Tianshou/test/discrete/test_a2c_with_il.py

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

from tianshou.env import DummyVectorEnv
from tianshou.data import Collector, ReplayBuffer
from tianshou.policy import A2CPolicy, ImitationPolicy
from tianshou.trainer import onpolicy_trainer, offpolicy_trainer
from tianshou.utils.net.discrete import Actor, Critic
from tianshou.utils.net.common import Net


def get_args():
    parser = argparse.ArgumentParser()
    parser.add_argument('--task', type=str, default='CartPole-v0')
    parser.add_argument('--seed', type=int, default=1)
    parser.add_argument('--buffer-size', type=int, default=20000)
    parser.add_argument('--lr', type=float, default=3e-4)
    parser.add_argument('--il-lr', type=float, default=1e-3)
    parser.add_argument('--gamma', type=float, default=0.9)
    parser.add_argument('--epoch', type=int, default=10)
    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=1)
    parser.add_argument('--batch-size', type=int, default=64)
    parser.add_argument('--layer-num', type=int, default=2)
    parser.add_argument('--training-num', type=int, default=8)
    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')
    # a2c special
    parser.add_argument('--vf-coef', type=float, default=0.5)
    parser.add_argument('--ent-coef', type=float, default=0.0)
    parser.add_argument('--max-grad-norm', type=float, default=None)
    parser.add_argument('--gae-lambda', type=float, default=1.)
    parser.add_argument('--rew-norm', type=bool, default=False)
    args = parser.parse_known_args()[0]
    return args


def test_a2c_with_il(args=get_args()):
    torch.set_num_threads(1)  # for poor CPU
    env = gym.make(args.task)
    args.state_shape = env.observation_space.shape or env.observation_space.n
    args.action_shape = env.action_space.shape or env.action_space.n
    # you can also use tianshou.env.SubprocVectorEnv
    # train_envs = gym.make(args.task)
    train_envs = DummyVectorEnv(
        [lambda: gym.make(args.task) for _ in range(args.training_num)])
    # test_envs = gym.make(args.task)
    test_envs = DummyVectorEnv(
        [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
    net = Net(args.layer_num, args.state_shape, device=args.device)
    actor = Actor(net, args.action_shape).to(args.device)
    critic = Critic(net).to(args.device)
    optim = torch.optim.Adam(list(
        actor.parameters()) + list(critic.parameters()), lr=args.lr)
    dist = torch.distributions.Categorical
    policy = A2CPolicy(
        actor, critic, optim, dist, args.gamma, gae_lambda=args.gae_lambda,
        vf_coef=args.vf_coef, ent_coef=args.ent_coef,
        max_grad_norm=args.max_grad_norm, reward_normalization=args.rew_norm)
    # collector
    train_collector = Collector(
        policy, train_envs, ReplayBuffer(args.buffer_size))
    test_collector = Collector(policy, test_envs)
    # log
    log_path = os.path.join(args.logdir, args.task, 'a2c')
    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'])
    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"]}')

    # here we define an imitation collector with a trivial policy
    if args.task == 'CartPole-v0':
        env.spec.reward_threshold = 190  # lower the goal
    net = Net(1, args.state_shape, device=args.device)
    net = Actor(net, args.action_shape).to(args.device)
    optim = torch.optim.Adam(net.parameters(), lr=args.il_lr)
    il_policy = ImitationPolicy(net, optim, mode='discrete')
    il_test_collector = Collector(
        il_policy,
        DummyVectorEnv(
            [lambda: gym.make(args.task) for _ in range(args.test_num)])
    )
    train_collector.reset()
    result = offpolicy_trainer(
        il_policy, train_collector, il_test_collector, args.epoch,
        args.step_per_epoch, args.collect_per_step, args.test_num,
        args.batch_size, stop_fn=stop_fn, save_fn=save_fn, writer=writer)
    assert stop_fn(result['best_reward'])
    if __name__ == '__main__':
        pprint.pprint(result)
        # Let's watch its performance!
        env = gym.make(args.task)
        collector = Collector(il_policy, env)
        result = collector.collect(n_episode=1, render=args.render)
        print(f'Final reward: {result["rew"]}, length: {result["len"]}')


if __name__ == '__main__':
    test_a2c_with_il()