main.py

import argparse
import torch
import time
import os
import numpy as np
from gym.spaces import Box, Discrete
from pathlib import Path
from torch.autograd import Variable
from tensorboardX import SummaryWriter
from utils.make_env import make_env
from utils.buffer import ReplayBuffer
from utils.env_wrappers import SubprocVecEnv, DummyVecEnv
from algorithms.maddpg import MADDPG

USE_CUDA = torch.cuda.is_available()

def make_parallel_env(env_id, n_rollout_threads, seed, discrete_action):
    def get_env_fn(rank):
        def init_env():
            env = make_env(env_id, discrete_action=discrete_action)
            env.seed(seed + rank * 1000)
            np.random.seed(seed + rank * 1000)
            return env
        return init_env
    if n_rollout_threads == 1:
        return DummyVecEnv([get_env_fn(0)])
    else:
        return SubprocVecEnv([get_env_fn(i) for i in range(n_rollout_threads)])

def run(config):
    model_dir = Path('./models') / config.env_id / config.model_name
    if not model_dir.exists():
        curr_run = 'run1'
    else:
        exst_run_nums = [int(str(folder.name).split('run')[1]) for folder in
                         model_dir.iterdir() if
                         str(folder.name).startswith('run')]
        if len(exst_run_nums) == 0:
            curr_run = 'run1'
        else:
            curr_run = 'run%i' % (max(exst_run_nums) + 1)
    run_dir = model_dir / curr_run
    log_dir = run_dir / 'logs'
    os.makedirs(log_dir)
    logger = SummaryWriter(str(log_dir))

    torch.manual_seed(config.seed)
    np.random.seed(config.seed)
    if not USE_CUDA:
        torch.set_num_threads(config.n_training_threads)
    env = make_parallel_env(config.env_id, config.n_rollout_threads, config.seed,
                            config.discrete_action)

    ##################### INITIALIZE FROM SAVED? ###########################
    if init_from_saved:
        if model_path is not None:
            maddpg = MADDPG.init_from_save(model_path)
            print("Initialized from saved model")
    # -------------------------------------------------------------------- #
    else:
        maddpg = MADDPG.init_from_env(env, agent_alg=config.agent_alg,
                                  adversary_alg=config.adversary_alg,
                                  tau=config.tau,
                                  lr=config.lr,
                                  hidden_dim=config.hidden_dim)
    # used for learning (updates)
    replay_buffer = ReplayBuffer(config.buffer_length, maddpg.nagents,
                                 [obsp.shape[0] for obsp in env.observation_space],
                                 [acsp.shape[0] if isinstance(acsp, Box) else acsp.n
                                  for acsp in env.action_space])

    # This is just to store the global rewards and not for updating the policies
    g_storage_buffer = ReplayBuffer(config.buffer_length, maddpg.nagents,
                                 [obsp.shape[0] for obsp in env.observation_space],
                                 [acsp.shape[0] if isinstance(acsp, Box) else acsp.n
                                 for acsp in env.action_space])

    t = 0
    for ep_i in range(0, config.n_episodes, config.n_rollout_threads):
        print("Episodes %i-%i of %i" % (ep_i + 1,
                                        ep_i + 1 + config.n_rollout_threads,
                                        config.n_episodes))
        obs = env.reset()
        # obs.shape = (n_rollout_threads, nagent)(nobs), nobs differs per agent so not tensor
        maddpg.prep_rollouts(device='cpu')

        explr_pct_remaining = max(0, config.n_exploration_eps - ep_i) / config.n_exploration_eps
        maddpg.scale_noise(config.final_noise_scale + (config.init_noise_scale - config.final_noise_scale) * explr_pct_remaining)
        maddpg.reset_noise()

        for et_i in range(config.episode_length):
            # rearrange observations to be per agent, and convert to torch Variable
            torch_obs = [Variable(torch.Tensor(np.vstack(obs[:, i])),
                                  requires_grad=False)
                         for i in range(maddpg.nagents)]
            # get actions as torch Variables
            torch_agent_actions = maddpg.step(torch_obs, explore=True)
            # convert actions to numpy arrays
            agent_actions = [ac.data.numpy() for ac in torch_agent_actions]
            # rearrange actions to be per environment
            actions = [[ac[i] for ac in agent_actions] for i in range(config.n_rollout_threads)]
            next_obs, rewards, dones, infos = env.step(actions, maddpg)

