def build_replay_buffer(self, config):
super().build_replay_buffer(config)
policy = self.get_policy()
self.virtual_replay = NumpyReplayBuffer(policy.observation_space,
policy.action_space,
config["virtual_buffer_size"])
self.virtual_replay.seed(config["seed"])
def build_replay_buffer(self, config):
"""Construct replay buffer to hold samples."""
policy = self.get_policy()
self.replay = NumpyReplayBuffer(policy.observation_space,
policy.action_space,
config["buffer_size"])
self.replay.seed(config["seed"])
def build_replay_buffer(self):
super().build_replay_buffer()
self.virtual_replay = NumpyReplayBuffer(
self.observation_space,
self.action_space,
self.config["virtual_buffer_size"],
)
self.virtual_replay.seed(self.config["seed"])
def build_replay_buffer(self):
"""Construct the experience replay buffer.
Should be called by subclasses on init.
"""
self.replay = NumpyReplayBuffer(self.observation_space,
self.action_space,
self.config["buffer_size"])
self.replay.seed(self.config["seed"])
def _init(self, config, env_creator):
self._validate_config(config)
self.workers = self._make_workers(env_creator,
self._policy,
config,
num_workers=config["num_workers"])
# Dummy optimizer to log stats since Trainer.collect_metrics is coupled with it
self.optimizer = PolicyOptimizer(self.workers)
policy = self.get_policy()
policy.set_reward_from_config(config["env"], config["env_config"])
self.replay = NumpyReplayBuffer(policy.observation_space,
policy.action_space,
config["buffer_size"])
self.replay.add_fields(ReplayField(SampleBatch.ACTION_LOGP))
self.replay.seed(config["seed"])
def numpy_replay(obs_space, action_space, size, sample_batch):
replay = NumpyReplayBuffer(obs_space, action_space, size)
replay.add(sample_batch)
return replay
def replay(obs_space, action_space, samples):
replay = NumpyReplayBuffer(obs_space, action_space, size=samples.count)
replay.add(samples)
return replay