"""Example of handling variable length and/or parametric action spaces.

This is a toy example of the action-embedding based approach for handling large
discrete action spaces (potentially infinite in size), similar to this:

    https://neuro.cs.ut.ee/the-use-of-embeddings-in-openai-five/

This currently works with RLlib's policy gradient style algorithms
(e.g., PG, PPO, IMPALA, A2C) and also DQN.

Note that since the model outputs now include "-inf" tf.float32.min
values, not all algorithm options are supported at the moment. For example,
algorithms might crash if they don't properly ignore the -inf action scores.
Working configurations are given below.
"""

import argparse
import os

import ray
from ray import tune
from ray.rllib.examples.env.parametric_actions_cartpole import \
    ParametricActionsCartPoleNoEmbeddings
from ray.rllib.examples.models.parametric_actions_model import \
    ParametricActionsModelThatLearnsEmbeddings
from ray.rllib.models import ModelCatalog
from ray.rllib.utils.test_utils import check_learning_achieved
from ray.tune.registry import register_env

parser = argparse.ArgumentParser()
parser.add_argument("--run", type=str, default="PPO")
parser.add_argument(
    "--framework",
    choices=["tf", "tf2", "tfe"],
    default="tf",
    help="The DL framework specifier (torch not supported yet "
    "due to lack of model).")
parser.add_argument("--as-test", action="store_true")
parser.add_argument("--stop-iters", type=int, default=200)
parser.add_argument("--stop-reward", type=float, default=150.0)
parser.add_argument("--stop-timesteps", type=int, default=100000)

if __name__ == "__main__":
    args = parser.parse_args()
    ray.init()

    register_env("pa_cartpole",
                 lambda _: ParametricActionsCartPoleNoEmbeddings(10))

    ModelCatalog.register_custom_model(
        "pa_model", ParametricActionsModelThatLearnsEmbeddings)

    if args.run == "DQN":
        cfg = {
            # TODO(ekl) we need to set these to prevent the masked values
            # from being further processed in DistributionalQModel, which
            # would mess up the masking. It is possible to support these if we
            # defined a custom DistributionalQModel that is aware of masking.
            "hiddens": [],
            "dueling": False,
        }
    else:
        cfg = {}

    config = dict(
        {
            "env": "pa_cartpole",
            "model": {
                "custom_model": "pa_model",
            },
            # Use GPUs iff `RLLIB_NUM_GPUS` env var set to > 0.
            "num_gpus": int(os.environ.get("RLLIB_NUM_GPUS", "0")),
            "num_workers": 0,
            "framework": args.framework,
        },
        **cfg)

    stop = {
        "training_iteration": args.stop_iters,
        "timesteps_total": args.stop_timesteps,
        "episode_reward_mean": args.stop_reward,
    }

    results = tune.run(args.run, stop=stop, config=config, verbose=2)

    if args.as_test:
        check_learning_achieved(results, args.stop_reward)

    ray.shutdown()