def trainer(_algorithm_cls, _parallel: bool, tmpdir: str, _convert_dataset: bool): logger.configure(tmpdir, ["tensorboard", "stdout"]) trajs = types.load( "tests/data/expert_models/cartpole_0/rollouts/final.pkl") if _convert_dataset: trans = rollout.flatten_trajectories(trajs) expert_data = datasets.TransitionsDictDatasetAdaptor(trans) else: expert_data = rollout.flatten_trajectories(trajs) venv = util.make_vec_env( "CartPole-v1", n_envs=2, parallel=_parallel, log_dir=tmpdir, ) gen_policy = util.init_rl(venv, verbose=1) return _algorithm_cls( venv=venv, expert_data=expert_data, gen_policy=gen_policy, log_dir=tmpdir, )
def test_train_disc_small_expert_data_warning(tmpdir, _algorithm_cls): logger.configure(tmpdir, ["tensorboard", "stdout"]) venv = util.make_vec_env( "CartPole-v1", n_envs=2, parallel=_parallel, ) gen_algo = util.init_rl(venv, verbose=1) small_data = rollout.generate_transitions(gen_algo, venv, n_timesteps=20) with pytest.raises(ValueError, match="Transitions.*expert_batch_size"): _algorithm_cls( venv=venv, expert_data=small_data, expert_batch_size=21, gen_algo=gen_algo, log_dir=tmpdir, ) with pytest.raises(ValueError, match="expert_batch_size.*positive"): _algorithm_cls( venv=venv, expert_data=small_data, expert_batch_size=-1, gen_algo=gen_algo, log_dir=tmpdir, )
def train_bc(task_name, batch_size, data_root, wrappers, train_epochs, n_traj, lr, policy_class, train_batches, log_interval, save_location, policy_path): # This code is designed to let you either train for a fixed number of batches, or for a fixed number of epochs assert train_epochs is None or train_batches is None, \ "Only one of train_batches or train_epochs should be set" assert not (train_batches is None and train_epochs is None), \ "You cannot have both train_epochs and train_epochs set to None" # This `get_data_pipeline_and_env` utility is designed to be shared across multiple baselines # It takes in a task name, data root, and set of wrappers and returns # (1) A "Dummy Env", i.e. an env with the same environment spaces as you'd getting from making the env associated # with this task and wrapping it in `wrappers`, but without having to actually start up Minecraft # (2) A MineRL DataPipeline that can be used to construct a batch_iter used by BC, and also as a handle to clean # up that iterator after training. data_pipeline, wrapped_dummy_env = utils.get_data_pipeline_and_env(task_name, data_root, wrappers) # This utility creates a data iterator that is basically a light wrapper around the baseline MineRL data iterator # that additionally: # (1) Applies all observation and action transformations specified by the wrappers in `wrappers`, and # (2) Calls `np.squeeze` recursively on all the nested dict spaces to remove the sequence dimension, since we're # just doing single-frame BC here data_iter = utils.create_data_iterator(wrapped_dummy_env, data_pipeline, batch_size, train_epochs, n_traj) if policy_class == SpaceFlatteningActorCriticPolicy: policy = policy_class(observation_space=wrapped_dummy_env.observation_space, action_space=wrapped_dummy_env.action_space, env=wrapped_dummy_env, lr_schedule=lambda _: 1e-4, features_extractor_class=MAGICALCNN) else: policy = policy_class(observation_space=wrapped_dummy_env.observation_space, action_space=wrapped_dummy_env.action_space, lr_schedule=lambda _: 1e-4, features_extractor_class=MAGICALCNN) run_save_location = os.path.join(save_location, str(round(time()))) os.mkdir(run_save_location) imitation_logger.configure(run_save_location, ["stdout", "tensorboard"]) bc_trainer = BC( observation_space=wrapped_dummy_env.observation_space, action_space=wrapped_dummy_env.action_space, policy_class= lambda **kwargs: policy, policy_kwargs=None, expert_data=data_iter, device='auto', optimizer_cls=th.optim.Adam, optimizer_kwargs=dict(lr=lr), ent_weight=1e-3, l2_weight=1e-5) bc_trainer.train(n_epochs=train_epochs, n_batches=train_batches, log_interval=log_interval) bc_trainer.save_policy(policy_path=os.path.join(run_save_location, policy_path)) print("Training complete; cleaning up data pipeline!") data_pipeline.close()
