File size: 3,083 Bytes
079c32c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
import os
import gym
from tensorboardX import SummaryWriter

from ding.config import compile_config
from ding.worker import BaseLearner, SampleSerialCollector, InteractionSerialEvaluator, AdvancedReplayBuffer
from ding.envs import BaseEnvManager
from ding.policy.d4pg import D4PGPolicy
from ding.model.template import QACDIST
from ding.utils import set_pkg_seed
from dizoo.classic_control.pendulum.envs import PendulumEnv
from dizoo.classic_control.pendulum.config.pendulum_d4pg_config import pendulum_d4pg_config


def main(cfg, seed=0):
    cfg = compile_config(
        cfg,
        BaseEnvManager,
        D4PGPolicy,
        BaseLearner,
        SampleSerialCollector,
        InteractionSerialEvaluator,
        AdvancedReplayBuffer,
        save_cfg=True
    )

    # Set up envs for collection and evaluation
    collector_env_num, evaluator_env_num = cfg.env.collector_env_num, cfg.env.evaluator_env_num
    collector_env = BaseEnvManager(
        env_fn=[lambda: PendulumEnv(cfg.env) for _ in range(collector_env_num)], cfg=cfg.env.manager
    )
    evaluator_env = BaseEnvManager(
        env_fn=[lambda: PendulumEnv(cfg.env) for _ in range(evaluator_env_num)], cfg=cfg.env.manager
    )

    # Set random seed for all package and instance
    collector_env.seed(seed)
    evaluator_env.seed(seed, dynamic_seed=False)
    set_pkg_seed(seed, use_cuda=cfg.policy.cuda)

    # Set up RL Policy
    model = QACDIST(**cfg.policy.model)
    policy = D4PGPolicy(cfg.policy, model=model)

    # Set up collection, training and evaluation utilities
    tb_logger = SummaryWriter(os.path.join('./{}/log/'.format(cfg.exp_name), 'serial'))
    learner = BaseLearner(cfg.policy.learn.learner, policy.learn_mode, tb_logger, exp_name=cfg.exp_name)
    collector = SampleSerialCollector(
        cfg.policy.collect.collector, collector_env, policy.collect_mode, tb_logger, exp_name=cfg.exp_name
    )
    evaluator = InteractionSerialEvaluator(
        cfg.policy.eval.evaluator, evaluator_env, policy.eval_mode, tb_logger, exp_name=cfg.exp_name
    )
    replay_buffer = AdvancedReplayBuffer(cfg.policy.other.replay_buffer, tb_logger, exp_name=cfg.exp_name)

    # Training & Evaluation loop
    while True:
        # Evaluate at the beginning and with specific frequency
        if evaluator.should_eval(learner.train_iter):
            stop, reward = evaluator.eval(learner.save_checkpoint, learner.train_iter, collector.envstep)
            if stop:
                break
        # Collect data from environments
        new_data = collector.collect(train_iter=learner.train_iter)
        replay_buffer.push(new_data, cur_collector_envstep=collector.envstep)
        # Train
        for i in range(cfg.policy.learn.update_per_collect):
            train_data = replay_buffer.sample(learner.policy.get_attribute('batch_size'), learner.train_iter)
            if train_data is None:
                break
            learner.train(train_data, collector.envstep)
            replay_buffer.update(learner.priority_info)


if __name__ == "__main__":
    main(pendulum_d4pg_config, seed=0)