ding.example.dqn¶
ding.example.dqn
¶
Example of DQN pipeline¶
Use the pipeline on a single process:
python3 -u ding/example/dqn.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.dqn.main --parallel-workers 4 --topology mesh
Second Example —— Execute on multiple machines.¶
- Execute 1 learner + 1 evaluator on one machine.
ditask --package . --main ding.example.dqn.main --parallel-workers 2 --topology mesh --node-ids 0 --ports 50515
- Execute 2 collectors on another machine. (Suppose the ip of the first machine is 127.0.0.1).
Here we use
alonetopology instead ofmeshbecause the collectors do not need communicate with each other. Remember thenode_idscannot be duplicated with the learner, evaluator processes. And remember to set theports(should not conflict with others) andattach_toparameters. The value of theattach_toparameter should be obtained from the log of the process started earlier (e.g. 'NNG listen on tcp://10.0.0.4:50515').
ditask --package . --main ding.example.dqn.main --parallel-workers 2 --topology alone --node-ids 2 --ports 50517 --attach-to tcp://10.0.0.4:50515,tcp://127.0.0.1:50516
- You can repeat step 2 to start more collectors on other machines.
Full Source Code
../ding/example/dqn.py
1""" 2# Example of DQN pipeline 3 4Use the pipeline on a single process: 5 6> python3 -u ding/example/dqn.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.dqn.main --parallel-workers 4 --topology mesh 18 19## Second Example —— Execute on multiple machines. 20 211. Execute 1 learner + 1 evaluator on one machine. 22 23> ditask --package . --main ding.example.dqn.main --parallel-workers 2 --topology mesh --node-ids 0 --ports 50515 24 252. Execute 2 collectors on another machine. (Suppose the ip of the first machine is 127.0.0.1). 26 Here we use `alone` topology instead of `mesh` because the collectors do not need communicate with each other. 27 Remember the `node_ids` cannot be duplicated with the learner, evaluator processes. 28 And remember to set the `ports` (should not conflict with others) and `attach_to` parameters. 29 The value of the `attach_to` parameter should be obtained from the log of the 30 process started earlier (e.g. 'NNG listen on tcp://10.0.0.4:50515'). 31 32> ditask --package . --main ding.example.dqn.main --parallel-workers 2 --topology alone --node-ids 2 \ 33 --ports 50517 --attach-to tcp://10.0.0.4:50515,tcp://127.0.0.1:50516 34 353. You can repeat step 2 to start more collectors on other machines. 36""" 37import gym 38from ditk import logging 39from ding.data.model_loader import FileModelLoader 40from ding.data.storage_loader import FileStorageLoader 41from ding.model import DQN 42from ding.policy import DQNPolicy 43from ding.envs import DingEnvWrapper, BaseEnvManagerV2 44from ding.data import DequeBuffer 45from ding.config import compile_config 46from ding.framework import task, ding_init 47from ding.framework.context import OnlineRLContext 48from ding.framework.middleware import OffPolicyLearner, StepCollector, interaction_evaluator, data_pusher, \ 49 eps_greedy_handler, CkptSaver, ContextExchanger, ModelExchanger, online_logger 50from ding.utils import set_pkg_seed 51from dizoo.classic_control.cartpole.config.cartpole_dqn_config import main_config, create_config 52 53 54def main(): 55 logging.getLogger().setLevel(logging.INFO) 56 cfg = compile_config(main_config, create_cfg=create_config, auto=True, save_cfg=task.router.node_id == 0) 57 ding_init(cfg) 58 with task.start(async_mode=False, ctx=OnlineRLContext()): 59 collector_env = BaseEnvManagerV2( 60 env_fn=[lambda: DingEnvWrapper(gym.make("CartPole-v0")) for _ in range(cfg.env.collector_env_num)], 61 cfg=cfg.env.manager 62 ) 63 evaluator_env = BaseEnvManagerV2( 64 env_fn=[lambda: DingEnvWrapper(gym.make("CartPole-v0")) for _ in range(cfg.env.evaluator_env_num)], 65 cfg=cfg.env.manager 66 ) 67 68 set_pkg_seed(cfg.seed, use_cuda=cfg.policy.cuda) 69 70 model = DQN(**cfg.policy.model) 71 buffer_ = DequeBuffer(size=cfg.policy.other.replay_buffer.replay_buffer_size) 72 policy = DQNPolicy(cfg.policy, model=model) 73 74 # Consider the case with multiple processes 75 if task.router.is_active: 76 # You can use labels to distinguish between workers with different roles, 77 # here we use node_id to distinguish. 78 if task.router.node_id == 0: 79 task.add_role(task.role.LEARNER) 80 elif task.router.node_id == 1: 81 task.add_role(task.role.EVALUATOR) 82 else: 83 task.add_role(task.role.COLLECTOR) 84 85 # Sync their context and model between each worker. 86 task.use(ContextExchanger(skip_n_iter=1)) 87 task.use(ModelExchanger(model)) 88 89 # Here is the part of single process pipeline. 90 task.use(interaction_evaluator(cfg, policy.eval_mode, evaluator_env)) 91 task.use(eps_greedy_handler(cfg)) 92 task.use(StepCollector(cfg, policy.collect_mode, collector_env)) 93 task.use(data_pusher(cfg, buffer_)) 94 task.use(OffPolicyLearner(cfg, policy.learn_mode, buffer_)) 95 task.use(online_logger(train_show_freq=10)) 96 task.use(CkptSaver(policy, cfg.exp_name, train_freq=100)) 97 98 task.run() 99 100 101if __name__ == "__main__": 102 main()