def test_no_accum(tmpdir): logger.configure(tmpdir, ["csv"]) sb_logger.record("A", 1) sb_logger.record("B", 1) sb_logger.dump() sb_logger.record("A", 2) sb_logger.dump() sb_logger.record("B", 3) sb_logger.dump() expect = {"A": [1, 2, ""], "B": [1, "", 3]} _compare_csv_lines(osp.join(tmpdir, "progress.csv"), expect)
def train_gail(env, n=0): venv = util.make_vec_env(env, n_envs=8) if isinstance(venv.action_space, Discrete): w = 64 else: w = 256 expert_data = make_sads_dataloader(env, max_trajs=5) logger.configure(os.path.join("learners", "GAIL")) for i in range(n): discrim_net = discrim_nets.ActObsMLP( action_space=venv.action_space, observation_space=venv.observation_space, hid_sizes=(w, w), ) gail_trainer = adversarial.GAIL( venv, expert_data=expert_data, expert_batch_size=32, gen_algo=PPO("MlpPolicy", venv, verbose=1, n_steps=1024, policy_kwargs=dict(net_arch=[w, w])), discrim_kwargs={'discrim_net': discrim_net}) mean_rewards = [] std_rewards = [] for train_steps in range(20): if train_steps > 0: if 'Bullet' in env: gail_trainer.train(total_timesteps=25000) else: gail_trainer.train(total_timesteps=16384) def get_policy(*args, **kwargs): return gail_trainer.gen_algo.policy model = PPO(get_policy, env, verbose=1) mean_reward, std_reward = evaluate_policy(model, model.env, n_eval_episodes=10) mean_rewards.append(mean_reward) std_rewards.append(std_reward) print("{0} Steps: {1}".format(train_steps, mean_reward)) np.savez(os.path.join("learners", env, "gail_rewards_{0}".format(i)), means=mean_rewards, stds=std_rewards)
def test_train_disc_small_expert_data_warning(tmpdir, _algorithm_cls): logger.configure(tmpdir, ["tensorboard", "stdout"]) venv = util.make_vec_env( "CartPole-v1", n_envs=2, parallel=_parallel, log_dir=tmpdir, ) gen_algo = util.init_rl(venv, verbose=1) small_data = rollout.generate_transitions(gen_algo, venv, n_timesteps=20) with pytest.warns(RuntimeWarning, match="discriminator batch size"): _algorithm_cls( venv=venv, expert_data=small_data, gen_algo=gen_algo, log_dir=tmpdir, )
def train_bc(env, experiment_path): """ Train GAIL on rollouts in the experiment path, save checkpoints and evaluate those checkpoints Based on code here https://github.com/HumanCompatibleAI/imitation/blob/master/src/imitation/scripts/train_adversarial.py """ rollout_file = os.path.join(experiment_path, ROLLOUTS_FILE) bc_model_directory = os.path.join(experiment_path, BC_MODEL_DIRECTORY) bc_log_directory = os.path.join(experiment_path, BC_LOG_DIRECTORY) if os.path.isdir(bc_log_directory): print("Skipping BC training (log directory exists)") return os.makedirs(bc_model_directory, exist_ok=True) os.makedirs(bc_log_directory, exist_ok=True) logger.configure(bc_log_directory) expert_trajs = types.load(rollout_file) expert_transitions = rollout.flatten_trajectories(expert_trajs) env = gym.make(env) trainer = BC(env.observation_space, env.action_space, expert_data=expert_transitions, policy_class=MlpPolicy, device="cpu", ent_weight=0.0) env.close() def callback(locals): path = os.path.join(bc_model_directory, "epoch_{}".format(locals["epoch_num"])) trainer.save_policy(path) trainer.save_policy(os.path.join(experiment_path, "start_bc")) trainer.train(BC_TRAIN_EPOCHS, on_epoch_end=callback) # Save trained policy trainer.save_policy(os.path.join(experiment_path, "final_bc"))
