Tianshou/test/continuous/test_ppo.py

import os
import gym
import torch
import pprint
import argparse
import numpy as np
from torch.utils.tensorboard import SummaryWriter

from tianshou.policy import PPOPolicy
from tianshou.env import VectorEnv
from tianshou.trainer import onpolicy_trainer
from tianshou.data import Collector, ReplayBuffer

if __name__ == '__main__':
    from net import ActorProb, Critic
else:  # pytest
    from test.continuous.net import ActorProb, Critic


def get_args():
    parser = argparse.ArgumentParser()
    parser.add_argument('--task', type=str, default='Pendulum-v0')
    parser.add_argument('--seed', type=int, default=0)
    parser.add_argument('--buffer-size', type=int, default=20000)
    parser.add_argument('--lr', type=float, default=3e-4)
    parser.add_argument('--gamma', type=float, default=0.9)
    parser.add_argument('--epoch', type=int, default=20)
    parser.add_argument('--step-per-epoch', type=int, default=1000)
    parser.add_argument('--collect-per-step', type=int, default=10)
    parser.add_argument('--repeat-per-collect', type=int, default=10)
    parser.add_argument('--batch-size', type=int, default=64)
    parser.add_argument('--layer-num', type=int, default=1)
    parser.add_argument('--training-num', type=int, default=16)
    parser.add_argument('--test-num', type=int, default=100)
    parser.add_argument('--logdir', type=str, default='log')
    parser.add_argument('--render', type=float, default=0.)
    parser.add_argument(
        '--device', type=str,
        default='cuda' if torch.cuda.is_available() else 'cpu')
    # ppo special
    parser.add_argument('--vf-coef', type=float, default=0.5)
    parser.add_argument('--ent-coef', type=float, default=0.0)
    parser.add_argument('--eps-clip', type=float, default=0.2)
    parser.add_argument('--max-grad-norm', type=float, default=0.5)
    args = parser.parse_known_args()[0]
    return args


def _test_ppo(args=get_args()):
    # just a demo, I have not made it work :(
    torch.set_num_threads(1)  # we just need only one thread for NN
    env = gym.make(args.task)
    if args.task == 'Pendulum-v0':
        env.spec.reward_threshold = -250
    args.state_shape = env.observation_space.shape or env.observation_space.n
    args.action_shape = env.action_space.shape or env.action_space.n
    args.max_action = env.action_space.high[0]
    # you can also use tianshou.env.SubprocVectorEnv
    # train_envs = gym.make(args.task)
    train_envs = VectorEnv(
        [lambda: gym.make(args.task) for _ in range(args.training_num)])
    # test_envs = gym.make(args.task)
    test_envs = VectorEnv(
        [lambda: gym.make(args.task) for _ in range(args.test_num)])
    # seed
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    train_envs.seed(args.seed)
    test_envs.seed(args.seed)
    # model
    actor = ActorProb(
        args.layer_num, args.state_shape, args.action_shape,
        args.max_action, args.device
    ).to(args.device)
    critic = Critic(
        args.layer_num, args.state_shape, device=args.device
    ).to(args.device)
    optim = torch.optim.Adam(list(
        actor.parameters()) + list(critic.parameters()), lr=args.lr)
    dist = torch.distributions.Normal
    policy = PPOPolicy(
        actor, critic, optim, dist, args.gamma,
        max_grad_norm=args.max_grad_norm,
        eps_clip=args.eps_clip,
        vf_coef=args.vf_coef,
        ent_coef=args.ent_coef,
        action_range=[env.action_space.low[0], env.action_space.high[0]])
    # collector
    train_collector = Collector(
        policy, train_envs, ReplayBuffer(args.buffer_size))
    test_collector = Collector(policy, test_envs)
    train_collector.collect(n_step=args.step_per_epoch)
    # log
    log_path = os.path.join(args.logdir, args.task, 'ppo')
    writer = SummaryWriter(log_path)

    def save_fn(policy):
        torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))

    def stop_fn(x):
        return x >= env.spec.reward_threshold

    # trainer
    result = onpolicy_trainer(
        policy, train_collector, test_collector, args.epoch,
        args.step_per_epoch, args.collect_per_step, args.repeat_per_collect,
        args.test_num, args.batch_size, stop_fn=stop_fn, save_fn=save_fn,
        writer=writer)
    assert stop_fn(result['best_reward'])
    train_collector.close()
    test_collector.close()
    if __name__ == '__main__':
        pprint.pprint(result)
        # Let's watch its performance!
        env = gym.make(args.task)
        collector = Collector(policy, env)
        result = collector.collect(n_episode=1, render=args.render)
        print(f'Final reward: {result["rew"]}, length: {result["len"]}')
        collector.close()


if __name__ == '__main__':
    _test_ppo()