ding.example.ppo_lunarlander

ding.example.ppo_lunarlander

Full Source Code

../ding/example/ppo_lunarlander.py

1import gym 2from ditk import logging 3from ding.model import VAC 4from ding.policy import PPOPolicy 5from ding.envs import DingEnvWrapper, BaseEnvManagerV2 6from ding.config import compile_config 7from ding.framework import task, ding_init 8from ding.framework.context import OnlineRLContext 9from ding.framework.middleware import multistep_trainer, StepCollector, interaction_evaluator, CkptSaver, \ 10 gae_estimator, online_logger 11from ding.utils import set_pkg_seed 12from dizoo.box2d.lunarlander.config.lunarlander_ppo_config import main_config, create_config 13 14 15def main(): 16 logging.getLogger().setLevel(logging.INFO) 17 cfg = compile_config(main_config, create_cfg=create_config, auto=True) 18 ding_init(cfg) 19 with task.start(async_mode=False, ctx=OnlineRLContext()): 20 collector_env = BaseEnvManagerV2( 21 env_fn=[lambda: DingEnvWrapper(gym.make("LunarLander-v2")) for _ in range(cfg.env.collector_env_num)], 22 cfg=cfg.env.manager 23 ) 24 evaluator_env = BaseEnvManagerV2( 25 env_fn=[lambda: DingEnvWrapper(gym.make("LunarLander-v2")) for _ in range(cfg.env.evaluator_env_num)], 26 cfg=cfg.env.manager 27 ) 28 29 set_pkg_seed(cfg.seed, use_cuda=cfg.policy.cuda) 30 31 model = VAC(**cfg.policy.model) 32 policy = PPOPolicy(cfg.policy, model=model) 33 34 task.use(interaction_evaluator(cfg, policy.eval_mode, evaluator_env)) 35 task.use(StepCollector(cfg, policy.collect_mode, collector_env)) 36 task.use(gae_estimator(cfg, policy.collect_mode)) 37 task.use(multistep_trainer(policy.learn_mode, log_freq=50)) 38 task.use(CkptSaver(policy, cfg.exp_name, train_freq=100)) 39 task.use(online_logger(train_show_freq=3)) 40 task.run() 41 42 43if __name__ == "__main__": 44 main()