            '''
            Reward Shaping using D++, D.
            The rewards now contain global as well as shaped rewards
            Keep the global for logging, and use the shaped rewards for updates
            '''
            # Choose which reward to use
            use_dpp = True

            # DIFFERENCE REWARDS
            d_rewards = []
            for n in range(maddpg.nagents):
                d_rewards.append([rewards[0][n][1]])
            d_rewards = [d_rewards]
            d_rewards = np.array(d_rewards)

            # GLOBAL REWARDS
            g_rewards = []
            for n in range(maddpg.nagents):
                g_rewards.append([rewards[0][n][0]])
            g_rewards = [g_rewards]
            g_rewards = np.array(g_rewards)

            if use_dpp:
                rewards = d_rewards
            else:
                rewards = g_rewards
            # ----------------------------------------------------------- #
            # Buffer used for updates
            replay_buffer.push(obs, agent_actions, rewards, next_obs, dones)
            # push global rewards into g_replay_buffer for plotting
            g_storage_buffer.push(obs, agent_actions, g_rewards, next_obs, dones)

            obs = next_obs
            t += config.n_rollout_threads
            if (len(replay_buffer) >= config.batch_size and
                (t % config.steps_per_update) < config.n_rollout_threads):
                if USE_CUDA:
                    maddpg.prep_training(device='gpu')
                else:
                    maddpg.prep_training(device='cpu')
                for u_i in range(config.n_rollout_threads):
                    for a_i in range(maddpg.nagents):
                        sample = replay_buffer.sample(config.batch_size,
                                                      to_gpu=USE_CUDA)
                        maddpg.update(sample, a_i, logger=logger)
                    maddpg.update_all_targets()
                maddpg.prep_rollouts(device='cpu')
        # Take out global reward from g_storage_buffer
        ep_rews = g_storage_buffer.get_average_rewards(
            config.episode_length * config.n_rollout_threads)

        for a_i, a_ep_rew in enumerate(ep_rews):
            logger.add_scalar('agent%i/mean_episode_rewards' % a_i, a_ep_rew, ep_i)

        if ep_i % config.save_interval < config.n_rollout_threads:
            os.makedirs(run_dir / 'incremental', exist_ok=True)
            maddpg.save(run_dir / 'incremental' / ('model_ep%i.pt' % (ep_i + 1)))
            maddpg.save(run_dir / 'model.pt')

    maddpg.save(run_dir / 'model.pt')
    env.close()
    logger.export_scalars_to_json(str(log_dir / 'summary.json'))
    logger.close()


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument("--env_id", default='simple_spread', help="Name of environment")
    parser.add_argument("--model_name", default='Exp',
                        help="Name of directory to store " +
                             "model/training contents")
    parser.add_argument("--seed",
                        default=1, type=int,
                        help="Random seed")
    parser.add_argument("--n_rollout_threads", default=1, type=int)
    parser.add_argument("--n_training_threads", default=6, type=int)
    parser.add_argument("--buffer_length", default=int(1e6), type=int)
    parser.add_argument("--n_episodes", default=100000, type=int)
    parser.add_argument("--episode_length", default=60, type=int)
    parser.add_argument("--steps_per_update", default=100, type=int)
    parser.add_argument("--batch_size",
                        default=1024, type=int,
                        help="Batch size for model training")
    parser.add_argument("--n_exploration_eps", default=100000, type=int)
    parser.add_argument("--init_noise_scale", default=0.3, type=float)
    parser.add_argument("--final_noise_scale", default=0.0, type=float)
    parser.add_argument("--save_interval", default=2000, type=int)
    parser.add_argument("--hidden_dim", default=64, type=int)
    parser.add_argument("--lr", default=0.01, type=float)
    parser.add_argument("--tau", default=0.01, type=float)
    parser.add_argument("--agent_alg",
                        default="MADDPG", type=str,
                        choices=['MADDPG', 'DDPG'])
    parser.add_argument("--adversary_alg",
                        default="MADDPG", type=str,
                        choices=['MADDPG', 'DDPG'])
    parser.add_argument("--discrete_action",
                        action='store_true')

    config = parser.parse_args()

    # INITIALIZE FROM SAVED MODEL ?
    init_from_saved = False
    config.discrete_action = False
    model_path = ""

    run(config)