examples/mujoco/train_ddpg_batch_gym.py

import argparse
import functools
import sys

import chainer
from chainer import optimizers
import gym
from gym import spaces
import numpy as np

import chainerrl
from chainerrl.agents.ddpg import DDPG
from chainerrl.agents.ddpg import DDPGModel
from chainerrl import experiments
from chainerrl import explorers
from chainerrl import misc
from chainerrl import policy
from chainerrl import q_functions
from chainerrl import replay_buffer


def main():
    import logging
    logging.basicConfig(level=logging.DEBUG)

    parser = argparse.ArgumentParser()
    parser.add_argument('--outdir', type=str, default='results',
                        help='Directory path to save output files.'
                             ' If it does not exist, it will be created.')
    parser.add_argument('--env', type=str, default='Humanoid-v2')
    parser.add_argument('--seed', type=int, default=0,
                        help='Random seed [0, 2 ** 32)')
    parser.add_argument('--gpu', type=int, default=0)
    parser.add_argument('--final-exploration-steps',
                        type=int, default=10 ** 6)
    parser.add_argument('--actor-lr', type=float, default=1e-4)
    parser.add_argument('--critic-lr', type=float, default=1e-3)
    parser.add_argument('--load', type=str, default='')
    parser.add_argument('--steps', type=int, default=10 ** 7)
    parser.add_argument('--n-hidden-channels', type=int, default=300)
    parser.add_argument('--n-hidden-layers', type=int, default=3)
    parser.add_argument('--replay-start-size', type=int, default=5000)
    parser.add_argument('--n-update-times', type=int, default=1)
    parser.add_argument('--target-update-interval',
                        type=int, default=1)
    parser.add_argument('--target-update-method',
                        type=str, default='soft', choices=['hard', 'soft'])
    parser.add_argument('--soft-update-tau', type=float, default=1e-2)
    parser.add_argument('--update-interval', type=int, default=4)
    parser.add_argument('--eval-n-runs', type=int, default=100)
    parser.add_argument('--eval-interval', type=int, default=10 ** 5)
    parser.add_argument('--gamma', type=float, default=0.995)
    parser.add_argument('--minibatch-size', type=int, default=200)
    parser.add_argument('--render', action='store_true')
    parser.add_argument('--demo', action='store_true')
    parser.add_argument('--use-bn', action='store_true', default=False)
    parser.add_argument('--monitor', action='store_true')
    parser.add_argument('--reward-scale-factor', type=float, default=1e-2)
    parser.add_argument('--num-envs', type=int, default=1)
    args = parser.parse_args()

    args.outdir = experiments.prepare_output_dir(
        args, args.outdir, argv=sys.argv)
    print('Output files are saved in {}'.format(args.outdir))

    # Set a random seed used in ChainerRL
    misc.set_random_seed(args.seed, gpus=(args.gpu,))

    def clip_action_filter(a):
        return np.clip(a, action_space.low, action_space.high)

    def reward_filter(r):
        return r * args.reward_scale_factor

    # Set different random seeds for different subprocesses.
    # If seed=0 and processes=4, subprocess seeds are [0, 1, 2, 3].
    # If seed=1 and processes=4, subprocess seeds are [4, 5, 6, 7].
    process_seeds = np.arange(args.num_envs) + args.seed * args.num_envs
    assert process_seeds.max() < 2 ** 32

    def make_env(idx, test):
        env = gym.make(args.env)
        # Use different random seeds for train and test envs
        process_seed = int(process_seeds[idx])
        env_seed = 2 ** 32 - 1 - process_seed if test else process_seed
        env.seed(env_seed)
        # Cast observations to float32 because our model uses float32
        env = chainerrl.wrappers.CastObservationToFloat32(env)
        if args.monitor:
            env = chainerrl.wrappers.Monitor(env, args.outdir)
        if isinstance(env.action_space, spaces.Box):
            misc.env_modifiers.make_action_filtered(env, clip_action_filter)
        if not test:
            # Scale rewards (and thus returns) to a reasonable range so that
            # training is easier
            env = chainerrl.wrappers.ScaleReward(env, args.reward_scale_factor)
        if args.render and not test:
            env = chainerrl.wrappers.Render(env)
        return env

