import argparse
import os
import gym

import torch
import torch.nn as nn

# import the skrl components to build the RL system
from skrl import logger
from skrl.agents.torch.ppo import PPO, PPO_CFG
from skrl.envs.wrappers.torch import wrap_env
from skrl.memories.torch import RandomMemory
from skrl.models.torch import DeterministicMixin, GaussianMixin, Model
from skrl.resources.preprocessors.torch import RunningStandardScaler
from skrl.resources.schedulers.torch import KLAdaptiveLR
from skrl.trainers.torch import SequentialTrainer
from skrl.utils import set_seed


# parse arguments
parser = argparse.ArgumentParser()
parser.add_argument("--num_envs", type=int, default=1, help="Number of environments")
parser.add_argument("--headless", action="store_true", help="Run in headless mode (no rendering)")
parser.add_argument("--seed", type=int, default=None, help="Random seed")
parser.add_argument("--checkpoint", type=str, default=None, help="Load checkpoint from path")
parser.add_argument("--eval", action="store_true", help="Run in evaluation mode (logging/checkpointing disabled)")
args, _ = parser.parse_known_args()


# seed for reproducibility
set_seed(args.seed)  # e.g. `set_seed(42)` for fixed seed


# define models (stochastic and deterministic models) using mixins
class Policy(GaussianMixin, Model):
    def __init__(
        self,
        observation_space,
        state_space,
        action_space,
        device,
        clip_actions=False,
        clip_log_std=True,
        min_log_std=-20,
        max_log_std=2,
        reduction="sum",
    ):
        Model.__init__(
            self, observation_space=observation_space, state_space=state_space, action_space=action_space, device=device
        )
        GaussianMixin.__init__(
            self,
            clip_actions=clip_actions,
            clip_log_std=clip_log_std,
            min_log_std=min_log_std,
            max_log_std=max_log_std,
            reduction=reduction,
        )

        self.net = nn.Sequential(
            nn.Linear(self.num_observations, 64),
            nn.ReLU(),
            nn.Linear(64, 64),
            nn.ReLU(),
            nn.Linear(64, self.num_actions),
            nn.Tanh(),
        )
        self.log_std_parameter = nn.Parameter(torch.zeros(self.num_actions))

    def compute(self, inputs, role):
        x = self.net(inputs["observations"])
        # Pendulum-v1 action_space is -2 to 2
        return 2.0 * x, {"log_std": self.log_std_parameter}


class Value(DeterministicMixin, Model):
    def __init__(self, observation_space, state_space, action_space, device):
        Model.__init__(
            self, observation_space=observation_space, state_space=state_space, action_space=action_space, device=device
        )
        DeterministicMixin.__init__(self)

        self.net = nn.Sequential(
            nn.Linear(self.num_observations, 64),
            nn.ReLU(),
            nn.Linear(64, 64),
            nn.ReLU(),
            nn.Linear(64, 1),
        )

    def compute(self, inputs, role):
        return self.net(inputs["observations"]), {}


# load the environment (note: the environment version may change depending on the gym version)
task_name = "Pendulum"
render_mode = "human" if not args.headless else None
env_id = [spec for spec in gym.envs.registry if spec.startswith(f"{task_name}-v")][-1]  # get latest environment version
if args.num_envs <= 1:
    env = gym.make(env_id, render_mode=render_mode)
else:
    env = gym.vector.make(env_id, num_envs=args.num_envs, render_mode=render_mode, asynchronous=False)
# wrap the environment
env = wrap_env(env)

device = env.device


# instantiate a memory as rollout buffer (any memory can be used for this)
memory = RandomMemory(memory_size=1024, num_envs=env.num_envs, device=device)


# instantiate the agent's models (function approximators).
# PPO requires 2 models, visit its documentation for more details
# https://skrl.readthedocs.io/en/latest/api/agents/ppo.html#models
models = {}
models["policy"] = Policy(env.observation_space, env.state_space, env.action_space, device)
models["value"] = Value(env.observation_space, env.state_space, env.action_space, device)


# configure and instantiate the agent (visit its documentation to see all the options)
# https://skrl.readthedocs.io/en/latest/api/agents/ppo.html#configuration-and-hyperparameters
cfg = PPO_CFG()
cfg.rollouts = 1024  # memory_size
cfg.learning_epochs = 10
cfg.mini_batches = 32
cfg.discount_factor = 0.9
cfg.gae_lambda = 0.95
cfg.learning_rate = 1e-3
cfg.learning_rate_scheduler = KLAdaptiveLR
cfg.learning_rate_scheduler_kwargs = {"kl_threshold": 0.008}
cfg.grad_norm_clip = 0.5
cfg.ratio_clip = 0.2
cfg.value_clip = 0.2
cfg.entropy_loss_scale = 0.0
cfg.value_loss_scale = 0.5
cfg.kl_threshold = 0
cfg.observation_preprocessor = RunningStandardScaler
cfg.observation_preprocessor_kwargs = {"size": env.observation_space, "device": device}
cfg.value_preprocessor = RunningStandardScaler
cfg.value_preprocessor_kwargs = {"size": 1, "device": device}
# logging to TensorBoard and write checkpoints (in timesteps)
cfg.experiment.write_interval = "auto" if not args.eval else 0
cfg.experiment.checkpoint_interval = "auto" if not args.eval else 0
cfg.experiment.directory = f"runs/torch/{task_name}"

agent = PPO(
    models=models,
    memory=memory,
    cfg=cfg,
    observation_space=env.observation_space,
    state_space=env.state_space,
    action_space=env.action_space,
    device=device,
)


# configure and instantiate the RL trainer
cfg_trainer = {"timesteps": 100000, "headless": args.headless}
trainer = SequentialTrainer(cfg=cfg_trainer, env=env, agents=agent)

if args.checkpoint:
    if not os.path.exists(args.checkpoint):
        logger.error(f"Checkpoint file not found: '{args.checkpoint}'")
        exit(1)
    agent.load(args.checkpoint)

trainer.train() if not args.eval else trainer.eval()