import os
import torch
from tensorboardX import SummaryWriter
from ding.config import compile_config
from ding.worker import BaseLearner, SampleSerialCollector, InteractionSerialEvaluator, AdvancedReplayBuffer
from ding.policy import FQFPolicy
from ding.model import FQF
from ding.utils import set_pkg_seed
from ding.rl_utils import get_epsilon_greedy_fn
from dizoo.atari.config.serial.qbert.qbert_fqf_config import qbert_fqf_config, create_config
from ding.utils import DistContext
from functools import partial
from ding.envs import get_vec_env_setting, create_env_manager


def main(cfg, create_cfg, seed=0):

    cfg = compile_config(cfg, seed=seed, env=None, auto=True, create_cfg=create_cfg, save_cfg=True)

    # Create main components: env, policy
    env_fn, collector_env_cfg, evaluator_env_cfg = get_vec_env_setting(cfg.env)
    collector_env = create_env_manager(cfg.env.manager, [partial(env_fn, cfg=c) for c in collector_env_cfg])
    evaluator_env = create_env_manager(cfg.env.manager, [partial(env_fn, cfg=c) for c in evaluator_env_cfg])

    # 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 = FQF(**cfg.policy.model)
    policy = FQFPolicy(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)

    # Set up other modules, etc. epsilon greedy
    eps_cfg = cfg.policy.other.eps
    epsilon_greedy = get_epsilon_greedy_fn(eps_cfg.start, eps_cfg.end, eps_cfg.decay, eps_cfg.type)

    # Training & Evaluation loop
    while True:
        # Evaluating 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
        # Update other modules
        eps = epsilon_greedy(collector.envstep)
        # Sampling data from environments
        new_data = collector.collect(train_iter=learner.train_iter, policy_kwargs={'eps': eps})
        replay_buffer.push(new_data, cur_collector_envstep=collector.envstep)
        # Training
        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)

        if collector.envstep >= 10000000:
            break


if __name__ == "__main__":
    # with DistContext():
    main(qbert_fqf_config, create_config)