    def make_batch_env(test):
        return chainerrl.envs.MultiprocessVectorEnv(
            [functools.partial(make_env, idx, test)
             for idx, env in enumerate(range(args.num_envs))])

    sample_env = make_env(0, test=False)
    timestep_limit = sample_env.spec.max_episode_steps

    obs_size = np.asarray(sample_env.observation_space.shape).prod()
    action_space = sample_env.action_space

    action_size = np.asarray(action_space.shape).prod()
    if args.use_bn:
        q_func = q_functions.FCBNLateActionSAQFunction(
            obs_size, action_size,
            n_hidden_channels=args.n_hidden_channels,
            n_hidden_layers=args.n_hidden_layers,
            normalize_input=True)
        pi = policy.FCBNDeterministicPolicy(
            obs_size, action_size=action_size,
            n_hidden_channels=args.n_hidden_channels,
            n_hidden_layers=args.n_hidden_layers,
            min_action=action_space.low, max_action=action_space.high,
            bound_action=True,
            normalize_input=True)
    else:
        q_func = q_functions.FCSAQFunction(
            obs_size, action_size,
            n_hidden_channels=args.n_hidden_channels,
            n_hidden_layers=args.n_hidden_layers)
        pi = policy.FCDeterministicPolicy(
            obs_size, action_size=action_size,
            n_hidden_channels=args.n_hidden_channels,
            n_hidden_layers=args.n_hidden_layers,
            min_action=action_space.low, max_action=action_space.high,
            bound_action=True)
    model = DDPGModel(q_func=q_func, policy=pi)
    opt_a = optimizers.Adam(alpha=args.actor_lr)
    opt_c = optimizers.Adam(alpha=args.critic_lr)
    opt_a.setup(model['policy'])
    opt_c.setup(model['q_function'])
    opt_a.add_hook(chainer.optimizer.GradientClipping(1.0), 'hook_a')
    opt_c.add_hook(chainer.optimizer.GradientClipping(1.0), 'hook_c')

    rbuf = replay_buffer.ReplayBuffer(5 * 10 ** 5)

    def random_action():
        a = action_space.sample()
        if isinstance(a, np.ndarray):
            a = a.astype(np.float32)
        return a

    ou_sigma = (action_space.high - action_space.low) * 0.2
    explorer = explorers.AdditiveOU(sigma=ou_sigma)
    agent = DDPG(model, opt_a, opt_c, rbuf, gamma=args.gamma,
                 explorer=explorer, replay_start_size=args.replay_start_size,
                 target_update_method=args.target_update_method,
                 target_update_interval=args.target_update_interval,
                 update_interval=args.update_interval,
                 soft_update_tau=args.soft_update_tau,
                 n_times_update=args.n_update_times,
                 gpu=args.gpu, minibatch_size=args.minibatch_size)

    if len(args.load) > 0:
        agent.load(args.load)

    if args.demo:
        eval_stats = experiments.eval_performance(
            env=make_batch_env(test=True),
            agent=agent,
            n_steps=None,
            n_episodes=args.eval_n_runs,
            max_episode_len=timestep_limit)
        print('n_runs: {} mean: {} median: {} stdev {}'.format(
            args.eval_n_runs, eval_stats['mean'], eval_stats['median'],
            eval_stats['stdev']))
    else:
        experiments.train_agent_batch_with_evaluation(
            agent=agent, env=make_batch_env(test=False), steps=args.steps,
            eval_env=make_batch_env(test=True), eval_n_steps=None,
            eval_n_episodes=args.eval_n_runs, eval_interval=args.eval_interval,
            outdir=args.outdir,
            max_episode_len=timestep_limit)


if __name__ == '__main__':
    main()