def trainer( _algorithm_cls, _parallel: bool, tmpdir: str, _convert_dataset: bool, expert_batch_size: int, expert_transitions: types.Transitions, ): logger.configure(tmpdir, ["tensorboard", "stdout"]) if _convert_dataset: expert_data = th_data.DataLoader( expert_transitions, batch_size=expert_batch_size, collate_fn=types.transitions_collate_fn, shuffle=True, drop_last=True, ) else: expert_data = expert_transitions venv = util.make_vec_env( "CartPole-v1", n_envs=2, parallel=_parallel, log_dir=tmpdir, ) gen_algo = util.init_rl(venv, verbose=1) trainer = _algorithm_cls( venv=venv, expert_data=expert_data, expert_batch_size=expert_batch_size, gen_algo=gen_algo, log_dir=tmpdir, ) try: yield trainer finally: venv.close()
def setup_logging( _run, log_format_strs: Sequence[str], ) -> Tuple[imit_logger.HierarchicalLogger, str]: """Builds the imitation logger. Args: log_format_strs: The types of formats to log to. Returns: The configured imitation logger and `log_dir`. Returning `log_dir` avoids the caller needing to capture this value. """ log_dir = make_log_dir() custom_logger = imit_logger.configure(os.path.join(log_dir, "log"), log_format_strs) return custom_logger, log_dir
def init_trainer( env_name: str, expert_trajectories: Sequence[types.Trajectory], *, log_dir: str, seed: int = 0, use_gail: bool = False, num_vec: int = 8, parallel: bool = False, max_episode_steps: Optional[int] = None, scale: bool = True, airl_entropy_weight: float = 1.0, discrim_kwargs: dict = {}, reward_kwargs: dict = {}, trainer_kwargs: dict = {}, init_rl_kwargs: dict = {}, ): """Builds an AdversarialTrainer, ready to be trained on expert demonstrations. Args: env_name: The string id of a gym environment. expert_trajectories: Demonstrations from expert. seed: Random seed. log_dir: Directory for logging output. Will generate a unique sub-directory within this directory for all output. use_gail: If True, then train using GAIL. If False, then train using AIRL. num_vec: The number of vectorized environments. parallel: If True, then use SubprocVecEnv; otherwise, DummyVecEnv. max_episode_steps: If specified, wraps VecEnv in TimeLimit wrapper with this episode length before returning. policy_dir: The directory containing the pickled experts for generating rollouts. scale: If True, then scale input Tensors to the interval [0, 1]. airl_entropy_weight: Only applicable for AIRL. The `entropy_weight` argument of `DiscrimNetAIRL.__init__`. trainer_kwargs: Arguments for the Trainer constructor. reward_kwargs: Arguments for the `*RewardNet` constructor. discrim_kwargs: Arguments for the `DiscrimNet*` constructor. init_rl_kwargs: Keyword arguments passed to `init_rl`, used to initialize the RL algorithm. """ logger.configure(folder=log_dir, format_strs=["tensorboard", "stdout"]) env = util.make_vec_env( env_name, num_vec, seed=seed, parallel=parallel, log_dir=log_dir, max_episode_steps=max_episode_steps, ) gen_policy = util.init_rl(env, verbose=1, **init_rl_kwargs) if use_gail: discrim = discrim_net.DiscrimNetGAIL(env.observation_space, env.action_space, scale=scale, **discrim_kwargs) else: rn = BasicShapedRewardNet(env.observation_space, env.action_space, scale=scale, **reward_kwargs) discrim = discrim_net.DiscrimNetAIRL( rn, entropy_weight=airl_entropy_weight, **discrim_kwargs) expert_demos = rollout.flatten_trajectories(expert_trajectories) trainer = AdversarialTrainer(env, gen_policy, discrim, expert_demos, log_dir=log_dir, **trainer_kwargs) return trainer
