def build_model(self):
# build the environments for the agent
self.build_env()
self.build_session()
self.build_policy_network()
self.fetch_policy_info()
self.session.run(tf.global_variables_initializer())
self.set_policy = utils.SetPolicyWeights(self.session,
self.policy_var_list)
self.get_policy = utils.GetPolicyWeights(self.session,
self.policy_var_list)
def build_models(self):
'''
@brief:
this is the function where the rollout agents and optimization
agent build their networks, set up the placeholders, and gather
the variable list.
'''
# make sure that the agent has a session
self.build_session()
# set the summary writer
self.summary_writer = summary_handler.gym_summary_handler(
self.session,
self.get_experiment_name(),
enable=self.args.write_summary,
summary_dir='1115' #self.args.output_dir
)
# build the policy network and baseline network
self.build_policy_network()
# the baseline function to reduce the variance
self.build_baseline_network()
# the training op and graphs
self.build_ppo_update_op()
self.update_parameters = self.update_ppo_parameters
# init the network parameters (xavier initializer)
self.session.run(tf.global_variables_initializer())
# the set weight policy ops
self.get_policy = utils.GetPolicyWeights(self.session,
self.policy_var_list)
# prepare the feed_dict info for the ppo minibatches
self.prepare_feed_dict_map()
# prepared the init kl divergence if needed
self.current_kl_lambda = 1
def build_models(self):
'''
@brief:
this is the function where the rollout agents and optimization
agent build their networks, set up the placeholders, and gather
the variable list.
'''
# make sure that the agent has a session
self.build_session()
# the baseline function to reduce the variance
self.build_baseline_network()
# init the network parameters (xavier initializer)
self.session.run(tf.global_variables_initializer())
# prepare the feed_dict info for the ppo minibatches
self._receive_idx, self._send_idx, self._node_type_idx, \
self._inverse_node_type_idx, self._batch_size_int = \
self.baseline_network.get_gnn_idx_placeholder()
self._num_nodes_ph = self.baseline_network._num_nodes_placeholder
self._input_parameters = \
self.baseline_network.get_input_parameters_placeholder()
self._target_returns = \
self.baseline_network.get_target_return_placeholder()
self._vpred = self.baseline_network.get_vpred_placeholder()
# prepared the init kl divergence if needed
self.current_kl_lambda = 1
self.policy_var_list, self.all_policy_var_list = \
self.baseline_network.get_var_list()
self.get_policy = \
utils.GetPolicyWeights(self.session, self.policy_var_list)
self.set_policy = \
utils.SetPolicyWeights(self.session, self.policy_var_list)
self.initial_weights = self.get_policy()
def build_models(self, received_data=None):
'''
@brief:
this is the function where the rollout agents and optimization
agent build their networks, set up the placeholders, and gather
the variable list.
'''
# make sure that the agent has a session
self.start_time = time.time()
self.build_session()
self.build_policy_network(adj_matrix=self.adj_matrix,
node_attr=self.node_attr)
self.baseline_network, self.target_return_placeholder, \
self.raw_obs_placeholder = \
agent_util.build_baseline_network(
self.args, self.session, self.name_scope, self.observation_size,
self.gnn_placeholder_list, self.obs_placeholder
)
# for key in tf.trainable_variables():
# print(key.name)
# from util import fpdb; fpdb.fpdb().set_trace()
# the training op and graphs
self.build_ppo_update_op()
self.update_parameters = self.update_ppo_parameters
# init the network parameters (xavier initializer)
self.session.run(tf.global_variables_initializer())
# the set weight policy ops
self.get_policy = \
utils.GetPolicyWeights(self.session, self.policy_var_list)
self.set_policy = \
utils.SetPolicyWeights(self.session, self.policy_var_list)
if received_data is not None:
# set the running_mean and network weights here
if received_data['reset']:
# no inheriting from the parents!
logger.info('Not Inheriting from parents!')
else:
if self.args.use_gnn_as_policy:
current_species_format = self.get_species_info()
processed_species_info = species2species.process_inherited_info(
raw_species_info=received_data,
current_species_format=current_species_format,
is_nervenet=self.args.use_gnn_as_policy)
self.set_policy(processed_species_info['policy_weights'])
self.set_running_means(
processed_species_info['running_mean_info'])
self.current_lr = received_data['lr']
else:
# old species of the fc baselines
if received_data[
'SpcID'] > 0 and self.args.fc_amortized_fitness:
self.set_policy(received_data['policy_weights'])
self.set_running_means(
received_data['running_mean_info'])
# prepare the feed_dict info for the ppo minibatches
self.prepare_feed_dict_map()
# prepared the init kl divergence if needed
self.current_kl_lambda = 1