def test_ppo(args=get_args()):
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 = 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.state_shape, hidden_sizes=args.hidden_sizes,
device=args.device)
actor = ActorProb(net, args.action_shape, max_action=args.max_action,
device=args.device).to(args.device)
critic = Critic(Net(
args.state_shape, hidden_sizes=args.hidden_sizes, device=args.device
), device=args.device).to(args.device)
# orthogonal initialization
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(set(
actor.parameters()).union(critic.parameters()), lr=args.lr)
# replace DiagGuassian with Independent(Normal) which is equivalent
# pass *logits to be consistent with policy.forward
def dist(*logits):
return Independent(Normal(*logits), 1)
policy = PPOPolicy(
actor, critic, optim, dist,
discount_factor=args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
advantage_normalization=args.norm_adv,
recompute_advantage=args.recompute_adv,
# dual_clip=args.dual_clip,
# dual clip cause monotonically increasing log_std :)
value_clip=args.value_clip,
gae_lambda=args.gae_lambda,
action_space=env.action_space)
# collector
train_collector = Collector(
policy, train_envs,
VectorReplayBuffer(args.buffer_size, len(train_envs)),
exploration_noise=True)
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'ppo')
writer = SummaryWriter(log_path)
logger = BasicLogger(writer)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(mean_rewards):
return mean_rewards >= env.spec.reward_threshold
# trainer
result = onpolicy_trainer(
policy, train_collector, test_collector, args.epoch,
args.step_per_epoch, args.repeat_per_collect, args.test_num, args.batch_size,
episode_per_collect=args.episode_per_collect, stop_fn=stop_fn, save_fn=save_fn,
logger=logger)
assert stop_fn(result['best_reward'])
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
policy.eval()
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
rews, lens = result["rews"], result["lens"]
print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
def test_reinforce(args=get_args()):
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
args.max_action = env.action_space.high[0]
print("Observations shape:", args.state_shape)
print("Actions shape:", args.action_shape)
print("Action range:", np.min(env.action_space.low),
np.max(env.action_space.high))
# train_envs = gym.make(args.task)
train_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)],
norm_obs=True)
# test_envs = gym.make(args.task)
test_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.test_num)],
norm_obs=True,
obs_rms=train_envs.obs_rms,
update_obs_rms=False)
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
# model
net_a = Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
activation=nn.Tanh,
device=args.device)
actor = ActorProb(net_a,
args.action_shape,
max_action=args.max_action,
unbounded=True,
device=args.device).to(args.device)
torch.nn.init.constant_(actor.sigma_param, -0.5)
for m in actor.modules():
if isinstance(m, torch.nn.Linear):
# orthogonal initialization
torch.nn.init.orthogonal_(m.weight, gain=np.sqrt(2))
torch.nn.init.zeros_(m.bias)
# do last policy layer scaling, this will make initial actions have (close to)
# 0 mean and std, and will help boost performances,
# see https://arxiv.org/abs/2006.05990, Fig.24 for details
for m in actor.mu.modules():
if isinstance(m, torch.nn.Linear):
torch.nn.init.zeros_(m.bias)
m.weight.data.copy_(0.01 * m.weight.data)
optim = torch.optim.Adam(actor.parameters(), lr=args.lr)
lr_scheduler = None
if args.lr_decay:
# decay learning rate to 0 linearly
max_update_num = np.ceil(
args.step_per_epoch / args.step_per_collect) * args.epoch
lr_scheduler = LambdaLR(
optim, lr_lambda=lambda epoch: 1 - epoch / max_update_num)
def dist(*logits):
return Independent(Normal(*logits), 1)
policy = PGPolicy(actor,
optim,
dist,
discount_factor=args.gamma,
reward_normalization=args.rew_norm,
action_scaling=True,
action_bound_method=args.action_bound_method,
lr_scheduler=lr_scheduler,
action_space=env.action_space)
# load a previous policy
if args.resume_path:
policy.load_state_dict(
torch.load(args.resume_path, map_location=args.device))
print("Loaded agent from: ", args.resume_path)
# collector
if args.training_num > 1:
buffer = VectorReplayBuffer(args.buffer_size, len(train_envs))
else:
buffer = ReplayBuffer(args.buffer_size)
train_collector = Collector(policy,
train_envs,
buffer,
exploration_noise=True)
test_collector = Collector(policy, test_envs)
# log
t0 = datetime.datetime.now().strftime("%m%d_%H%M%S")
log_file = f'seed_{args.seed}_{t0}-{args.task.replace("-", "_")}_reinforce'
log_path = os.path.join(args.logdir, args.task, 'reinforce', log_file)
writer = SummaryWriter(log_path)
writer.add_text("args", str(args))
logger = BasicLogger(writer, update_interval=10, train_interval=100)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
if not args.watch:
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.repeat_per_collect,
args.test_num,
args.batch_size,
step_per_collect=args.step_per_collect,
save_fn=save_fn,
logger=logger,
test_in_train=False)
pprint.pprint(result)
