask * mixed_input[:, :, :, 0]
self.gen_voice_phase = self.phase_mask * tf.squeeze(
mixed_phase, axis=3)
self.voice_phase = tf.angle(
tf.complex(self.voice_input[:, :, :, 1],
self.voice_input[:, :, :, 2]))
self.gen_voice = mf.concat(
tf.expand_dims(self.gen_voice_mag, axis=3),
tf.expand_dims(self.gen_voice_phase, axis=3))
self.mag_loss = mf.l1_loss(self.gen_voice_mag,
voice_input[:, :, :, 0])
self.phase_loss = mf.l1_phase_loss(
self.gen_voice_phase, self.voice_phase, phase_loss_masking,
phase_loss_approximation,
self.gen_voice_mag) * phase_weight
self.cost = (self.mag_loss + self.phase_loss) / 2
elif data_type == 'mag_phase_real_imag':
self.gen_voice = self.voice_mask * mixed_input[:, :, :, 2:4]
self.mag_loss = mf.l1_loss(self.gen_voice[:, :, :, 0],
voice_input[:, :, :, 2])
self.phase_loss = mf.l1_phase_loss(
self.gen_voice[:, :, :, 1], voice_input[:, :, :, 3],
phase_loss_masking, phase_loss_approximation,
self.gen_voice[:, :, :, 0]) * phase_weight
self.cost = (self.mag_loss + self.phase_loss) / 2
elif data_type == 'complex_to_mag_phase':
self.gen_voice = self.voice_mask * mixed_input[:, :, :, 2:4]
self.mag_loss = mf.l1_loss(self.gen_voice[:, :, :, 0],
voice_input[:, :, :, 2])
self.phase_loss = mf.l1_phase_loss(
self.gen_voice[:, :, :, 1], voice_input[:, :, :, 3],
phase_loss_masking, phase_loss_approximation,
self.gen_voice[:, :, :, 0]) * phase_weight
self.cost = (self.mag_loss + self.phase_loss) / 2
self.optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate,
beta1=0.5,
)
self.train_op = self.optimizer.minimize(self.cost)
def __init__(self, mixed_input, voice_input, mixed_phase, mixed_audio,
voice_audio, background_audio, variant, is_training,
learning_rate, data_type, phase_weight, name):
with tf.variable_scope(name):
self.mixed_input = mixed_input
self.voice_input = voice_input
self.mixed_phase = mixed_phase
self.mixed_audio = mixed_audio
self.voice_audio = voice_audio
self.background_audio = background_audio
self.variant = variant
self.is_training = is_training
if self.variant in ['unet', 'capsunet']:
self.voice_mask_network = UNet(mixed_input,
variant,
data_type,
is_training=is_training,
reuse=False,
name='voice-mask-unet')
elif self.variant == 'basic_capsnet':
self.voice_mask_network = BasicCapsnet(
mixed_input, name='SegCaps_CapsNetBasic')
elif self.variant == 'conv_net':
self.voice_mask_network = conv_net(mixed_input,
is_training=is_training,
reuse=None,
name='basic_cnn')
self.voice_mask = self.voice_mask_network.output
if data_type == 'mag':
self.gen_voice = self.voice_mask * mixed_input
self.cost = mf.l1_loss(self.gen_voice, voice_input)
elif data_type in ['mag_phase']:
self.gen_voice = self.voice_mask * mixed_input
self.mag_loss = mf.l1_loss(self.gen_voice[:, :, :, 0],
voice_input[:, :, :, 0])
self.phase_loss = mf.l1_phase_loss(
self.gen_voice[:, :, :, 1], voice_input[:, :, :,
1]) * phase_weight
self.cost = (self.mag_loss + self.phase_loss) / 2
elif data_type == 'mag_phase_diff':
self.gen_voice_mag = tf.expand_dims(
self.voice_mask[:, :, :, 0] * mixed_input[:, :, :, 0],
axis=3)
self.mag_loss = mf.l1_loss(self.gen_voice_mag[:, :, :, 0],
voice_input[:, :, :, 0])
self.phase_loss = mf.l1_phase_loss(
mf.phase_difference(mixed_input[:, :, :, 1],
voice_input[:, :, :, 1]),
self.voice_mask[:, :, :, 1]) * 0.00001
self.cost = (self.mag_loss + self.phase_loss) / 2
self.gen_voice_phase = tf.expand_dims(
self.voice_mask[:, :, :, 1] + mixed_input[:, :, :, 1],
axis=3)
self.gen_voice = mf.concat(self.gen_voice_mag,
self.gen_voice_phase)
elif data_type == 'real_imag':
self.gen_voice = self.voice_mask * mixed_input
self.real_loss = mf.l1_loss(self.gen_voice[:, :, :, 0],
voice_input[:, :, :, 0])
self.imag_loss = mf.l1_loss(self.gen_voice[:, :, :, 1],
