ding.example.trex

ding.example.trex

Full Source Code

../ding/example/trex.py

1import gym 2from tensorboardX import SummaryWriter 3import copy 4import easydict 5import os 6from ditk import logging 7 8from ding.model import DQN 9from ding.policy import DQNPolicy 10from ding.envs import DingEnvWrapper, BaseEnvManagerV2 11from ding.data import DequeBuffer 12from ding.config import compile_config 13from ding.framework import task 14from ding.framework.context import OnlineRLContext 15from ding.framework.middleware import OffPolicyLearner, StepCollector, interaction_evaluator, \ 16 eps_greedy_handler, CkptSaver, eps_greedy_masker, sqil_data_pusher, data_pusher 17from ding.utils import set_pkg_seed 18from ding.entry import trex_collecting_data 19from ding.reward_model import create_reward_model 20from dizoo.classic_control.cartpole.config.cartpole_trex_dqn_config import main_config, create_config 21 22 23def main(): 24 logging.getLogger().setLevel(logging.INFO) 25 demo_arg = easydict.EasyDict({'cfg': [main_config, create_config], 'seed': 0}) 26 trex_collecting_data(demo_arg) 27 cfg = compile_config(main_config, create_cfg=create_config, auto=True, renew_dir=False) 28 29 with task.start(async_mode=False, ctx=OnlineRLContext()): 30 collector_env = BaseEnvManagerV2( 31 env_fn=[lambda: DingEnvWrapper(gym.make("CartPole-v0")) for _ in range(cfg.env.collector_env_num)], 32 cfg=cfg.env.manager 33 ) 34 35 evaluator_env = BaseEnvManagerV2( 36 env_fn=[lambda: DingEnvWrapper(gym.make("CartPole-v0")) for _ in range(cfg.env.evaluator_env_num)], 37 cfg=cfg.env.manager 38 ) 39 40 set_pkg_seed(cfg.seed, use_cuda=cfg.policy.cuda) 41 model = DQN(**cfg.policy.model) 42 buffer_ = DequeBuffer(size=cfg.policy.other.replay_buffer.replay_buffer_size) 43 policy = DQNPolicy(cfg.policy, model=model) 44 45 tb_logger = SummaryWriter(os.path.join('./{}/log/'.format(cfg.exp_name), 'serial')) 46 reward_model = create_reward_model(copy.deepcopy(cfg), policy.collect_mode.get_attribute('device'), tb_logger) 47 reward_model.train() 48 49 task.use(interaction_evaluator(cfg, policy.eval_mode, evaluator_env)) 50 task.use(eps_greedy_handler(cfg)) 51 task.use(StepCollector(cfg, policy.collect_mode, collector_env)) 52 task.use(data_pusher(cfg, buffer_)) 53 task.use(OffPolicyLearner(cfg, policy.learn_mode, buffer_, reward_model)) 54 task.use(CkptSaver(policy, cfg.exp_name, train_freq=100)) 55 task.run() 56 57 58if __name__ == "__main__": 59 main()