def __init__(self, mixed_mag, is_training, reuse, name):
"""
input_tensor: Tensor with shape [batch_size, height, width, channels]
is_training: Boolean - should the model be trained on the current input or not
name: Model instance name
"""
with tf.variable_scope(name):
self.mixed_mag = mixed_mag
with tf.variable_scope('Convolution'):
net = mf.relu(mixed_mag)
net = mf.conv(net, filters=128, kernel_size=5, stride=(1, 1))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.conv1 = net
with tf.variable_scope('Primary_Caps'):
net = mf.relu(net)
net = mf.conv(net, filters=128, kernel_size=5, stride=(1, 1))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.primary_caps = net
with tf.variable_scope('Seg_Caps'):
net = mf.relu(net)
net = mf.conv(net, filters=16, kernel_size=5, stride=(1, 1))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.seg_caps = net
with tf.variable_scope('Mask'):
net = mf.relu(net)
net = mf.conv(mixed_mag, filters=1, kernel_size=5, stride=(1, 1))
self.voice_mask = net
self.output = net
def __init__(self, input_tensor, is_training, reuse):
self.input_tensor = input_tensor
with tf.variable_scope('Encoder'):
# conv1 = layers.Conv2D(filters=16, kernel_size=5, strides=1, padding='same', activation='relu', name='conv1')(x)
_, H, W, C = self.input_tensor.get_shape()
net = layers.Reshape((H.value, W.value, 1, C.value))(self.input_tensor)
net = capsule_layers.ConvCapsuleLayer(kernel_size=5, num_capsule=2, num_atoms=2, strides=2, padding='same',
routings=3, name='caps_conv1')(net)
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.caps_conv1 = net
net = capsule_layers.ConvCapsuleLayer(kernel_size=5, num_capsule=4, num_atoms=4, strides=2, padding='same',
routings=3, name='caps_conv2')(net)
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.caps_conv2 = net
net = capsule_layers.ConvCapsuleLayer(kernel_size=5, num_capsule=8, num_atoms=8, strides=2, padding='same',
routings=3, name='caps_conv3')(net)
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.caps_conv3 = net
net = capsule_layers.ConvCapsuleLayer(kernel_size=5, num_capsule=16, num_atoms=16, strides=2, padding='same',
routings=3, name='caps_conv4')(net)
#self.caps_conv4 = net
#net = capsule_layers.ConvCapsuleLayer(kernel_size=5, num_capsule=32, num_atoms=2, strides=2, padding='same',
# routings=3, name='caps_conv5')(net)
#self.caps_conv5 = net
#net = capsule_layers.ConvCapsuleLayer(kernel_size=5, num_capsule=64, num_atoms=2, strides=2, padding='same',
# routings=3, name='caps_conv6')(net)
self.output = net
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, is_training, reuse):
self.input_tensor = input_tensor
self.input_depth = self.input_tensor.shape[3]
with tf.variable_scope('encoder'):
with tf.variable_scope('layer-1'):
net = mf.conv3d(self.input_tensor,
filters=8,
kernel_size=(5, 5, self.input_depth),
stride=(2, 2))
self.l1 = net
with tf.variable_scope('layer-2'):
net = mf.lrelu(net)
net = mf.conv3d(net,
filters=16,
kernel_size=(5, 5, self.input_depth),
stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.l2 = net
with tf.variable_scope('layer-3'):
net = mf.lrelu(net)
net = mf.conv3d(net,
filters=32,
kernel_size=(5, 5, self.input_depth),
stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.l3 = net
with tf.variable_scope('layer-4'):
net = mf.lrelu(net)
net = mf.conv3d(net,
filters=64,
kernel_size=(5, 5, self.input_depth),
stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.l4 = net
with tf.variable_scope('layer-5'):
net = mf.lrelu(net)
net = mf.conv3d(net,
filters=128,
kernel_size=(5, 5, self.input_depth),
stride=(2, 2))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.l5 = net
with tf.variable_scope('layer-6'):
net = mf.lrelu(net)
net = mf.conv3d(net,
filters=256,
kernel_size=(5, 5, self.input_depth),
