ding.entry.serial_entry_preference_based_irl_onpolicy

ding.entry.serial_entry_preference_based_irl_onpolicy

serial_pipeline_preference_based_irl_onpolicy(input_cfg, seed=0, env_setting=None, model=None, max_train_iter=int(10000000000.0), max_env_step=int(10000000000.0))

Overview

Serial pipeline entry for preference based irl of on-policy algorithm(such as PPO).

Arguments: - input_cfg (:obj:Union[str, Tuple[dict, dict]]): Config in dict type. str type means config file path. Tuple[dict, dict] type means [user_config, create_cfg]. - seed (:obj:int): Random seed. - env_setting (:obj:Optional[List[Any]]): A list with 3 elements: BaseEnv subclass, collector env config, and evaluator env config. - model (:obj:Optional[torch.nn.Module]): Instance of torch.nn.Module. - max_train_iter (:obj:Optional[int]): Maximum policy update iterations in training. - max_env_step (:obj:Optional[int]): Maximum collected environment interaction steps. Returns: - policy (:obj:Policy): Converged policy.

Full Source Code

../ding/entry/serial_entry_preference_based_irl_onpolicy.py

1from typing import Union, Optional, List, Any, Tuple 2import os 3import torch 4from ditk import logging 5from functools import partial 6from tensorboardX import SummaryWriter 7from copy import deepcopy 8 9from ding.envs import get_vec_env_setting, create_env_manager 10from ding.worker import BaseLearner, InteractionSerialEvaluator, BaseSerialCommander, create_buffer, \ 11 create_serial_collector 12from ding.config import read_config, compile_config 13from ding.policy import create_policy, PolicyFactory 14from ding.reward_model import create_reward_model 15from ding.utils import set_pkg_seed 16 17 18def serial_pipeline_preference_based_irl_onpolicy( 19 input_cfg: Union[str, Tuple[dict, dict]], 20 seed: int = 0, 21 env_setting: Optional[List[Any]] = None, 22 model: Optional[torch.nn.Module] = None, 23 max_train_iter: Optional[int] = int(1e10), 24 max_env_step: Optional[int] = int(1e10), 25) -> 'Policy': # noqa 26 """ 27 Overview: 28 Serial pipeline entry for preference based irl of on-policy algorithm(such as PPO). 29 Arguments: 30 - input_cfg (:obj:`Union[str, Tuple[dict, dict]]`): Config in dict type. \ 31 ``str`` type means config file path. \ 32 ``Tuple[dict, dict]`` type means [user_config, create_cfg]. 33 - seed (:obj:`int`): Random seed. 34 - env_setting (:obj:`Optional[List[Any]]`): A list with 3 elements: \ 35 ``BaseEnv`` subclass, collector env config, and evaluator env config. 36 - model (:obj:`Optional[torch.nn.Module]`): Instance of torch.nn.Module. 37 - max_train_iter (:obj:`Optional[int]`): Maximum policy update iterations in training. 38 - max_env_step (:obj:`Optional[int]`): Maximum collected environment interaction steps. 39 Returns: 40 - policy (:obj:`Policy`): Converged policy. 41 """ 42 if isinstance(input_cfg, str): 43 cfg, create_cfg = read_config(input_cfg) 44 else: 45 cfg, create_cfg = deepcopy(input_cfg) 46 create_cfg.policy.type = create_cfg.policy.type + '_command' 47 create_cfg.reward_model = dict(type=cfg.reward_model.type) 48 env_fn = None if env_setting is None else env_setting[0] 49 cfg = compile_config(cfg, seed=seed, env=env_fn, auto=True, create_cfg=create_cfg, save_cfg=True, renew_dir=False) 50 # Create main components: env, policy 51 if env_setting is None: 52 env_fn, collector_env_cfg, evaluator_env_cfg = get_vec_env_setting(cfg.env) 53 else: 54 env_fn, collector_env_cfg, evaluator_env_cfg = env_setting 55 collector_env = create_env_manager(cfg.env.manager, [partial(env_fn, cfg=c) for c in collector_env_cfg]) 56 evaluator_env = create_env_manager(cfg.env.manager, [partial(env_fn, cfg=c) for c in evaluator_env_cfg]) 57 collector_env.seed(cfg.seed) 58 evaluator_env.seed(cfg.seed, dynamic_seed=False) 59 set_pkg_seed(cfg.seed, use_cuda=cfg.policy.cuda) 60 policy = create_policy(cfg.policy, model=model, enable_field=['learn', 'collect', 'eval', 'command']) 61 62 # Create worker components: learner, collector, evaluator, replay buffer, commander. 63 tb_logger = SummaryWriter(os.path.join('./{}/log/'.format(cfg.exp_name), 'serial')) 64 learner = BaseLearner(cfg.policy.learn.learner, policy.learn_mode, tb_logger, exp_name=cfg.exp_name) 65 collector = create_serial_collector( 66 cfg.policy.collect.collector, 67 env=collector_env, 68 policy=policy.collect_mode, 69 tb_logger=tb_logger, 70 exp_name=cfg.exp_name 71 ) 72 evaluator = InteractionSerialEvaluator( 73 cfg.policy.eval.evaluator, evaluator_env, policy.eval_mode, tb_logger, exp_name=cfg.exp_name 74 ) 75 commander = BaseSerialCommander( 76 cfg.policy.other.commander, learner, collector, evaluator, None, policy.command_mode 77 ) 78 reward_model = create_reward_model(cfg, policy.collect_mode.get_attribute('device'), tb_logger) 79 reward_model.train() 80 # ========== 81 # Main loop 82 # ========== 83 # Learner's before_run hook. 84 learner.call_hook('before_run') 85 86 while True: 87 collect_kwargs = commander.step() 88 # Evaluate policy performance 89 if evaluator.should_eval(learner.train_iter): 90 stop, reward = evaluator.eval(learner.save_checkpoint, learner.train_iter, collector.envstep) 91 if stop: 92 break 93 # Collect data by default config n_sample/n_episode 94 new_data = collector.collect(train_iter=learner.train_iter) 95 train_data = new_data 96 # update train_data reward using the augmented reward 97 train_data_augmented = reward_model.estimate(train_data) 98 learner.train(train_data_augmented, collector.envstep) 99 if collector.envstep >= max_env_step or learner.train_iter >= max_train_iter: 100 break 101 102 # Learner's after_run hook. 103 learner.call_hook('after_run') 104 return policy