def play(self,
callbacks: Union[List[core.AgentCallback], core.AgentCallback,
None] = None,
num_episodes: int = 1,
max_steps_per_episode: int = 1000,
play_context: core.PlayContext = None,
default_plots: bool = None):
"""Plays num_episodes with the current policy.
Args:
callbacks: list of callbacks called during each episode play
num_episodes: number of episodes to play
max_steps_per_episode: max steps per episode
play_context: play configuration to be used. If set override all other play context arguments
default_plots: if set addes a set of default callbacks (plot.State, plot.Rewards, ...)
Returns:
play_context containg the actions taken and the rewards received during training
"""
if play_context is None:
play_context = core.PlayContext()
play_context.max_steps_per_episode = max_steps_per_episode
play_context.num_episodes = num_episodes
self._play(play_context=play_context,
callbacks=callbacks,
default_plots=default_plots)
return play_context
def play(self, play_context: core.PlayContext,
callbacks: List[core.AgentCallback]):
"""Forwarding to play_implementation overriden by the subclass.
Args:
play_context: play configuration to be used
callbacks: list of callbacks called during play.
"""
assert callbacks is not None, "callbacks not set"
assert play_context, "play_context not set"
assert self._agent_context.play is None, "play_context already set in agent_context"
play_context._reset()
play_context._validate()
self._agent_context.play = play_context
old_callbacks = self._callbacks
self._callbacks = callbacks
try:
monitor._MonitorEnv._register_backend_agent(self)
self._on_play_begin()
self.play_implementation(self._agent_context.play)
self._on_play_end()
finally:
monitor._MonitorEnv._register_backend_agent(None)
self._callbacks = old_callbacks
self._agent_context.play = None
def play(self,
callbacks: Union[List[core.AgentCallback], core.AgentCallback, None] = None,
num_episodes: int = 1,
max_steps_per_episode: int = 1000,
play_context: core.PlayContext = None,
default_plots: bool = None):
"""Plays num_episodes with the current policy.
Args:
callbacks: list of callbacks called during each episode play
num_episodes: number of episodes to play
max_steps_per_episode: max steps per episode
play_context: play configuration to be used. If set override all other play context arguments
default_plots: if set addes a set of default callbacks (plot.State, plot.Rewards, ...)
Returns:
play_context containg the actions taken and the rewards received during training
"""
assert self._backend_agent._agent_context._is_policy_trained, "No trained policy available. Call train() first."
if play_context is None:
play_context = core.PlayContext()
play_context.max_steps_per_episode = max_steps_per_episode
play_context.num_episodes = num_episodes
callbacks = self._to_callback_list(callbacks=callbacks)
callbacks = self._add_plot_callbacks(callbacks, default_plots, [plot.Steps(), plot.Rewards()])
self._backend_agent.play(play_context=play_context, callbacks=callbacks)
return play_context
def test_agentcontext_play(self):
b = easyagents.backends.debug.DebugAgentFactory()
a = b.create_agent(PpoAgent, ModelConfig(_env_name))
c = AgentContextTest.PlayCallback()
pc = PlayContext()
pc.num_episodes = 10
pc.max_steps_per_episode = 10
a.play(callbacks=[Fast(), c], play_context=pc)
assert not c.train_called
assert c.play_called