def offline_gym(
env_name: str,
pkl_path: str,
num_train_transitions: int,
max_steps: Optional[int],
seed: Optional[int] = None,
):
"""
Generate samples from a DiscreteRandomPolicy on the Gym environment and
saves results in a pandas df parquet.
"""
initialize_seed(seed)
env = Gym(env_name=env_name)
replay_buffer = ReplayBuffer(replay_capacity=num_train_transitions,
batch_size=1)
fill_replay_buffer(env, replay_buffer, num_train_transitions)
if isinstance(env.action_space, gym.spaces.Discrete):
is_discrete_action = True
else:
assert isinstance(env.action_space, gym.spaces.Box)
is_discrete_action = False
df = replay_buffer_to_pre_timeline_df(is_discrete_action, replay_buffer)
logger.info(f"Saving dataset with {len(df)} samples to {pkl_path}")
df.to_pickle(pkl_path)
def _offline_gym(
env: Gym,
agent: Agent,
pkl_path: str,
num_train_transitions: int,
max_steps: Optional[int],
seed: int = 1,
):
initialize_seed(seed)
replay_buffer = ReplayBuffer(replay_capacity=num_train_transitions,
batch_size=1)
fill_replay_buffer(env, replay_buffer, num_train_transitions, agent)
if isinstance(env.action_space, gym.spaces.Discrete):
is_discrete_action = True
else:
assert isinstance(env.action_space, gym.spaces.Box)
is_discrete_action = False
df = replay_buffer_to_pre_timeline_df(is_discrete_action, replay_buffer)
logger.info(f"Saving dataset with {len(df)} samples to {pkl_path}")
df.to_pickle(pkl_path)