def rollouts_and_policy( _run, _seed: int, env_name: str, total_timesteps: int, *, log_dir: str, num_vec: int, parallel: bool, max_episode_steps: Optional[int], normalize: bool, normalize_kwargs: dict, init_rl_kwargs: dict, n_episodes_eval: int, reward_type: Optional[str], reward_path: Optional[str], rollout_save_interval: int, rollout_save_final: bool, rollout_save_n_timesteps: Optional[int], rollout_save_n_episodes: Optional[int], policy_save_interval: int, policy_save_final: bool, init_tensorboard: bool, ) -> dict: """Trains an expert policy from scratch and saves the rollouts and policy. Checkpoints: At applicable training steps `step` (where step is either an integer or "final"): - Policies are saved to `{log_dir}/policies/{step}/`. - Rollouts are saved to `{log_dir}/rollouts/{step}.pkl`. Args: env_name: The gym.Env name. Loaded as VecEnv. total_timesteps: Number of training timesteps in `model.learn()`. log_dir: The root directory to save metrics and checkpoints to. num_vec: Number of environments in VecEnv. parallel: If True, then use DummyVecEnv. Otherwise use SubprocVecEnv. max_episode_steps: If not None, then environments are wrapped by TimeLimit so that they have at most `max_episode_steps` steps per episode. normalize: If True, then rescale observations and reward. normalize_kwargs: kwargs for `VecNormalize`. init_rl_kwargs: kwargs for `init_rl`. n_episodes_eval: The number of episodes to average over when calculating the average ground truth reward return of the final policy. reward_type: If provided, then load the serialized reward of this type, wrapping the environment in this reward. This is useful to test whether a reward model transfers. For more information, see `imitation.rewards.serialize.load_reward`. reward_path: A specifier, such as a path to a file on disk, used by reward_type to load the reward model. For more information, see `imitation.rewards.serialize.load_reward`. rollout_save_interval: The number of training updates in between intermediate rollout saves. If the argument is nonpositive, then don't save intermediate updates. rollout_save_final: If True, then save rollouts right after training is finished. rollout_save_n_timesteps: The minimum number of timesteps saved in every file. Could be more than `rollout_save_n_timesteps` because trajectories are saved by episode rather than by transition. Must set exactly one of `rollout_save_n_timesteps` and `rollout_save_n_episodes`. rollout_save_n_episodes: The number of episodes saved in every file. Must set exactly one of `rollout_save_n_timesteps` and `rollout_save_n_episodes`. policy_save_interval: The number of training updates between saves. Has the same semantics are `rollout_save_interval`. policy_save_final: If True, then save the policy right after training is finished. init_tensorboard: If True, then write tensorboard logs to {log_dir}/sb_tb and "output/summary/...". Returns: The return value of `rollout_stats()` using the final policy. """ os.makedirs(log_dir, exist_ok=True) sacred_util.build_sacred_symlink(log_dir, _run) sample_until = rollout.make_sample_until(rollout_save_n_timesteps, rollout_save_n_episodes) eval_sample_until = rollout.min_episodes(n_episodes_eval) with networks.make_session(): tf.logging.set_verbosity(tf.logging.INFO) logger.configure(folder=osp.join(log_dir, "rl"), format_strs=["tensorboard", "stdout"]) rollout_dir = osp.join(log_dir, "rollouts") policy_dir = osp.join(log_dir, "policies") os.makedirs(rollout_dir, exist_ok=True) os.makedirs(policy_dir, exist_ok=True) if init_tensorboard: sb_tensorboard_dir = osp.join(log_dir, "sb_tb") # Convert sacred's ReadOnlyDict to dict so we can modify on next line. init_rl_kwargs = dict(init_rl_kwargs) init_rl_kwargs["tensorboard_log"] = sb_tensorboard_dir venv = util.make_vec_env( env_name, num_vec, seed=_seed, parallel=parallel, log_dir=log_dir, max_episode_steps=max_episode_steps, ) log_callbacks = [] with contextlib.ExitStack() as stack: if reward_type is not None: reward_fn_ctx = load_reward(reward_type, reward_path, venv) reward_fn = stack.enter_context(reward_fn_ctx) venv = RewardVecEnvWrapper(venv, reward_fn) log_callbacks.append(venv.log_callback) tf.logging.info( f"Wrapped env in reward {reward_type} from {reward_path}.") vec_normalize = None if normalize: venv = vec_normalize = VecNormalize(venv, **normalize_kwargs) policy = util.init_rl(venv, verbose=1, **init_rl_kwargs) # Make callback to save intermediate artifacts during training. step = 0 def callback(locals_: dict, _) -> bool: nonlocal step step += 1 policy = locals_["self"] # TODO(adam): make logging frequency configurable for callback in log_callbacks: callback(sb_logger) if rollout_save_interval > 0 and step % rollout_save_interval == 0: save_path = osp.join(rollout_dir, f"{step}.pkl") rollout.rollout_and_save(save_path, policy, venv, sample_until) if policy_save_interval > 0 and step % policy_save_interval == 0: output_dir = os.path.join(policy_dir, f"{step:05d}") serialize.save_stable_model(output_dir, policy, vec_normalize) policy.learn(total_timesteps, callback=callback) # Save final artifacts after training is complete. if rollout_save_final: save_path = osp.join(rollout_dir, "final.pkl") rollout.rollout_and_save(save_path, policy, venv, sample_until) if policy_save_final: output_dir = os.path.join(policy_dir, "final") serialize.save_stable_model(output_dir, policy, vec_normalize) # Final evaluation of expert policy. trajs = rollout.generate_trajectories(policy, venv, eval_sample_until) stats = rollout.rollout_stats(trajs) return stats
def test_hard(tmpdir): logger.configure(tmpdir) # Part One: Test logging outside of the accumulating scope, and within scopes # with two different different logging keys (including a repeat). sb_logger.record("no_context", 1) with logger.accumulate_means("disc"): sb_logger.record("C", 2) sb_logger.record("D", 2) sb_logger.dump() sb_logger.record("C", 4) sb_logger.dump() with logger.accumulate_means("gen"): sb_logger.record("E", 2) sb_logger.dump() sb_logger.record("E", 0) sb_logger.dump() with logger.accumulate_means("disc"): sb_logger.record("C", 3) sb_logger.dump() sb_logger.dump() # Writes 1 mean each from "gen" and "disc". expect_raw_gen = {"raw/gen/E": [2, 0]} expect_raw_disc = { "raw/disc/C": [2, 4, 3], "raw/disc/D": [2, "", ""], } expect_default = { "mean/gen/E": [1], "mean/disc/C": [3], "mean/disc/D": [2], "no_context": [1], } _compare_csv_lines(osp.join(tmpdir, "progress.csv"), expect_default) _compare_csv_lines(osp.join(tmpdir, "raw", "gen", "progress.csv"), expect_raw_gen) _compare_csv_lines(osp.join(tmpdir, "raw", "disc", "progress.csv"), expect_raw_disc) # Part Two: # Check that we append to the same logs after the first dump to "means/*". with logger.accumulate_means("disc"): sb_logger.record("D", 100) sb_logger.dump() sb_logger.record("no_context", 2) sb_logger.dump() # Writes 1 mean from "disc". "gen" is blank. expect_raw_gen = {"raw/gen/E": [2, 0]} expect_raw_disc = { "raw/disc/C": [2, 4, 3, ""], "raw/disc/D": [2, "", "", 100], } expect_default = { "mean/gen/E": [1, ""], "mean/disc/C": [3, ""], "mean/disc/D": [2, 100], "no_context": [1, 2], } _compare_csv_lines(osp.join(tmpdir, "progress.csv"), expect_default) _compare_csv_lines(osp.join(tmpdir, "raw", "gen", "progress.csv"), expect_raw_gen) _compare_csv_lines(osp.join(tmpdir, "raw", "disc", "progress.csv"), expect_raw_disc)