voice_input[:, :, :, 1])
self.cost = (self.real_loss + self.imag_loss) / 2
elif data_type == 'mag_real_imag':
self.gen_voice = self.voice_mask * mixed_input
self.mag_loss = mf.l1_loss(self.gen_voice[:, :, :, 0],
voice_input[:, :, :, 0])
self.real_loss = mf.l1_loss(self.gen_voice[:, :, :, 1],
voice_input[:, :, :, 1])
self.imag_loss = mf.l1_loss(self.gen_voice[:, :, :, 2],
voice_input[:, :, :, 2])
self.cost = (self.mag_loss + self.real_loss +
self.imag_loss) / 3
elif data_type == 'mag_phase2':
self.mag_mask = self.voice_mask[:, :, :, 0]
self.phase_mask = tf.angle(
tf.complex(self.voice_mask[:, :, :, 1],
self.voice_mask[:, :, :, 2]))
self.voice_mask = mf.concat(
tf.expand_dims(self.mag_mask, axis=3),
tf.expand_dims(self.phase_mask, axis=3))
self.gen_mag = self.mag_mask * mixed_input[:, :, :, 0]
self.gen_phase = self.phase_mask * tf.squeeze(mixed_phase,
axis=3)
self.voice_phase = tf.angle(
tf.complex(self.voice_input[:, :, :, 1],
self.voice_input[:, :, :, 2]))
self.gen_voice = mf.concat(
tf.expand_dims(self.gen_mag, axis=3),
tf.expand_dims(self.gen_phase, axis=3))
self.mag_loss = mf.l1_loss(self.gen_mag, voice_input[:, :, :,
0])
self.phase_loss = mf.l1_phase_loss(
self.gen_phase, self.voice_phase) * phase_weight
self.cost = (self.mag_loss + self.phase_loss) / 2
elif data_type in ['mag_phase_real_imag']:
self.gen_voice = self.voice_mask * mixed_input[:, :, :, 2:4]
self.mag_loss = mf.l1_loss(self.gen_voice[:, :, :, 0],
voice_input[:, :, :, 2])
self.phase_loss = mf.l1_phase_loss(
self.gen_voice[:, :, :, 1], voice_input[:, :, :,
3]) * phase_weight
self.cost = (self.mag_loss + self.phase_loss) / 2
self.optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate,
beta1=0.5,
)
self.train_op = self.optimizer.minimize(self.cost)
def __init__(self, mixed_input, voice_input, mixed_phase, mixed_audio,
voice_audio, background_audio, variant, is_training,
learning_rate, data_type, phase_weight, phase_loss_function,
name):
with tf.variable_scope(name):
self.mixed_input = mixed_input
self.voice_input = voice_input
self.mixed_phase = mixed_phase
self.mixed_audio = mixed_audio
self.voice_audio = voice_audio
self.background_audio = background_audio
self.variant = variant
self.is_training = is_training
# Set the loss function
if phase_loss_function == 'l1':
self.phase_loss_function = mf.l1_loss
elif phase_loss_function == 'l2':
self.phase_loss_function = mf.l2_loss
elif phase_loss_function == 'l1_crcular':
self.phase_loss_function = mf.l1_phase_loss
elif phase_loss_function == 'l2_circular':
self.phase_loss_function = mf.l2_phase_loss
# Initialise the selected model variant
if self.variant in ['unet', 'capsunet', 'noconvcapsunet'
] and data_type == 'complex_to_mag_phase':
self.voice_mask_network = UNet(mixed_input[:, :, :, 0:2],
variant,
data_type,
is_training=is_training,
reuse=False,
name='voice-mask-unet')
elif self.variant in ['unet', 'capsunet', 'noconvcapsunet']:
self.voice_mask_network = UNet(mixed_input,
variant,
data_type,
is_training=is_training,
reuse=False,
name='voice-mask-unet')
elif self.variant == 'basic_capsnet':
self.voice_mask_network = BasicCapsNet(mixed_input,
name='basic_capsnet')
elif self.variant == 'basic_convnet':
self.voice_mask_network = BasicConvNet(mixed_input,
is_training=is_training,
reuse=None,
name='basic_convnet')
self.voice_mask = self.voice_mask_network.output
# Depending on the data_type, setup the loss functions and optimisation
if data_type == 'mag':
self.gen_voice = self.voice_mask * mixed_input
self.cost = mf.l1_loss(self.gen_voice, voice_input)
elif data_type == 'mag_phase':
self.gen_voice = self.voice_mask * mixed_input
self.mag_loss = mf.l1_loss(self.gen_voice[:, :, :, 0],
voice_input[:, :, :, 0])
