def encode(x, hparams, is_training=True, reuse=False):
"""Autoencoder encoder network.
Args:
x: Tensor. The observed variables.
hparams: HParams. Hyperparameters.
is_training: bool. Whether batch normalization should be computed in
training mode. Defaults to True.
reuse: bool. Whether the variable scope should be reused.
Defaults to False.
Returns:
The output of the encoder, i.e. a synthetic z computed from x.
"""
with tf.variable_scope("encoder", reuse=reuse):
h = utils.conv2d(x, [5, 5], [2, 2],
128,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="0")
h = utils.conv2d(h, [4, 4], [2, 2],
128,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="1")
h = utils.conv2d(h, [4, 4], [2, 2],
128,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="2")
h = utils.conv2d(h, [4, 4], [2, 2],
256,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="3")
h = utils.conv2d(h, [4, 4], [2, 2],
256,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="4")
h = utils.conv2d(h, [4, 4], [2, 2],
256,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="5")
h = utils.conv2d(h, [4, 4], [2, 2],
512,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="6")
h = utils.conv2d(h, [4, 4], [2, 2],
512,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="7")
h = utils.conv2d(h, [4, 4], [2, 1],
512,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="7_1")
h = utils.conv2d(h, [1, 1], [1, 1],
1024,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="8")
z = utils.conv2d(h, [1, 1], [1, 1],
hparams.num_latent,
is_training,
activation_fn=None,
batch_norm=True,
scope="z")
return z
def encode(x, hparams, is_training=True, reuse=False):
"""Autoencoder encoder network.
Args:
x: Tensor. The observed variables.
hparams: HParams. Hyperparameters.
is_training: bool. Whether batch normalization should be computed in
training mode. Defaults to True.
reuse: bool. Whether the variable scope should be reused.
Defaults to False.
Returns:
The output of the encoder, i.e. a synthetic z computed from x.
"""
with tf.variable_scope("encoder", reuse=reuse):
h = utils.conv2d(
x, [5, 5], [2, 2],
128,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="0")
h = utils.conv2d(
h, [4, 4], [2, 2],
128,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="1")
h = utils.conv2d(
h, [4, 4], [2, 2],
128,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="2")
h = utils.conv2d(
h, [4, 4], [2, 2],
256,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="3")
h = utils.conv2d(
h, [4, 4], [2, 2],
256,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="4")
h = utils.conv2d(
h, [4, 4], [2, 2],
256,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="5")
h = utils.conv2d(
h, [4, 4], [2, 2],
512,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="6")
h = utils.conv2d(
h, [4, 4], [2, 2],
512,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="7")
h = utils.conv2d(
h, [4, 4], [2, 1],
512,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="7_1")
h = utils.conv2d(
h, [1, 1], [1, 1],
1024,
is_training,
activation_fn=utils.leaky_relu(),
batch_norm=True,
scope="8")
z = utils.conv2d(
h, [1, 1], [1, 1],
hparams.num_latent,
is_training,
activation_fn=None,
batch_norm=True,
scope="z")
return z
def decode(z, batch, hparams, is_training=True, reuse=False):
"""Autoencoder decoder network.
Args:
z: Tensor. The latent variables.
batch: NSynthReader batch for pitch information.
hparams: HParams. Hyperparameters (unused).
is_training: bool. Whether batch normalization should be computed in
training mode. Defaults to True.
reuse: bool. Whether the variable scope should be reused.
Defaults to False.
Returns:
The output of the decoder, i.e. a synthetic x computed from z.
"""
del hparams
with tf.variable_scope("decoder", reuse=reuse):
z_pitch = utils.pitch_embeddings(batch, reuse=reuse)
z = tf.concat([z, z_pitch], 3)
h = utils.conv2d(z, [1, 1], [1, 1],
1024,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="0")
h = utils.conv2d(h, [4, 4], [2, 2],
512,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="1")
h = utils.conv2d(h, [4, 4], [2, 2],
512,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="2")
h = utils.conv2d(h, [4, 4], [2, 2],
256,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="3")
h = utils.conv2d(h, [4, 4], [2, 2],
256,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="4")
h = utils.conv2d(h, [4, 4], [2, 2],
256,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="5")
h = utils.conv2d(h, [4, 4], [2, 2],
128,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="6")
h = utils.conv2d(h, [4, 4], [2, 2],
128,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="7")
h = utils.conv2d(h, [5, 5], [2, 2],
128,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="8")
h = utils.conv2d(h, [5, 5], [2, 1],
128,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="8_1")
xhat = utils.conv2d(h, [1, 1], [1, 1],
1,
is_training,
activation_fn=tf.nn.sigmoid,
batch_norm=False,
scope="mag")
return xhat
def decode(z, batch, hparams, is_training=True, reuse=False):
"""Autoencoder decoder network.
Args:
z: Tensor. The latent variables.
batch: NSynthReader batch for pitch information.
hparams: HParams. Hyperparameters (unused).
is_training: bool. Whether batch normalization should be computed in
training mode. Defaults to True.
reuse: bool. Whether the variable scope should be reused.
Defaults to False.
Returns:
The output of the decoder, i.e. a synthetic x computed from z.
"""
del hparams
with tf.variable_scope("decoder", reuse=reuse):
z_pitch = utils.pitch_embeddings(batch, reuse=reuse)
z = tf.concat([z, z_pitch], 3)
h = utils.conv2d(
z, [1, 1], [1, 1],
1024,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="0")
h = utils.conv2d(
h, [4, 4], [2, 2],
512,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="1")
h = utils.conv2d(
h, [4, 4], [2, 2],
512,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="2")
h = utils.conv2d(
h, [4, 4], [2, 2],
256,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="3")
h = utils.conv2d(
h, [4, 4], [2, 2],
256,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="4")
h = utils.conv2d(
h, [4, 4], [2, 2],
256,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="5")
h = utils.conv2d(
h, [4, 4], [2, 2],
128,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="6")
h = utils.conv2d(
h, [4, 4], [2, 2],
128,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="7")
h = utils.conv2d(
h, [5, 5], [2, 2],
128,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="8")
h = utils.conv2d(
h, [5, 5], [2, 1],
128,
is_training,
activation_fn=utils.leaky_relu(),
transpose=True,
batch_norm=True,
scope="8_1")
xhat = utils.conv2d(
h, [1, 1], [1, 1],
1,
is_training,
activation_fn=tf.nn.sigmoid,
batch_norm=False,
scope="mag")
return xhat