# Convert List[types.Trajectory] to an instance of `imitation.data.types.Transitions`. # This is a more general dataclass containing unordered # (observation, actions, next_observation) transitions. transitions = rollout.flatten_trajectories(trajectories) venv = util.make_vec_env("CartPole-v1", n_envs=2) tempdir = tempfile.TemporaryDirectory(prefix="quickstart") tempdir_path = pathlib.Path(tempdir.name) print( f"All Tensorboards and logging are being written inside {tempdir_path}/.") # Train BC on expert data. # BC also accepts as `expert_data` any PyTorch-style DataLoader that iterates over # dictionaries containing observations and actions. logger.configure(tempdir_path / "BC/") bc_trainer = bc.BC(venv.observation_space, venv.action_space, expert_data=transitions) bc_trainer.train(n_epochs=1) # Train GAIL on expert data. # GAIL, and AIRL also accept as `expert_data` any Pytorch-style DataLoader that # iterates over dictionaries containing observations, actions, and next_observations. logger.configure(tempdir_path / "GAIL/") gail_trainer = adversarial.GAIL( venv, expert_data=transitions, expert_batch_size=32, gen_algo=sb3.PPO("MlpPolicy", venv, verbose=1, n_steps=1024), )
def imitation_learning(expert_traj_path, imitation_algo_name, rl_algo_name, env_name): # Load pickled expert demonstrations. with open(expert_traj_path, "rb") as f: # This is a list of `imitation.data.types.Trajectory`, where # every instance contains observations and actions for a single expert # demonstration. trajectories = pickle.load(f) # Convert List[types.Trajectory] to an instance of `imitation.data.types.Transitions`. # This is a more general dataclass containing unordered # (observation, actions, next_observation) transitions. transitions = rollout.flatten_trajectories(trajectories) venv = util.make_vec_env(env_name, n_envs=2) # tempdir = tempfile.TemporaryDirectory(prefix="il_results/{}_{}".format(rl_algo_name, env_name)) # tempdir_path = pathlib.Path(tempdir.name) # print(f"All Tensorboards and logging are being written inside {tempdir_path}/.") log_path = "il_results/{}_{}/{}/".format(rl_algo_name, env_name, imitation_algo_name) if imitation_algo_name == 'BC': # Train BC on expert data. # BC also accepts as `expert_data` any PyTorch-style DataLoader that iterates over # dictionaries containing observations and actions. logger.configure(log_path, format_strs=["stdout", "tensorboard"]) trainer = bc.BC(venv.observation_space, venv.action_space, expert_data=transitions) trainer.train(n_epochs=100, log_interval=1) elif imitation_algo_name == 'GAIL': logger.configure(log_path, format_strs=["stdout", "tensorboard"]) gail_trainer = adversarial.GAIL( venv, expert_data=transitions, expert_batch_size=32, gen_algo=sb3.PPO("MlpPolicy", venv, verbose=1, n_steps=1024), discrim_kwargs={ 'discrim_net': ActObsMLP( action_space=venv.action_space, observation_space=venv.observation_space, hid_sizes=(32, 32), ) }) gail_trainer.train(total_timesteps=2048) trainer = gail_trainer.gen_algo elif imitation_algo_name == 'AIRL': # Train AIRL on expert data. logger.configure(log_path) airl_trainer = adversarial.AIRL( venv, expert_data=transitions, expert_batch_size=32, gen_algo=sb3.PPO("MlpPolicy", venv, verbose=1, n_steps=1024), ) airl_trainer.train(total_timesteps=2048) sample_until = rollout.min_episodes(15) trained_ret_mean = rollout.mean_return(trainer.policy, venv, sample_until) # trainer.save_policy("{}/bc_policy.pth.tar".format(log_path)) th.save(trainer.policy, "{}/{}_policy.pth.tar".format(log_path, imitation_algo_name)) return trained_ret_mean