stride=(2, 2))
self.output = net
def __init__(self, input_tensor, encoder, is_training, reuse):
net = input_tensor
with tf.variable_scope('Decoder'):
# Layer 1 Up: Deconvolutional capsules, skip connection, convolutional capsules
net = capsule_layers.DeconvCapsuleLayer(kernel_size=4, num_capsule=8, num_atoms=16, upsamp_type='deconv',
scaling=2, padding='same', routings=3, name='deconv_cap_1')(net)
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.upcap_1 = net
net = layers.Concatenate(axis=-2, name='skip_1')([net, encoder.conv_cap_3])
# Layer 2 Up: Deconvolutional capsules, skip connection, convolutional capsules
net = capsule_layers.DeconvCapsuleLayer(kernel_size=4, num_capsule=4, num_atoms=8, upsamp_type='deconv',
scaling=2, padding='same', routings=3, name='deconv_cap_2')(net)
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.upcap_2 = net
net = layers.Concatenate(axis=-2, name='skip_2')([net, encoder.conv_cap_2])
# Layer 3 Up: Deconvolutional capsules, skip connection
net = capsule_layers.DeconvCapsuleLayer(kernel_size=4, num_capsule=2, num_atoms=8, upsamp_type='deconv',
scaling=2, padding='same', routings=3, name='deconv_cap_3')(net)
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.upcap_3 = net
net = layers.Concatenate(axis=-2, name='skip_3')([net, encoder.primary_caps])
# Layer 4 Up: Deconvolutional capsules, skip connection
net = capsule_layers.DeconvCapsuleLayer(kernel_size=4, num_capsule=1, num_atoms=16, upsamp_type='deconv',
scaling=2, padding='same', routings=3, name='deconv_cap_4')(net)
self.upcap_4 = net
# Reconstruction - Reshape, skip connection + 3x conventional Conv2D layers
_, H, W, C, D = net.get_shape()
net = layers.Reshape((H.value, W.value, D.value))(net)
net = layers.Concatenate(axis=-1, name='skip_4')([net, encoder.conv1])
net = layers.Conv2D(filters=64, kernel_size=1, padding='same', kernel_initializer='he_normal',
activation='relu', name='recon_1')(net)
net = layers.Conv2D(filters=128, kernel_size=1, padding='same', kernel_initializer='he_normal',
activation='relu', name='recon_2')(net)
if tf.rank(encoder.input_tensor) == 3:
self.out_depth = 1
else:
self.out_depth = encoder.input_tensor.shape[3].value
net = layers.Conv2D(filters=self.out_depth, kernel_size=1, padding='same', kernel_initializer='he_normal',
activation='sigmoid', name='out_recon')(net)
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, is_training, reuse, name):
"""
input_tensor: Tensor with shape [batch_size, height, width, channels]
is_training: Boolean - should the model be trained on the current input or not
name: Model instance name
"""
with tf.variable_scope(name):
self.input_tensor = input_tensor
if tf.rank(self.input_tensor) == 3:
self.out_depth = 1
else:
self.out_depth = input_tensor.shape[3].value
with tf.variable_scope('layer_1'):
net = mf.relu(input_tensor)
net = mf.conv(net, filters=128, kernel_size=5, stride=(1, 1))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.l1 = net
with tf.variable_scope('layer_2'):
net = mf.relu(net)
net = mf.conv(net, filters=128, kernel_size=5, stride=(1, 1))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.l2 = net
with tf.variable_scope('layer_3'):
net = mf.relu(net)
net = mf.conv(net, filters=16, kernel_size=5, stride=(1, 1))
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.l3 = net
with tf.variable_scope('mask'):
net = mf.relu(net)
net = mf.conv(net,
filters=self.out_depth,
kernel_size=5,
stride=(1, 1))
self.voice_mask = net
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, input_tensor, is_training, reuse):
# net = layers.Input(shape=input_tensor)
self.input_tensor = input_tensor
with tf.variable_scope('Encoder'):
with tf.variable_scope('Convolution'):
