def __init__(self,
policy,
env,
n_envs,
replay_pool_size,
max_path_length,
sampling_method,
save_rollouts=False,
save_rollouts_observations=True,
save_env_infos=False,
env_str=None,
replay_pool_params={}):
self._policy = policy
self._n_envs = n_envs
assert (self._n_envs == 1) # b/c policy reset
self._replay_pools = [
RNNCriticReplayPool(
env.spec,
env.horizon,
policy.N,
policy.gamma,
replay_pool_size // n_envs,
obs_history_len=policy.obs_history_len,
sampling_method=sampling_method,
save_rollouts=save_rollouts,
save_rollouts_observations=save_rollouts_observations,
save_env_infos=save_env_infos,
replay_pool_params=replay_pool_params) for _ in range(n_envs)
]
try:
envs = [
pickle.loads(pickle.dumps(env)) for _ in range(self._n_envs)
] if self._n_envs > 1 else [env]
except:
envs = [create_env(env_str) for _ in range(self._n_envs)
] if self._n_envs > 1 else [env]
### need to seed each environment if it is GymEnv
seed = get_seed()
if seed is not None and isinstance(utils.inner_env(env), GymEnv):
for i, env in enumerate(envs):
utils.inner_env(env).env.seed(seed + i)
self._vec_env = VecEnvExecutor(envs=envs,
max_path_length=max_path_length)
self._curr_observations = self._vec_env.reset()
def statistics(self):
return RNNCriticReplayPool.statistics_pools(self._replay_pools)
def get_recent_paths(self):
return RNNCriticReplayPool.get_recent_paths_pools(self._replay_pools)
def log(self, prefix=''):
RNNCriticReplayPool.log_pools(self._replay_pools, self._nb_target_lost, prefix=prefix)
def sample(self, batch_size):
return RNNCriticReplayPool.sample_pools(self._replay_pools, batch_size,
only_completed_episodes=self._policy.only_completed_episodes)
def log(self, prefix=''):
RNNCriticReplayPool.log_pools(self._replay_pools, prefix=prefix)