ding.example.r2d2

ding.example.r2d2

Full Source Code

../ding/example/r2d2.py

1import gym 2from ditk import logging 3from ding.model import DRQN 4from ding.policy import R2D2Policy 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 eps_greedy_handler, CkptSaver, nstep_reward_enhancer 12from ding.utils import set_pkg_seed 13from dizoo.classic_control.cartpole.config.cartpole_r2d2_config import main_config, create_config 14 15 16def main(): 17 logging.getLogger().setLevel(logging.INFO) 18 cfg = compile_config(main_config, create_cfg=create_config, auto=True) 19 with task.start(async_mode=False, ctx=OnlineRLContext()): 20 collector_env = BaseEnvManagerV2( 21 env_fn=[lambda: DingEnvWrapper(gym.make("CartPole-v0")) 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("CartPole-v0")) 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 = DRQN(**cfg.policy.model) 32 buffer_ = DequeBuffer(size=cfg.policy.other.replay_buffer.replay_buffer_size) 33 policy = R2D2Policy(cfg.policy, model=model) 34 35 task.use(interaction_evaluator(cfg, policy.eval_mode, evaluator_env)) 36 task.use(eps_greedy_handler(cfg)) 37 task.use(StepCollector(cfg, policy.collect_mode, collector_env)) 38 task.use(nstep_reward_enhancer(cfg)) 39 task.use(data_pusher(cfg, buffer_, group_by_env=True)) 40 task.use(OffPolicyLearner(cfg, policy.learn_mode, buffer_)) 41 task.use(CkptSaver(policy, cfg.exp_name, train_freq=100)) 42 task.run() 43 44 45if __name__ == "__main__": 46 main()