ding.example.ppo

ding.example.ppo

Example of PPO pipeline

Use the pipeline on a single process:

python3 -u ding/example/ppo.py

Use the pipeline on multiple processes:

We surpose there are N processes (workers) = 1 learner + 1 evaluator + (N-2) collectors

First Example —— Execute on one machine with multi processes.

Execute 4 processes with 1 learner + 1 evaluator + 2 collectors Remember to keep them connected by mesh to ensure that they can exchange information with each other.

ditask --package . --main ding.example.ppo.main --parallel-workers 4 --topology mesh

Full Source Code

../ding/example/ppo.py

1""" 2# Example of PPO pipeline 3 4Use the pipeline on a single process: 5 6> python3 -u ding/example/ppo.py 7 8Use the pipeline on multiple processes: 9 10We surpose there are N processes (workers) = 1 learner + 1 evaluator + (N-2) collectors 11 12## First Example —— Execute on one machine with multi processes. 13 14Execute 4 processes with 1 learner + 1 evaluator + 2 collectors 15Remember to keep them connected by mesh to ensure that they can exchange information with each other. 16 17> ditask --package . --main ding.example.ppo.main --parallel-workers 4 --topology mesh 18""" 19import gym 20from ditk import logging 21from ding.model import VAC 22from ding.policy import PPOPolicy 23from ding.envs import DingEnvWrapper, BaseEnvManagerV2 24from ding.data import DequeBuffer 25from ding.config import compile_config 26from ding.framework import task, ding_init 27from ding.framework.context import OnlineRLContext 28from ding.framework.middleware import multistep_trainer, StepCollector, interaction_evaluator, CkptSaver, \ 29 gae_estimator, online_logger, ContextExchanger, ModelExchanger 30from ding.utils import set_pkg_seed 31from dizoo.classic_control.cartpole.config.cartpole_ppo_config import main_config, create_config 32 33 34def main(): 35 logging.getLogger().setLevel(logging.INFO) 36 cfg = compile_config(main_config, create_cfg=create_config, auto=True, save_cfg=task.router.node_id == 0) 37 ding_init(cfg) 38 with task.start(async_mode=False, ctx=OnlineRLContext()): 39 collector_env = BaseEnvManagerV2( 40 env_fn=[lambda: DingEnvWrapper(gym.make("CartPole-v0")) for _ in range(cfg.env.collector_env_num)], 41 cfg=cfg.env.manager 42 ) 43 evaluator_env = BaseEnvManagerV2( 44 env_fn=[lambda: DingEnvWrapper(gym.make("CartPole-v0")) for _ in range(cfg.env.evaluator_env_num)], 45 cfg=cfg.env.manager 46 ) 47 48 set_pkg_seed(cfg.seed, use_cuda=cfg.policy.cuda) 49 50 model = VAC(**cfg.policy.model) 51 policy = PPOPolicy(cfg.policy, model=model) 52 53 # Consider the case with multiple processes 54 if task.router.is_active: 55 # You can use labels to distinguish between workers with different roles, 56 # here we use node_id to distinguish. 57 if task.router.node_id == 0: 58 task.add_role(task.role.LEARNER) 59 elif task.router.node_id == 1: 60 task.add_role(task.role.EVALUATOR) 61 else: 62 task.add_role(task.role.COLLECTOR) 63 64 # Sync their context and model between each worker. 65 task.use(ContextExchanger(skip_n_iter=1)) 66 task.use(ModelExchanger(model)) 67 68 task.use(interaction_evaluator(cfg, policy.eval_mode, evaluator_env)) 69 task.use(StepCollector(cfg, policy.collect_mode, collector_env)) 70 task.use(gae_estimator(cfg, policy.collect_mode)) 71 task.use(multistep_trainer(policy.learn_mode, log_freq=50)) 72 task.use(CkptSaver(policy, cfg.exp_name, train_freq=100)) 73 task.use(online_logger(train_show_freq=3)) 74 task.run() 75 76 77if __name__ == "__main__": 78 main()