def build_policy_network(self):
if self.args.use_gnn_as_policy:
self.policy_network = gated_graph_policy_network.GGNN(
session=self.session,
name_scope=self.name_scope + '_policy',
input_size=self.observation_size,
output_size=self.action_size,
ob_placeholder=None,
trainable=True,
build_network_now=True,
is_baseline=False,
placeholder_list=None,
args=self.args)
self.raw_obs_placeholder = None
self.node_info = self.policy_network.get_node_info()
else:
self.policy_network = policy_network.policy_network(
session=self.session,
name_scope=self.name_scope + '_policy',
input_size=self.observation_size,
output_size=self.action_size,
ob_placeholder=None,
trainable=True,
build_network_now=True,
define_std=True,
is_baseline=False,
args=self.args)
self.fetch_policy_info()
def build_policy_network(self, adj_matrix=None, node_attr=None):
if self.args.nervenetplus:
assert self.args.use_gnn_as_policy and self.args.use_nervenet
if self.args.use_gnn_as_policy:
if self.args.use_nervenet:
if self.args.nervenetplus:
if self.args.tree_net:
self.policy_network = treenet_policy.nervenet(
session=self.session,
name_scope=self.name_scope + '_policy',
input_size=self.observation_size,
output_size=self.action_size,
adj_matrix=adj_matrix,
node_attr=node_attr,
args=self.args,
is_rollout_agent=self.is_rollout_agent)
self.step_policy_network = treenet_policy.nervenet(
session=self.session,
name_scope=self.name_scope + 'step_policy',
input_size=self.observation_size,
output_size=self.action_size,
adj_matrix=self.adj_matrix,
node_attr=self.node_attr,
args=self.args,
is_rollout_agent=True)
self.node_info = self.policy_network.get_node_info()
self.step_policy_var_list, _ = \
self.step_policy_network.get_var_list()
self.set_step_policy = \
utils.SetPolicyWeights(self.session,
self.step_policy_var_list)
else:
self.policy_network = nervenetplus_policy.nervenet(
session=self.session,
name_scope=self.name_scope + '_policy',
input_size=self.observation_size,
output_size=self.action_size,
adj_matrix=adj_matrix,
node_attr=node_attr,
args=self.args,
is_rollout_agent=self.is_rollout_agent)
self.step_policy_network = nervenetplus_policy.nervenet(
session=self.session,
name_scope=self.name_scope + 'step_policy',
input_size=self.observation_size,
output_size=self.action_size,
adj_matrix=self.adj_matrix,
node_attr=self.node_attr,
args=self.args,
is_rollout_agent=True)
self.node_info = self.policy_network.get_node_info()
self.step_policy_var_list, _ = \
self.step_policy_network.get_var_list()
self.set_step_policy = \
utils.SetPolicyWeights(self.session,
self.step_policy_var_list)
else:
self.policy_network = nervenet_policy.nervenet(
session=self.session,
name_scope=self.name_scope + '_policy',
input_size=self.observation_size,
output_size=self.action_size,
adj_matrix=adj_matrix,
node_attr=node_attr,
args=self.args)
else:
self.policy_network = gated_graph_policy_network.GGNN(
session=self.session,
name_scope=self.name_scope + '_policy',
input_size=self.observation_size,
output_size=self.action_size,
ob_placeholder=None,
trainable=True,
build_network_now=True,
is_baseline=False,
placeholder_list=None,
args=self.args)
self.raw_obs_placeholder = None
self.node_info = self.policy_network.get_node_info()
else:
self.policy_network = policy_network.policy_network(
session=self.session,
name_scope=self.name_scope + '_policy',
input_size=self.observation_size,
output_size=self.action_size,
ob_placeholder=None,
trainable=True,
build_network_now=True,
define_std=True,
is_baseline=False,
args=self.args)
# if use the nervenetplus model
# if self.args.nervenetplus:
# # build the action model
# with tf.variable_scope('', reuse=True):
# self.step_policy_network = nervenetplus_policy.nervenet(
# session=self.session,
# name_scope=self.name_scope + '_policy',
# input_size=self.observation_size,
# output_size=self.action_size,
# adj_matrix=self.adj_matrix,
# node_attr=self.node_attr,
# args=self.args,
# is_rollout_agent=True
# )
# self.node_info = self.policy_network.get_node_info()
self.fetch_policy_info()