def prepare(self):
self._mu_encoder = Dense(self.size,
'linear',
init=GaussianInitializer(),
random_bias=True).init(self.input_dim)
self._log_sigma_encoder = Dense(self.size,
'linear',
init=GaussianInitializer(),
random_bias=True).init(self.input_dim)
self.register_inner_layers(self._mu_encoder, self._log_sigma_encoder)
class ReparameterizationLayer(NeuralLayer):
"""
Reparameterization layer in a Variational encoder.
Only binary output cost function is supported now.
The prior value is recorded after the computation graph created.
"""
def __init__(self, size, sample=False):
"""
:param size: the size of latent variable
:param sample: whether to get a clean latent variable
"""
super(ReparameterizationLayer, self).__init__("VariationalEncoder")
self.size = size
self.output_dim = size
self.sample = sample
self._prior = None
def prepare(self):
self._mu_encoder = Dense(self.size,
'linear',
init=GaussianInitializer(),
random_bias=True).initialize(self.input_dim)
self._log_sigma_encoder = Dense(self.size,
'linear',
init=GaussianInitializer(),
random_bias=True).initialize(
self.input_dim)
self.register_inner_layers(self._mu_encoder, self._log_sigma_encoder)
def compute_tensor(self, x):
# Compute p(z|x)
mu = self._mu_encoder.compute_tensor(x)
log_sigma = 0.5 * self._log_sigma_encoder.compute_tensor(x)
self._prior = 0.5 * T.sum(1 + 2 * log_sigma - mu**2 -
T.exp(2 * log_sigma))
# Reparameterization
eps = global_theano_rand.normal((x.shape[0], self.size))
if self.sample:
z = mu
else:
z = mu + T.exp(log_sigma) * eps
return z
def prior(self):
"""
Get the prior value.
"""
return self._prior
class ReparameterizationLayer(NeuralLayer):
"""
Reparameterization layer in a Variational encoder.
Only binary output cost function is supported now.
The prior value is recorded after the computation graph created.
"""
def __init__(self, size, sample=False):
"""
:param size: the size of latent variable
:param sample: whether to get a clean latent variable
"""
super(ReparameterizationLayer, self).__init__("VariationalEncoder")
self.size = size
self.output_dim = size
self.sample = sample
self._prior = None
def prepare(self):
self._mu_encoder = Dense(self.size, 'linear', init=GaussianInitializer(), random_bias=True).initialize(
self.input_dim)
self._log_sigma_encoder = Dense(self.size, 'linear', init=GaussianInitializer(), random_bias=True).initialize(
self.input_dim)
self.register_inner_layers(self._mu_encoder, self._log_sigma_encoder)
def compute_tensor(self, x):
# Compute p(z|x)
mu = self._mu_encoder.compute_tensor(x)
log_sigma = 0.5 * self._log_sigma_encoder.compute_tensor(x)
self._prior = 0.5* T.sum(1 + 2*log_sigma - mu**2 - T.exp(2*log_sigma))
# Reparameterization
eps = global_theano_rand.normal((x.shape[0], self.size))
if self.sample:
z = mu
else:
z = mu + T.exp(log_sigma) * eps
return z
def prior(self):
"""
Get the prior value.
"""
return self._prior
def setup(self):
self._mu_encoder = Dense(self.size, 'linear', init=GaussianInitializer(), random_bias=True).connect(self.input_dim)
self._log_sigma_encoder = Dense(self.size, 'linear', init=GaussianInitializer(), random_bias=True).connect(self.input_dim)
self.register_inner_layers(self._mu_encoder, self._log_sigma_encoder)
