ding.example.cql

ding.example.cql

Full Source Code

../ding/example/cql.py

1import gym 2from ditk import logging 3from ding.model import QAC 4from ding.policy import CQLPolicy 5from ding.envs import DingEnvWrapper, BaseEnvManagerV2 6from ding.data import create_dataset 7from ding.config import compile_config 8from ding.framework import task, ding_init 9from ding.framework.context import OfflineRLContext 10from ding.framework.middleware import interaction_evaluator, trainer, CkptSaver, offline_data_fetcher, offline_logger 11from ding.utils import set_pkg_seed 12from dizoo.classic_control.pendulum.envs.pendulum_env import PendulumEnv 13from dizoo.classic_control.pendulum.config.pendulum_cql_config import main_config, create_config 14 15 16def main(): 17 # If you don't have offline data, you need to prepare if first and set the data_path in config 18 # For demostration, we also can train a RL policy (e.g. SAC) and collect some data 19 logging.getLogger().setLevel(logging.INFO) 20 cfg = compile_config(main_config, create_cfg=create_config, auto=True) 21 ding_init(cfg) 22 with task.start(async_mode=False, ctx=OfflineRLContext()): 23 evaluator_env = BaseEnvManagerV2( 24 env_fn=[lambda: PendulumEnv(cfg.env) for _ in range(cfg.env.evaluator_env_num)], cfg=cfg.env.manager 25 ) 26 27 set_pkg_seed(cfg.seed, use_cuda=cfg.policy.cuda) 28 29 dataset = create_dataset(cfg) 30 model = QAC(**cfg.policy.model) 31 policy = CQLPolicy(cfg.policy, model=model) 32 33 task.use(interaction_evaluator(cfg, policy.eval_mode, evaluator_env)) 34 task.use(offline_data_fetcher(cfg, dataset)) 35 task.use(trainer(cfg, policy.learn_mode)) 36 task.use(CkptSaver(policy, cfg.exp_name, train_freq=100)) 37 task.use(offline_logger()) 38 task.run() 39 40 41if __name__ == "__main__": 42 main()