def _encode_state(self, state, reuse):
with tf.variable_scope('state_encoder', reuse=reuse):
x = self._dense_norm_activation(
state,
512,
kernel_initializer=tf_utils.kaiming_initializer(),
activation=tf.nn.relu)
return x
def _critic(self, state, action, reuse):
x = self._encode_state(state, True)
with tf.variable_scope('critic', reuse=reuse):
x = tf.concat([x, action], 1)
x = self._dense_norm_activation(
x,
256,
kernel_initializer=tf_utils.kaiming_initializer(),
activation=tf.nn.relu)
x = self._dense(x, 1)
return x
def _actor(self, state):
x = self._encode_state(state, self.reuse)
with tf.variable_scope('actor', reuse=self.reuse):
x = self._dense_norm_activation(
x,
256,
kernel_initializer=tf_utils.kaiming_initializer(),
activation=tf.nn.relu)
x = self._dense_norm_activation(x, self.action_size)
x = tf.clip_by_value(x, -1, 1, name='action')
return x
def _dense_ln_relu(self,
x,
units,
kernel_initializer=tf_utils.kaiming_initializer(),
trainable=True,
name=None,
reuse=None):
x = self._dense(x,
units,
kernel_initializer=kernel_initializer,
trainable=trainable,
name=name,
reuse=reuse)
x = tf_utils.ln_relu(x)
return x
def _convtrans_bn_relu(self,
x,
filters,
kernel_size,
strides=1,
padding='same',
kernel_initializer=tf_utils.kaiming_initializer(),
trainable=True,
name=None,
reuse=None):
x = self._convtrans(x,
filters,
kernel_size,
strides=strides,
padding=padding,
kernel_initializer=kernel_initializer,
trainable=trainable,
name=name,
reuse=reuse)
x = tf_utils.bn_relu(x, self._training)
return x
