def _save_snapshot(self, policy: TFPolicy) -> None:
weights = policy.get_weights()
try:
self.policy_snapshots[self.snapshot_counter] = weights
except IndexError:
self.policy_snapshots.append(weights)
self.policy_elos[self.snapshot_counter] = self.current_elo
self.snapshot_counter = (self.snapshot_counter + 1) % self.window
def add_policy(self, name_behavior_id: str, policy: TFPolicy) -> None:
# for saving/swapping snapshots
policy.init_load_weights()
self.policies[name_behavior_id] = policy
# First policy encountered
if not self.learning_behavior_name:
weights = policy.get_weights()
self.current_policy_snapshot = weights
self._save_snapshot(policy)
self.trainer.add_policy(name_behavior_id, policy)
self.learning_behavior_name = name_behavior_id
def add_policy(self, name_behavior_id: str, policy: TFPolicy) -> None:
"""
Adds policy to trainer. For the first policy added, add a trainer
to the policy and set the learning behavior name to name_behavior_id.
:param name_behavior_id: Behavior ID that the policy should belong to.
:param policy: Policy to associate with name_behavior_id.
"""
self.policies[name_behavior_id] = policy
policy.create_tf_graph()
# First policy encountered
if not self.learning_behavior_name:
weights = policy.get_weights()
self.current_policy_snapshot = weights
self.trainer.add_policy(name_behavior_id, policy)
self._save_snapshot(policy) # Need to save after trainer initializes policy
self.learning_behavior_name = name_behavior_id
else:
# for saving/swapping snapshots
policy.init_load_weights()