#self.phase_loss = mf.l1_phase_loss(self.gen_voice[:, :, :, 1], voice_input[:, :, :, 1]) * phase_weight
self.phase_loss = self.phase_loss_function(
self.gen_voice[:, :, :, 1], voice_input[:, :, :,
1]) * phase_weight
self.cost = (self.mag_loss + self.phase_loss) / 2
elif data_type == 'mag_phase_diff2':
self.gen_voice_mag = tf.expand_dims(
self.voice_mask[:, :, :, 0] * mixed_input[:, :, :, 0],
axis=3)
self.mag_loss = mf.l1_loss(self.gen_voice_mag[:, :, :, 0],
voice_input[:, :, :, 0])
self.phase_loss = self.phase_loss_function(
mf.phase_difference(mixed_input[:, :, :, 1],
voice_input[:, :, :, 1]),
self.voice_mask[:, :, :, 1]) * phase_weight
self.cost = (self.mag_loss + self.phase_loss) / 2
self.gen_voice_phase = tf.expand_dims(
self.voice_mask[:, :, :, 1] + mixed_input[:, :, :, 1],
axis=3)
self.gen_voice = mf.concat(self.gen_voice_mag,
self.gen_voice_phase)
elif data_type == 'mag_phase_diff':
self.gen_voice_mag = tf.expand_dims(
self.voice_mask[:, :, :, 0] * mixed_input[:, :, :, 0],
axis=3)
self.gen_voice_phase = tf.expand_dims(
self.voice_mask[:, :, :, 1] + mixed_input[:, :, :, 1],
axis=3)
self.gen_voice = mf.concat(self.gen_voice_mag,
self.gen_voice_phase)
self.mag_loss = mf.l1_loss(self.gen_voice[:, :, :, 0],
voice_input[:, :, :, 0])
self.phase_loss = self.phase_loss_function(
self.gen_voice[:, :, :, 1], voice_input[:, :, :,
1]) * phase_weight
self.cost = (self.mag_loss + self.phase_loss) / 2
elif data_type == 'real_imag':
self.gen_voice = self.voice_mask * mixed_input
self.real_loss = mf.l1_loss(self.gen_voice[:, :, :, 0],
voice_input[:, :, :, 0])
self.imag_loss = mf.l1_loss(self.gen_voice[:, :, :, 1],
voice_input[:, :, :, 1])
self.cost = (self.real_loss + self.imag_loss) / 2
elif data_type == 'mag_real_imag':
self.gen_voice = self.voice_mask * mixed_input
self.mag_loss = mf.l1_loss(self.gen_voice[:, :, :, 0],
voice_input[:, :, :, 0])
self.real_loss = mf.l1_loss(self.gen_voice[:, :, :, 1],
voice_input[:, :, :, 1])
self.imag_loss = mf.l1_loss(self.gen_voice[:, :, :, 2],
voice_input[:, :, :, 2])
self.cost = (self.mag_loss + self.real_loss +
self.imag_loss) / 3
elif data_type == 'mag_phase2':
self.mag_mask = self.voice_mask[:, :, :, 0]
self.phase_mask = tf.angle(
tf.complex(self.voice_mask[:, :, :, 1],
self.voice_mask[:, :, :, 2]))
self.voice_mask = mf.concat(
tf.expand_dims(self.mag_mask, axis=3),
tf.expand_dims(self.phase_mask, axis=3))
self.gen_mag = self.mag_mask * mixed_input[:, :, :, 0]
self.gen_phase = self.phase_mask * tf.squeeze(mixed_phase,
axis=3)
self.voice_phase = tf.angle(
tf.complex(self.voice_input[:, :, :, 1],
self.voice_input[:, :, :, 2]))
self.gen_voice = mf.concat(
tf.expand_dims(self.gen_mag, axis=3),
tf.expand_dims(self.gen_phase, axis=3))
self.mag_loss = mf.l1_loss(self.gen_mag, voice_input[:, :, :,
0])
self.phase_loss = self.phase_loss_function(
self.gen_phase, self.voice_phase) * phase_weight
self.cost = (self.mag_loss + self.phase_loss) / 2
elif data_type == 'mag_phase_real_imag':
self.gen_voice = self.voice_mask * mixed_input[:, :, :, 2:4]
self.mag_loss = mf.l1_loss(self.gen_voice[:, :, :, 0],
voice_input[:, :, :, 2])
self.phase_loss = self.phase_loss_function(
self.gen_voice[:, :, :, 1], voice_input[:, :, :,
3]) * phase_weight
self.cost = (self.mag_loss + self.phase_loss) / 2
elif data_type == 'complex_to_mag_phase':
self.gen_voice = self.voice_mask * mixed_input[:, :, :, 2:4]
self.mag_loss = mf.l1_loss(self.gen_voice[:, :, :, 0],
voice_input[:, :, :, 2])
self.phase_loss = self.phase_loss_function(
self.gen_voice[:, :, :, 1], voice_input[:, :, :,
3]) * phase_weight
self.cost = (self.mag_loss + self.phase_loss) / 2
self.optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate,
beta1=0.5,
)
self.train_op = self.optimizer.minimize(self.cost)
