ding.example.ddpg

ding.example.ddpg

Full Source Code

../ding/example/ddpg.py

1import gym 2from ditk import logging 3from ding.model.template.qac import ContinuousQAC 4from ding.policy import DDPGPolicy 5from ding.envs import DingEnvWrapper, BaseEnvManagerV2 6from ding.data import DequeBuffer 7from ding.config import compile_config 8from ding.framework import task 9from ding.framework.context import OnlineRLContext 10from ding.framework.middleware import OffPolicyLearner, StepCollector, interaction_evaluator, data_pusher, \ 11 CkptSaver, termination_checker 12from ding.utils import set_pkg_seed 13from dizoo.classic_control.pendulum.envs.pendulum_env import PendulumEnv 14from dizoo.classic_control.pendulum.config.pendulum_ddpg_config import main_config, create_config 15 16 17def main(): 18 logging.getLogger().setLevel(logging.INFO) 19 cfg = compile_config(main_config, create_cfg=create_config, auto=True) 20 with task.start(async_mode=False, ctx=OnlineRLContext()): 21 collector_env = BaseEnvManagerV2( 22 env_fn=[lambda: PendulumEnv(cfg.env) for _ in range(cfg.env.collector_env_num)], cfg=cfg.env.manager 23 ) 24 evaluator_env = BaseEnvManagerV2( 25 env_fn=[lambda: PendulumEnv(cfg.env) for _ in range(cfg.env.evaluator_env_num)], cfg=cfg.env.manager 26 ) 27 28 set_pkg_seed(cfg.seed, use_cuda=cfg.policy.cuda) 29 30 model = ContinuousQAC(**cfg.policy.model) 31 buffer_ = DequeBuffer(size=cfg.policy.other.replay_buffer.replay_buffer_size) 32 policy = DDPGPolicy(cfg.policy, model=model) 33 34 task.use(interaction_evaluator(cfg, policy.eval_mode, evaluator_env)) 35 task.use( 36 StepCollector(cfg, policy.collect_mode, collector_env, random_collect_size=cfg.policy.random_collect_size) 37 ) 38 task.use(data_pusher(cfg, buffer_)) 39 task.use(OffPolicyLearner(cfg, policy.learn_mode, buffer_)) 40 task.use(CkptSaver(policy, cfg.exp_name, train_freq=100)) 41 task.use(termination_checker(max_train_iter=10000)) 42 task.run() 43 44 45if __name__ == "__main__": 46 main()