ding.entry.serial_entry

ding.entry.serial_entry

serial_pipeline(input_cfg, seed=0, env_setting=None, model=None, max_train_iter=int(10000000000.0), max_env_step=int(10000000000.0), dynamic_seed=True)

Overview

Serial pipeline entry for off-policy RL.

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. - dynamic_seed(:obj:Optional[bool]): set dynamic seed for collector. Returns: - policy (:obj:Policy): Converged policy.

Full Source Code

../ding/entry/serial_entry.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, create_serial_evaluator 12from ding.config import read_config, compile_config 13from ding.policy import create_policy 14from ding.utils import set_pkg_seed, get_rank 15from .utils import random_collect, maybe_init_wandb, maybe_finish_wandb 16 17 18def serial_pipeline( 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 dynamic_seed: Optional[bool] = True, 26) -> 'Policy': # noqa 27 """ 28 Overview: 29 Serial pipeline entry for off-policy RL. 30 Arguments: 31 - input_cfg (:obj:`Union[str, Tuple[dict, dict]]`): Config in dict type. \ 32 ``str`` type means config file path. \ 33 ``Tuple[dict, dict]`` type means [user_config, create_cfg]. 34 - seed (:obj:`int`): Random seed. 35 - env_setting (:obj:`Optional[List[Any]]`): A list with 3 elements: \ 36 ``BaseEnv`` subclass, collector env config, and evaluator env config. 37 - model (:obj:`Optional[torch.nn.Module]`): Instance of torch.nn.Module. 38 - max_train_iter (:obj:`Optional[int]`): Maximum policy update iterations in training. 39 - max_env_step (:obj:`Optional[int]`): Maximum collected environment interaction steps. 40 - dynamic_seed(:obj:`Optional[bool]`): set dynamic seed for collector. 41 Returns: 42 - policy (:obj:`Policy`): Converged policy. 43 """ 44 if isinstance(input_cfg, str): 45 cfg, create_cfg = read_config(input_cfg) 46 else: 47 cfg, create_cfg = deepcopy(input_cfg) 48 create_cfg.policy.type = create_cfg.policy.type + '_command' 49 env_fn = None if env_setting is None else env_setting[0] 50 cfg = compile_config( 51 cfg, 52 seed=seed, 53 env=env_fn, 54 auto=True, 55 create_cfg=create_cfg, 56 save_cfg=True, 57 renew_dir=not cfg.policy.learn.get('resume_training', False) 58 ) 59 # Create main components: env, policy 60 if env_setting is None: 61 env_fn, collector_env_cfg, evaluator_env_cfg = get_vec_env_setting(cfg.env) 62 else: 63 env_fn, collector_env_cfg, evaluator_env_cfg = env_setting 64 collector_env = create_env_manager(cfg.env.manager, [partial(env_fn, cfg=c) for c in collector_env_cfg]) 65 evaluator_env = create_env_manager(cfg.env.manager, [partial(env_fn, cfg=c) for c in evaluator_env_cfg]) 66 collector_env.seed(cfg.seed, dynamic_seed=dynamic_seed) 67 evaluator_env.seed(cfg.seed, dynamic_seed=False) 68 set_pkg_seed(cfg.seed, use_cuda=cfg.policy.cuda) 69 policy = create_policy(cfg.policy, model=model, enable_field=['learn', 'collect', 'eval', 'command']) 70 71 # Create worker components: learner, collector, evaluator, replay buffer, commander. 72 wandb_run = maybe_init_wandb(cfg) if get_rank() == 0 else None 73 tb_logger = SummaryWriter(os.path.join('./{}/log/'.format(cfg.exp_name), 'serial')) if get_rank() == 0 else None 74 learner = BaseLearner(cfg.policy.learn.learner, policy.learn_mode, tb_logger, exp_name=cfg.exp_name) 75 collector = create_serial_collector( 76 cfg.policy.collect.collector, 77 env=collector_env, 78 policy=policy.collect_mode, 79 tb_logger=tb_logger, 80 exp_name=cfg.exp_name 81 ) 82 evaluator = create_serial_evaluator( 83 cfg.policy.eval.evaluator, 84 env=evaluator_env, 85 policy=policy.eval_mode, 86 tb_logger=tb_logger, 87 exp_name=cfg.exp_name 88 ) 89 replay_buffer = create_buffer(cfg.policy.other.replay_buffer, tb_logger=tb_logger, exp_name=cfg.exp_name) 90 commander = BaseSerialCommander( 91 cfg.policy.other.commander, learner, collector, evaluator, replay_buffer, policy.command_mode 92 ) 93 # ========== 94 # Main loop 95 # ========== 96 # Learner's before_run hook. 97 learner.call_hook('before_run') 98 if cfg.policy.learn.get('resume_training', False): 99 collector.envstep = learner.collector_envstep 100 101 # Accumulate plenty of data at the beginning of training. 102 if cfg.policy.get('random_collect_size', 0) > 0: 103 random_collect(cfg.policy, policy, collector, collector_env, commander, replay_buffer) 104 while True: 105 collect_kwargs = commander.step() 106 # Evaluate policy performance 107 if evaluator.should_eval(learner.train_iter): 108 stop, eval_info = evaluator.eval(learner.save_checkpoint, learner.train_iter, collector.envstep) 109 if stop: 110 break 111 # Collect data by default config n_sample/n_episode 112 new_data = collector.collect(train_iter=learner.train_iter, policy_kwargs=collect_kwargs) 113 replay_buffer.push(new_data, cur_collector_envstep=collector.envstep) 114 # Learn policy from collected data 115 for i in range(cfg.policy.learn.update_per_collect): 116 # Learner will train ``update_per_collect`` times in one iteration. 117 train_data = replay_buffer.sample(learner.policy.get_attribute('batch_size'), learner.train_iter) 118 if train_data is None: 119 # It is possible that replay buffer's data count is too few to train ``update_per_collect`` times 120 logging.warning( 121 "Replay buffer's data can only train for {} steps. ".format(i) + 122 "You can modify data collect config, e.g. increasing n_sample, n_episode." 123 ) 124 break 125 learner.train(train_data, collector.envstep) 126 if learner.policy.get_attribute('priority'): 127 replay_buffer.update(learner.priority_info) 128 if collector.envstep >= max_env_step or learner.train_iter >= max_train_iter: 129 break 130 131 # Learner's after_run hook. 132 learner.call_hook('after_run') 133 if get_rank() == 0: 134 import time 135 import pickle 136 import numpy as np 137 with open(os.path.join(cfg.exp_name, 'result.pkl'), 'wb') as f: 138 eval_value_raw = eval_info['eval_episode_return'] 139 final_data = { 140 'stop': stop, 141 'env_step': collector.envstep, 142 'train_iter': learner.train_iter, 143 'eval_value': np.mean(eval_value_raw), 144 'eval_value_raw': eval_value_raw, 145 'finish_time': time.ctime(), 146 } 147 pickle.dump(final_data, f) 148 maybe_finish_wandb(wandb_run) 149 return policy