def set_agent(self,
agent_class: Callable,
agent_config: ConfigBase,
agent_kwargs: Dict[str, any],
gpu_memory_limit: int = 512):
self.gpu_memory_limit = gpu_memory_limit
VirtualGPU(gpu_device_id=self.device_id,
gpu_memory_limit=self.gpu_memory_limit)
self.agent_class = agent_class
config = agent_config.build()
config.update(agent_kwargs)
self.agent_config = config
self.agent = agent_class(**self.agent_config)
def example(cls,
config: ConfigBase,
render: bool = True,
n_episodes: int = 500,
max_episode_steps: int = 500) -> "RandomAgent":
"""Create, train, and save agent for a given config."""
config_dict = config.build()
agent = cls(**config_dict)
agent.train(verbose=True,
render=render,
n_episodes=n_episodes,
max_episode_steps=max_episode_steps)
agent.save()
return agent
def _fit_agent(agent_class: Callable, agent_config: ConfigBase, training_options: Dict[str, Any],
gpu_memory_per_agent: int = 512) -> AgentBase:
VirtualGPU(gpu_device_id=0, gpu_memory_limit=gpu_memory_per_agent)
with warnings.catch_warnings():
warnings.simplefilter('ignore', FutureWarning)
config_dict = agent_config.build()
# Give each agent a unique name for easier tracking with verbose and multiprocessing
config_dict["name"] = f"{config_dict.get('name', 'Agent')}_{np.random.randint(0, 2 ** 16)}"
agent: AgentBase = agent_class(**config_dict)
agent.train(**training_options)
# Might as well save agent. This will also unready and save buffers, models, etc.
agent.save()
agent.unready()
return agent
def example(cls,
config: ConfigBase,
render: bool = True,
n_episodes: int = 500,
max_episode_steps: int = 500,
update_every: int = 1) -> "ReinforceAgent":
"""Create, train, and save agent for a given config."""
VirtualGPU(config.gpu_memory)
config_dict = config.build()
agent = cls(**config_dict)
agent.train(verbose=True,
render=render,
n_episodes=n_episodes,
max_episode_steps=max_episode_steps,
update_every=update_every)
agent.save()
return agent
def example(cls, config: ConfigBase, render: bool = True,
n_episodes: int = 500, max_episode_steps: int = 500, update_every: int = 10,
checkpoint_every: int = 100) -> "DeepQAgent":
"""For a given config, create new, or load existing agent. Then train and save agent."""
VirtualGPU(config.gpu_memory)
config_dict = config.build()
if os.path.exists(config_dict['name']):
agent = cls.load(config_dict['name'])
warnings.warn('Loaded existing agent.')
else:
agent = cls(**config_dict)
agent.train(verbose=True, render=render,
n_episodes=n_episodes, max_episode_steps=max_episode_steps, update_every=update_every,
checkpoint_every=checkpoint_every)
agent.save()
return agent
def example(cls, config: ConfigBase, render: bool = True,
n_episodes: int = 500, max_episode_steps: int = 500,
checkpoint_every: int = 100, **kwargs) -> "AgentBase":
"""
Default example runner.
kwargs should contain algo specific settings and will be passed to .train. eg, update_every for dqn.
"""
VirtualGPU(config.gpu_memory)
config_dict = config.build()
if os.path.exists(config_dict['name']):
agent = cls.load(config_dict['name'])
warnings.warn('Loaded existing agent.')
else:
agent = cls(**config_dict) # noqa - __init__ will be available in non abstract children
agent.train(verbose=True, render=render,
n_episodes=n_episodes, max_episode_steps=max_episode_steps,
checkpoint_every=checkpoint_every, **kwargs)
agent.save()
return agent