def train( _run, _seed: int, algorithm: str, env_name: str, num_vec: int, parallel: bool, max_episode_steps: Optional[int], rollout_path: str, n_expert_demos: Optional[int], log_dir: str, total_timesteps: int, n_episodes_eval: int, init_tensorboard: bool, checkpoint_interval: int, gen_batch_size: int, init_rl_kwargs: Mapping, algorithm_kwargs: Mapping[str, Mapping], discrim_net_kwargs: Mapping[str, Mapping], ) -> dict: """Train an adversarial-network-based imitation learning algorithm. Checkpoints: - DiscrimNets are saved to `f"{log_dir}/checkpoints/{step}/discrim/"`, where step is either the training round or "final". - Generator policies are saved to `f"{log_dir}/checkpoints/{step}/gen_policy/"`. Args: _seed: Random seed. algorithm: A case-insensitive string determining which adversarial imitation learning algorithm is executed. Either "airl" or "gail". env_name: The environment to train in. num_vec: Number of `gym.Env` to vectorize. parallel: Whether to use "true" parallelism. If True, then use `SubProcVecEnv`. Otherwise, use `DummyVecEnv` which steps through environments serially. max_episode_steps: If not None, then a TimeLimit wrapper is applied to each environment to artificially limit the maximum number of timesteps in an episode. rollout_path: Path to pickle containing list of Trajectories. Used as expert demonstrations. n_expert_demos: The number of expert trajectories to actually use after loading them from `rollout_path`. If None, then use all available trajectories. If `n_expert_demos` is an `int`, then use exactly `n_expert_demos` trajectories, erroring if there aren't enough trajectories. If there are surplus trajectories, then use the first `n_expert_demos` trajectories and drop the rest. log_dir: Directory to save models and other logging to. total_timesteps: The number of transitions to sample from the environment during training. n_episodes_eval: The number of episodes to average over when calculating the average episode reward of the imitation policy for return. init_tensorboard: If True, then write tensorboard logs to `{log_dir}/sb_tb`. checkpoint_interval: Save the discriminator and generator models every `checkpoint_interval` rounds and after training is complete. If 0, then only save weights after training is complete. If <0, then don't save weights at all. gen_batch_size: Batch size for generator updates. Sacred automatically uses this to calculate `n_steps` in `init_rl_kwargs`. In the script body, this is only used in sanity checks. init_rl_kwargs: Keyword arguments for `init_rl`, the RL algorithm initialization utility function. algorithm_kwargs: Keyword arguments for the `GAIL` or `AIRL` constructor that can apply to either constructor. Unlike a regular kwargs argument, this argument can only have the following keys: "shared", "airl", and "gail". `algorithm_kwargs["airl"]`, if it is provided, is a kwargs `Mapping` passed to the `AIRL` constructor when `algorithm == "airl"`. Likewise `algorithm_kwargs["gail"]` is passed to the `GAIL` constructor when `algorithm == "gail"`. `algorithm_kwargs["shared"]`, if provided, is passed to both the `AIRL` and `GAIL` constructors. Duplicate keyword argument keys between `algorithm_kwargs["shared"]` and `algorithm_kwargs["airl"]` (or "gail") leads to an error. discrim_net_kwargs: Keyword arguments for the `DiscrimNet` constructor. Unlike a regular kwargs argument, this argument can only have the following keys: "shared", "airl", "gail". These keys