# Layer 1: A conventional Conv2D layer
net = layers.Conv2D(filters=16, kernel_size=5, strides=1, padding='same', activation='relu',
name='conv1')(self.input_tensor)
self.conv1 = net
# Reshape layer to be 1 capsule x [filters] atoms
_, H, W, C = net.get_shape()
net = layers.Reshape((H.value, W.value, 1, C.value))(net)
# Layer 1: Primary Capsule: Conv cap with routing 1
net = capsule_layers.ConvCapsuleLayer(kernel_size=5, num_capsule=2, num_atoms=8, strides=2, padding='same',
routings=1, name='primarycaps')(net)
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.primary_caps = net
# Layer 2: Convolutional Capsules
net = capsule_layers.ConvCapsuleLayer(kernel_size=5, num_capsule=4, num_atoms=8, strides=2, padding='same',
routings=3, name='conv_cap_2')(net)
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.conv_cap_2 = net
# Layer 3: Convolutional Capsules
net = capsule_layers.ConvCapsuleLayer(kernel_size=5, num_capsule=8, num_atoms=16, strides=2, padding='same',
routings=3, name='conv_cap_3')(net)
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.conv_cap_3 = net
# Layer 4: Convolutional Capsules
net = capsule_layers.ConvCapsuleLayer(kernel_size=5, num_capsule=8, num_atoms=32, strides=2, padding='same',
routings=3, name='conv_cap_4')(net)
self.output = net
def __init__(self, input_tensor, encoder, is_training, reuse):
net = input_tensor
with tf.variable_scope('Decoder'):
#net = capsule_layers.DeconvCapsuleLayer(kernel_size=5, num_capsule=32, num_atoms=2, upsamp_type='deconv',
# scaling=2, padding='same', routings=3,
# name='caps_deconv5')(net)
#self.upcap_1 = net
#net = layers.Concatenate(axis=3, name='up4')([net, encoder.caps_conv5])
#net = capsule_layers.DeconvCapsuleLayer(kernel_size=5, num_capsule=16, num_atoms=2, upsamp_type='deconv',
# scaling=2, padding='same', routings=3,
# name='caps_deconv4')(net)
#self.upcap_2 = net
#net = layers.Concatenate(axis=3, name='up3')([net, encoder.caps_conv4])
net = capsule_layers.DeconvCapsuleLayer(kernel_size=5,
num_capsule=8,
num_atoms=8,
upsamp_type='deconv',
scaling=2,
padding='same',
routings=3,
name='caps_deconv3')(net)
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.upcap_3 = net
net = layers.Concatenate(axis=3,
name='up2')([net, encoder.caps_conv3])
net = capsule_layers.DeconvCapsuleLayer(kernel_size=5,
num_capsule=4,
num_atoms=4,
upsamp_type='deconv',
scaling=2,
padding='same',
routings=3,
name='caps_deconv2')(net)
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.upcap_4 = net
net = layers.Concatenate(axis=3,
name='up1')([net, encoder.caps_conv2])
net = capsule_layers.DeconvCapsuleLayer(kernel_size=5,
num_capsule=2,
num_atoms=2,
upsamp_type='deconv',
scaling=2,
padding='same',
routings=3,
name='caps_deconv1')(net)
net = mf.batch_norm(net, is_training=is_training, reuse=reuse)
self.upcap_5 = net
net = layers.Concatenate(axis=3,
name='up0')([net, encoder.caps_conv1])
net = capsule_layers.DeconvCapsuleLayer(kernel_size=5,
num_capsule=1,
num_atoms=2,
upsamp_type='deconv',
scaling=2,
padding='same',
routings=3,
name='caps_deconv0')(net)
self.upcap_6 = net
_, H, W, C, A = net.get_shape()
net = layers.Reshape((H.value, W.value, A.value))(net)
#net = layers.Concatenate(axis=-1, name='skip_4')([net, encoder.conv1])
net = layers.Conv2D(filters=64,
kernel_size=1,
padding='same',
kernel_initializer='he_normal',
activation='relu',
name='recon_1')(net)
net = layers.Conv2D(filters=128,
kernel_size=1,
padding='same',
kernel_initializer='he_normal',
activation='relu',
name='recon_2')(net)
if tf.rank(encoder.input_tensor) == 3:
self.out_depth = 1
else:
self.out_depth = encoder.input_tensor.shape[3].value
net = layers.Conv2D(filters=self.out_depth,
kernel_size=1,
padding='same',
kernel_initializer='he_normal',
activation='sigmoid',
name='out_recon')(net)
self.output = net
