def _create_graph(self):
Dx = self._Dx
if len(self._cond_t_lst) == 0:
mu_and_logsig_t = tf.get_variable(
'params', self._layer_sizes[-1],
initializer=tf.random_normal_initializer(0, 0.1)
)
else:
mu_and_logsig_t = mlp(
inputs=self._cond_t_lst,
layer_sizes=self._layer_sizes,
output_nonlinearity=None,
) # ... x K*Dx*2+K
self._mu_t = mu_and_logsig_t[..., :Dx]
self._log_sig_t = tf.clip_by_value(mu_and_logsig_t[..., Dx:], LOG_SIG_CAP_MIN, LOG_SIG_CAP_MAX)
# Tensorflow's multivariate normal distribution supports reparameterization
ds = tf.contrib.distributions
dist = ds.MultivariateNormalDiag(loc=self._mu_t, scale_diag=tf.exp(self._log_sig_t))
x_t = dist.sample()
if not self._reparameterize:
x_t = tf.stop_gradient(x_t)
log_pi_t = dist.log_prob(x_t)
self._dist = dist
self._x_t = x_t
self._log_pi_t = log_pi_t
reg_loss_t = self._reg * 0.5 * tf.reduce_mean(self._log_sig_t ** 2)
reg_loss_t += self._reg * 0.5 * tf.reduce_mean(self._mu_t ** 2)
self._reg_loss_t = reg_loss_t
def _create_p_xz_params(self):
K = self._K
Dx = self._Dx
if len(self._cond_t_lst) == 0:
w_and_mu_and_logsig_t = tf.get_variable(
'params',
self._layer_sizes[-1],
initializer=tf.random_normal_initializer(0, 0.1))
else:
w_and_mu_and_logsig_t = mlp(
inputs=self._cond_t_lst,
layer_sizes=self._layer_sizes,
output_nonlinearity=None,
) # ... x K*Dx*2+K
w_and_mu_and_logsig_t = tf.reshape(w_and_mu_and_logsig_t,
shape=(-1, K, 2 * Dx + 1))
log_w_t = w_and_mu_and_logsig_t[..., 0]
mu_t = w_and_mu_and_logsig_t[..., 1:1 + Dx]
log_sig_t = w_and_mu_and_logsig_t[..., 1 + Dx:]
log_sig_t = tf.clip_by_value(log_sig_t, LOG_SIG_CAP_MIN,
LOG_SIG_CAP_MAX)
return log_w_t, mu_t, log_sig_t
def _create_graph(self):
logits = mlp(
inputs=self._cond_t_lst,
layer_sizes=self._layer_sizes,
output_nonlinearity=tf.nn.sigmoid,
)
l_logits = [
logits[:, self._slice[i]:self._slice[i + 1]]
for i in range(self._actnum)
]
self._logits = []
for i in range(self._actnum):
up_tri = tf.constant(
np.triu(np.ones((self._Dx[i], self._Dx[i]), dtype=np.float32)))
low_tri = tf.constant(
np.tril(np.ones((self._Dx[i], self._Dx[i]), dtype=np.float32),
-1))
ordinal_logits = tf.matmul(
tf.log(l_logits[i] + EPS), up_tri) + tf.matmul(
tf.log(1 - l_logits[i] + EPS), low_tri)
self._logits.append(ordinal_logits)
#self._logits = [logits[:,self._slice[i]:self._slice[i+1]] for i in range(self._actnum)]
self._probs = [tf.nn.softmax(logit) for logit in self._logits]
self._logprobs = [tf.nn.log_softmax(logit) for logit in self._logits]
def get_output_for(self, *inputs, reuse=False):
with tf.variable_scope(self._name, reuse=reuse):
value_t = mlp(
inputs=inputs,
output_nonlinearity=None,
layer_sizes=self._layer_sizes,
) # N
return value_t
def _create_graph(self):
logits = mlp(
inputs=self._cond_t_lst,
layer_sizes=self._layer_sizes,
output_nonlinearity=None,
)
self._logits = logits
self._probs = tf.nn.softmax(logits)
self._logprobs = tf.nn.log_softmax(logits)
def get_output_for(self, *inputs, **kwargs):
if 'reuse' not in kwargs.keys():
reuse = False
else:
reuse = kwargs['reuse']
with tf.variable_scope(self._name, reuse=reuse):
value_t = mlp(
inputs=inputs,
output_nonlinearity=None,
layer_sizes=self._layer_sizes,
) # N
return value_t
