def save(self, directory: Optional[str] = None,
callbacks: Union[List[core.AgentCallback], core.AgentCallback, None] = None) -> str:
"""Saves the currently trained actor policy in directory.
If save is called before a trained policy is created, eg by calling train, an exception is raised.
Args:
directory: the directory to save the policy weights to.
if the directory does not exist yet, a new directory is created. if None the policy is saved
in a temp directory.
callbacks: list of callbacks called during save (eg log.Agent)
Returns:
the absolute path to the directory containing the saved policy.
"""
if directory is None:
directory = bcore._get_temp_path()
assert directory
assert self._backend_agent._agent_context._is_policy_trained, "No trained policy available. Call train() first."
directory = bcore._mkdir(directory)
agent_json_path = os.path.join(directory, EasyAgent._KEY_EASYAGENT_FILENAME)
with open(agent_json_path, 'w') as jsonfile:
agent_dict = self._to_dict()
json.dump(agent_dict, jsonfile, sort_keys=True, indent=2)
callbacks = self._to_callback_list(callbacks=callbacks)
policy_directory = os.path.join(directory, EasyAgent._KEY_POLICY_DIRECTORY)
policy_directory = bcore._mkdir(policy_directory)
self._backend_agent.save(directory=policy_directory, callbacks=callbacks)
return directory
def test_save_load(self):
model_config = core.ModelConfig(_lineworld_name)
random_agent = tfagents.TfRandomAgent(model_config=model_config)
tempdir = bcore._get_temp_path()
bcore._mkdir(tempdir)
random_agent.save(directory=tempdir, callbacks=[])
random_agent.load(directory=tempdir, callbacks=[])
bcore._rmpath(tempdir)
def test_save_(self):
from easyagents.backends import tforce
model_config = core.ModelConfig(_cartpole_name)
tc = core.PpoTrainContext()
tc.num_iterations = 3
ppo_agent = tforce.TforcePpoAgent(model_config=model_config)
ppo_agent.train(train_context=tc, callbacks=[log.Iteration(), log.Agent()])
tempdir = bcore._get_temp_path()
bcore._mkdir(tempdir)
ppo_agent.save(tempdir, [])
loaded_agent = tforce.TforcePpoAgent(model_config=model_config)
loaded_agent.load(tempdir, [])
bcore._rmpath(tempdir)
def test_save_load(self):
model_config = core.ModelConfig(_lineworld_name)
tc = core.PpoTrainContext()
ppo_agent = tfagents.TfPpoAgent(model_config=model_config)
ppo_agent.train(
train_context=tc,
callbacks=[duration._SingleIteration(),
log.Iteration()])
tempdir = bcore._get_temp_path()
bcore._mkdir(tempdir)
ppo_agent.save(tempdir, [])
ppo_agent = tfagents.TfPpoAgent(model_config=model_config)
ppo_agent.load(tempdir, [])
pc = core.PlayContext()
pc.max_steps_per_episode = 10
pc.num_episodes = 1
ppo_agent.play(play_context=pc, callbacks=[])
bcore._rmpath(tempdir)
def __init__(self, directory: str = None):
"""Saves the best policies (along with the agent definition) in directory.
If directory is None the policies are written in a temp directory.
Args:
directory: the directory to save to, if None a temp directory is created.
"""
directory = directory if directory else bcore._get_temp_path()
self.directory: str = bcore._mkdir(directory)
self.saved_agents: List[Tuple[int, float, str]] = []