ding.example.ppg_offpolicy

ding.example.ppg_offpolicy

Full Source Code

../ding/example/ppg_offpolicy.py

1import gym 2from ditk import logging 3from ding.model import PPG 4from ding.policy import PPGOffPolicy 5from ding.envs import DingEnvWrapper, BaseEnvManagerV2 6from ding.data import DequeBuffer 7from ding.data.buffer.middleware import use_time_check, sample_range_view 8from ding.config import compile_config 9from ding.framework import task 10from ding.framework.context import OnlineRLContext 11from ding.framework.middleware import OffPolicyLearner, StepCollector, interaction_evaluator, data_pusher, \ 12 CkptSaver, gae_estimator 13from ding.utils import set_pkg_seed 14from dizoo.classic_control.cartpole.config.cartpole_ppg_offpolicy_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: DingEnvWrapper(gym.make("CartPole-v0")) for _ in range(cfg.env.collector_env_num)], 23 cfg=cfg.env.manager 24 ) 25 evaluator_env = BaseEnvManagerV2( 26 env_fn=[lambda: DingEnvWrapper(gym.make("CartPole-v0")) for _ in range(cfg.env.evaluator_env_num)], 27 cfg=cfg.env.manager 28 ) 29 30 set_pkg_seed(cfg.seed, use_cuda=cfg.policy.cuda) 31 32 model = PPG(**cfg.policy.model) 33 buffer_cfg = cfg.policy.other.replay_buffer 34 max_size = max(buffer_cfg.policy.replay_buffer_size, buffer_cfg.value.replay_buffer_size) 35 buffer_ = DequeBuffer(size=max_size) 36 policy_buffer = buffer_.view() # shallow copy 37 policy_buffer.use(use_time_check(policy_buffer, max_use=buffer_cfg.policy.max_use)) 38 policy_buffer.use(sample_range_view(policy_buffer, start=-buffer_cfg.policy.replay_buffer_size)) 39 value_buffer = buffer_.view() 40 value_buffer.use(use_time_check(value_buffer, max_use=buffer_cfg.value.max_use)) 41 value_buffer.use(sample_range_view(value_buffer, start=-buffer_cfg.value.replay_buffer_size)) 42 policy = PPGOffPolicy(cfg.policy, model=model) 43 44 task.use(interaction_evaluator(cfg, policy.eval_mode, evaluator_env)) 45 task.use(StepCollector(cfg, policy.collect_mode, collector_env)) 46 task.use(gae_estimator(cfg, policy.collect_mode, buffer_)) 47 task.use(OffPolicyLearner(cfg, policy.learn_mode, {'policy': policy_buffer, 'value': value_buffer})) 48 task.use(CkptSaver(policy, cfg.exp_name, train_freq=100)) 49 task.run() 50 51 52if __name__ == "__main__": 53 main()