DI-engine/ding/entry/serial_entry_preference_based_irl_onpolicy.py

from typing import Union, Optional, List, Any, Tuple
import os
import torch
from ditk import logging
from functools import partial
from tensorboardX import SummaryWriter
from copy import deepcopy

from ding.envs import get_vec_env_setting, create_env_manager
from ding.worker import BaseLearner, InteractionSerialEvaluator, BaseSerialCommander, create_buffer, \
    create_serial_collector
from ding.config import read_config, compile_config
from ding.policy import create_policy, PolicyFactory
from ding.reward_model import create_reward_model
from ding.utils import set_pkg_seed


def serial_pipeline_preference_based_irl_onpolicy(
        input_cfg: Union[str, Tuple[dict, dict]],
        seed: int = 0,
        env_setting: Optional[List[Any]] = None,
        model: Optional[torch.nn.Module] = None,
        max_train_iter: Optional[int] = int(1e10),
        max_env_step: Optional[int] = int(1e10),
) -> 'Policy':  # noqa
    """
    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.
    """
    if isinstance(input_cfg, str):
        cfg, create_cfg = read_config(input_cfg)
    else:
        cfg, create_cfg = deepcopy(input_cfg)
    create_cfg.policy.type = create_cfg.policy.type + '_command'
    create_cfg.reward_model = dict(type=cfg.reward_model.type)
    env_fn = None if env_setting is None else env_setting[0]
    cfg = compile_config(cfg, seed=seed, env=env_fn, auto=True, create_cfg=create_cfg, save_cfg=True, renew_dir=False)
    # Create main components: env, policy
    if env_setting is None:
        env_fn, collector_env_cfg, evaluator_env_cfg = get_vec_env_setting(cfg.env)
    else:
        env_fn, collector_env_cfg, evaluator_env_cfg = env_setting
    collector_env = create_env_manager(cfg.env.manager, [partial(env_fn, cfg=c) for c in collector_env_cfg])
    evaluator_env = create_env_manager(cfg.env.manager, [partial(env_fn, cfg=c) for c in evaluator_env_cfg])
    collector_env.seed(cfg.seed)
    evaluator_env.seed(cfg.seed, dynamic_seed=False)
    set_pkg_seed(cfg.seed, use_cuda=cfg.policy.cuda)
    policy = create_policy(cfg.policy, model=model, enable_field=['learn', 'collect', 'eval', 'command'])

    # Create worker components: learner, collector, evaluator, replay buffer, commander.
    tb_logger = SummaryWriter(os.path.join('./{}/log/'.format(cfg.exp_name), 'serial'))
    learner = BaseLearner(cfg.policy.learn.learner, policy.learn_mode, tb_logger, exp_name=cfg.exp_name)
    collector = create_serial_collector(
        cfg.policy.collect.collector,
        env=collector_env,
        policy=policy.collect_mode,
        tb_logger=tb_logger,
        exp_name=cfg.exp_name
    )
    evaluator = InteractionSerialEvaluator(
        cfg.policy.eval.evaluator, evaluator_env, policy.eval_mode, tb_logger, exp_name=cfg.exp_name
    )
    commander = BaseSerialCommander(
        cfg.policy.other.commander, learner, collector, evaluator, None, policy.command_mode
    )
    reward_model = create_reward_model(cfg, policy.collect_mode.get_attribute('device'), tb_logger)
    reward_model.train()
    # ==========
    # Main loop
    # ==========
    # Learner's before_run hook.
    learner.call_hook('before_run')

    while True:
        collect_kwargs = commander.step()
        # Evaluate policy performance
        if evaluator.should_eval(learner.train_iter):
            stop, reward = evaluator.eval(learner.save_checkpoint, learner.train_iter, collector.envstep)
            if stop:
                break
        # Collect data by default config n_sample/n_episode
        new_data = collector.collect(train_iter=learner.train_iter)
        train_data = new_data
        # update train_data reward using the augmented reward
        train_data_augmented = reward_model.estimate(train_data)
        learner.train(train_data_augmented, collector.envstep)
        if collector.envstep >= max_env_step or learner.train_iter >= max_train_iter:
            break

    # Learner's after_run hook.
    learner.call_hook('after_run')
    return policy