have the same meaning as they do in `algorithm_kwargs`. Returns: A dictionary with two keys. "imit_stats" gives the return value of `rollout_stats()` on rollouts test-reward-wrapped environment, using the final policy (remember that the ground-truth reward can be recovered from the "monitor_return" key). "expert_stats" gives the return value of `rollout_stats()` on the expert demonstrations loaded from `rollout_path`. """ if gen_batch_size % num_vec != 0: raise ValueError( f"num_vec={num_vec} must evenly divide gen_batch_size={gen_batch_size}." ) allowed_keys = {"shared", "gail", "airl"} if not discrim_net_kwargs.keys() <= allowed_keys: raise ValueError( f"Invalid discrim_net_kwargs.keys()={discrim_net_kwargs.keys()}. " f"Allowed keys: {allowed_keys}" ) if not algorithm_kwargs.keys() <= allowed_keys: raise ValueError( f"Invalid discrim_net_kwargs.keys()={algorithm_kwargs.keys()}. " f"Allowed keys: {allowed_keys}" ) if not os.path.exists(rollout_path): raise ValueError(f"File at rollout_path={rollout_path} does not exist.") expert_trajs = types.load(rollout_path) if n_expert_demos is not None: if not len(expert_trajs) >= n_expert_demos: raise ValueError( f"Want to use n_expert_demos={n_expert_demos} trajectories, but only " f"{len(expert_trajs)} are available via {rollout_path}." ) expert_trajs = expert_trajs[:n_expert_demos] expert_transitions = rollout.flatten_trajectories(expert_trajs) total_timesteps = int(total_timesteps) logging.info("Logging to %s", log_dir) logger.configure(log_dir, ["tensorboard", "stdout"]) os.makedirs(log_dir, exist_ok=True) sacred_util.build_sacred_symlink(log_dir, _run) venv = util.make_vec_env( env_name, num_vec, seed=_seed, parallel=parallel, log_dir=log_dir, max_episode_steps=max_episode_steps, ) # if init_tensorboard: # tensorboard_log = osp.join(log_dir, "sb_tb") # else: # tensorboard_log = None gen_algo = util.init_rl( # FIXME(sam): ignoring tensorboard_log is a hack to prevent SB3 from # re-configuring the logger (SB3 issue #109). See init_rl() for details. # TODO(shwang): Let's get rid of init_rl after SB3 issue #109 is fixed? # Besides sidestepping #109, init_rl is just a stub function. venv, **init_rl_kwargs, ) discrim_kwargs_shared = discrim_net_kwargs.get("shared", {}) discrim_kwargs_algo = discrim_net_kwargs.get(algorithm, {}) final_discrim_kwargs = dict(**discrim_kwargs_shared, **discrim_kwargs_algo) algorithm_kwargs_shared = algorithm_kwargs.get("shared", {}) algorithm_kwargs_algo = algorithm_kwargs.get(algorithm, {}) final_algorithm_kwargs = dict( **algorithm_kwargs_shared, **algorithm_kwargs_algo, ) if algorithm.lower() == "gail": algo_cls = adversarial.GAIL elif algorithm.lower() == "airl": algo_cls = adversarial.AIRL else: raise ValueError(f"Invalid value algorithm={algorithm}.") trainer = algo_cls( venv=venv, expert_data=expert_transitions, gen_algo=gen_algo, log_dir=log_dir, discrim_kwargs=final_discrim_kwargs, **final_algorithm_kwargs, ) def callback(round_num): if checkpoint_interval > 0 and round_num % checkpoint_interval == 0: save(trainer, os.path.join(log_dir, "checkpoints", f"{round_num:05d}")) trainer.train(total_timesteps, callback) # Save final artifacts. if checkpoint_interval >= 0: save(trainer, os.path.join(log_dir, "checkpoints", "final")) # Final evaluation of imitation policy. results = {} sample_until_eval = rollout.min_episodes(n_episodes_eval) trajs = rollout.generate_trajectories( trainer.gen_algo, trainer.venv_train_norm, sample_until=sample_until_eval ) results["expert_stats"] = rollout.rollout_stats(expert_trajs) results["imit_stats"] = rollout.rollout_stats(trajs) return results