# Let's watch its performance!
policy.eval()
test_envs.seed(args.seed)
test_collector.reset()
result = test_collector.collect(n_episode=args.test_num,
render=args.render)
print(
f'Final reward: {result["rews"].mean()}, length: {result["lens"].mean()}'
)
def test_trpo(args=get_args()):
env, train_envs, test_envs = make_mujoco_env(args.task,
args.seed,
args.training_num,
args.test_num,
obs_norm=True)
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]
print("Observations shape:", args.state_shape)
print("Actions shape:", args.action_shape)
print("Action range:", np.min(env.action_space.low),
np.max(env.action_space.high))
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
# model
net_a = Net(
args.state_shape,
hidden_sizes=args.hidden_sizes,
activation=nn.Tanh,
device=args.device,
)
actor = ActorProb(
net_a,
args.action_shape,
max_action=args.max_action,
unbounded=True,
device=args.device,
).to(args.device)
net_c = Net(
args.state_shape,
hidden_sizes=args.hidden_sizes,
activation=nn.Tanh,
device=args.device,
)
critic = Critic(net_c, device=args.device).to(args.device)
torch.nn.init.constant_(actor.sigma_param, -0.5)
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
# orthogonal initialization
torch.nn.init.orthogonal_(m.weight, gain=np.sqrt(2))
torch.nn.init.zeros_(m.bias)
# do last policy layer scaling, this will make initial actions have (close to)
# 0 mean and std, and will help boost performances,
# see https://arxiv.org/abs/2006.05990, Fig.24 for details
for m in actor.mu.modules():
if isinstance(m, torch.nn.Linear):
torch.nn.init.zeros_(m.bias)
m.weight.data.copy_(0.01 * m.weight.data)
optim = torch.optim.Adam(critic.parameters(), lr=args.lr)
lr_scheduler = None
if args.lr_decay:
# decay learning rate to 0 linearly
max_update_num = np.ceil(
args.step_per_epoch / args.step_per_collect) * args.epoch
lr_scheduler = LambdaLR(
optim, lr_lambda=lambda epoch: 1 - epoch / max_update_num)
def dist(*logits):
return Independent(Normal(*logits), 1)
policy = TRPOPolicy(
actor,
critic,
optim,
dist,
discount_factor=args.gamma,
gae_lambda=args.gae_lambda,
reward_normalization=args.rew_norm,
action_scaling=True,
action_bound_method=args.bound_action_method,
lr_scheduler=lr_scheduler,
action_space=env.action_space,
advantage_normalization=args.norm_adv,
optim_critic_iters=args.optim_critic_iters,
max_kl=args.max_kl,
backtrack_coeff=args.backtrack_coeff,
max_backtracks=args.max_backtracks,
)
# load a previous policy
if args.resume_path:
ckpt = torch.load(args.resume_path, map_location=args.device)
policy.load_state_dict(ckpt["model"])
train_envs.set_obs_rms(ckpt["obs_rms"])
test_envs.set_obs_rms(ckpt["obs_rms"])
print("Loaded agent from: ", args.resume_path)
# collector
if args.training_num > 1:
buffer = VectorReplayBuffer(args.buffer_size, len(train_envs))
else:
buffer = ReplayBuffer(args.buffer_size)
train_collector = Collector(policy,
train_envs,
buffer,
exploration_noise=True)
test_collector = Collector(policy, test_envs)
# log
now = datetime.datetime.now().strftime("%y%m%d-%H%M%S")
args.algo_name = "trpo"
log_name = os.path.join(args.task, args.algo_name, str(args.seed), now)
log_path = os.path.join(args.logdir, log_name)
