def initialize_objective(self):
H, ds, da = self.horizon, self.ds, self.da
if self.time_varying:
A = T.concatenate([T.eye(ds), T.zeros([ds, da])], -1)
self.A = T.variable(A[None] + 1e-2 * T.random_normal([H - 1, ds, ds + da]))
self.Q_log_diag = T.variable(T.random_normal([H - 1, ds]) + 1)
self.Q = T.matrix_diag(T.exp(self.Q_log_diag))
else:
A = T.concatenate([T.eye(ds), T.zeros([ds, da])], -1)
self.A = T.variable(A + 1e-2 * T.random_normal([ds, ds + da]))
self.Q_log_diag = T.variable(T.random_normal([ds]) + 1)
self.Q = T.matrix_diag(T.exp(self.Q_log_diag))
def log1pexp(x):
return T.log(1 + T.exp(x))
def log_likelihood(self, x):
m, b = Stats.X(m), Stats.X(b)
x = Stats.X(x)
z = (x - m) / b
return 1 / b * T.exp(-z - T.exp(-z))
from deepx import T
from deepx.nn import Linear
from deepx import stats
__all__ = ['Gaussian', 'Bernoulli']
log1pexp = lambda x: T.log(1. + T.exp(x) + 1e-4)
class Gaussian(Linear):
def __init__(self, *args, **kwargs):
self.cov_type = kwargs.pop('cov_type', 'diagonal')
super(Gaussian, self).__init__(*args, **kwargs)
assert not self.elementwise
def get_dim_out(self):
return [self.dim_out[0] * 2]
def activate(self, X):
import tensorflow as tf
if self.cov_type == 'diagonal':
sigma, mu = T.split(X, 2, axis=-1)
sigma = T.matrix_diag(log1pexp(sigma))
return stats.Gaussian([sigma, mu], parameter_type='regular')
raise Exception("Undefined covariance type: %s" % self.cov_type)
def __str__(self):
return "Gaussian(%s)" % self.dim_out
class GaussianStats(Linear):
def natural_to_regular(cls, natural_parameters):
eta = natural_parameters[Stats.X]
return T.exp(eta)
def log_likelihood(self, batch_z, batch):
x = Vector(self.input_size, placeholder=batch, is_input=False)
mu, sigma = (x >> self.q_network).get_graph_outputs()
sigma = T.sqrt(T.exp(sigma))
return T.mean(
log_normal(batch_z, mu, sigma, self.embedding_size, dim=2))
def sample_z(self, batch, batch_noise, feed_dict={}):
x = Vector(self.input_size, placeholder=batch, is_input=False)
mu, sigma = (x >> self.q_network).get_graph_outputs()
sigma = T.sqrt(T.exp(sigma))
return mu + sigma * batch_noise