def __init__(self, model, optimizer, hyper):
super().__init__(model=model, optimizer=optimizer, hyper=hyper)
self.temp = tf.constant(value=hyper['temp'], dtype=tf.float32)
self.prior_file = hyper['prior_file']
self.mu_0 = tf.constant(value=0., dtype=tf.float32, shape=(1, 1, 1, 1))
self.xi_0 = tf.constant(value=0., dtype=tf.float32, shape=(1, 1, 1, 1))
self.ng = GaussianSoftmaxDist(mu=self.mu_0, xi=self.xi_0)
self.load_prior_values()
def reparameterize(self, params_broad):
mu, xi = params_broad
self.ng = GaussianSoftmaxDist(mu=mu,
xi=xi,
temp=self.temp,
sample_size=self.sample_size)
self.ng.do_reparameterization_trick()
z_discrete = [self.ng.log_psi]
return z_discrete
def determine_distribution(model_type, params, temp, samples_n):
if model_type == 'GSMDis':
dist = GaussianSoftmaxDist(mu=params[0],
xi=params[1],
sample_size=samples_n,
temp=temp)
elif model_type == 'ExpGSDis':
dist = ExpGSDist(log_pi=params[0], sample_size=samples_n, temp=temp)
else:
raise RuntimeError
return dist
def reparameterize(self, params_broad):
mean, log_var, mu, xi = params_broad
z_norm = sample_normal(mean=mean, log_var=log_var)
self.ng = GaussianSoftmaxDist(mu=mu,
xi=xi,
temp=self.temp,
sample_size=self.sample_size)
self.ng.do_reparameterization_trick()
z_discrete = self.ng.psi
self.n_required = z_discrete.shape[1]
z = [z_norm, z_discrete]
return z
def reparameterize(self, params_broad):
mu, xi = params_broad
epsilon = tf.random.normal(shape=mu.shape)
self.ng = GaussianSoftmaxDist(mu=mu,
xi=xi,
temp=self.temp,
sample_size=self.sample_size)
sigma = tf.math.exp(xi)
self.ng.lam = self.model.planar_flow(mu + sigma * epsilon)
self.ng.log_psi = self.ng.lam - tf.math.reduce_logsumexp(
self.ng.lam, axis=1, keepdims=True)
# psi = tf.math.softmax(lam / self.temp, axis=1)
z_discrete = [self.ng.log_psi]
return z_discrete