# logger
if args.logger == "wandb":
logger = WandbLogger(
save_interval=1,
name=log_name.replace(os.path.sep, "__"),
run_id=args.resume_id,
config=args,
project=args.wandb_project,
)
writer = SummaryWriter(log_path)
writer.add_text("args", str(args))
if args.logger == "tensorboard":
logger = TensorboardLogger(writer)
else: # wandb
logger.load(writer)
def save_best_fn(policy):
state = {
"model": policy.state_dict(),
"obs_rms": train_envs.get_obs_rms()
}
torch.save(state, os.path.join(log_path, "policy.pth"))
if not args.watch:
# trainer
result = onpolicy_trainer(
policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.repeat_per_collect,
args.test_num,
args.batch_size,
step_per_collect=args.step_per_collect,
save_best_fn=save_best_fn,
logger=logger,
test_in_train=False,
)
pprint.pprint(result)
# Let's watch its performance!
policy.eval()
test_envs.seed(args.seed)
test_collector.reset()
result = test_collector.collect(n_episode=args.test_num,
render=args.render)
print(
f'Final reward: {result["rews"].mean()}, length: {result["lens"].mean()}'
)
def test_acktr(args=get_args()):
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 = 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.state_shape,
hidden_sizes=args.hidden_sizes,
activation=nn.Tanh,
device=args.device)
actor = ActorProb(net,
args.action_shape,
max_action=args.max_action,
unbounded=True,
device=args.device).to(args.device)
critic = Critic(Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device,
activation=nn.Tanh),
device=args.device).to(args.device)
torch.nn.init.constant_(actor.sigma_param._bias, -0.5)
# orthogonal initialization
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = KFACOptimizer(actor, critic, lr=0.25)
# replace DiagGuassian with Independent(Normal) which is equivalent
# pass *logits to be consistent with policy.forward
def dist(*logits):
return Independent(Normal(*logits), 1)
policy = ACKTRPolicy(actor,
critic,
optim,
dist,
discount_factor=args.gamma,
reward_normalization=args.rew_norm,
advantage_normalization=args.norm_adv,
gae_lambda=args.gae_lambda,
action_space=env.action_space)
# collector
train_collector = Collector(
policy, train_envs,
VectorReplayBuffer(args.buffer_size, len(train_envs)))
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'acktr')
writer = SummaryWriter(log_path)
logger = BasicLogger(writer)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(mean_rewards):
return mean_rewards >= env.spec.reward_threshold
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.repeat_per_collect,
args.test_num,
args.batch_size,
step_per_collect=args.step_per_collect,
stop_fn=stop_fn,
save_fn=save_fn,
logger=logger)
assert stop_fn(result['best_reward'])
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
policy.eval()
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
rews, lens = result["rews"], result["lens"]
print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
def get_agents(
args: argparse.Namespace = get_args(),
agents: Optional[List[BasePolicy]] = None,
optims: Optional[List[torch.optim.Optimizer]] = None,
) -> Tuple[BasePolicy, List[torch.optim.Optimizer], List]:
env = get_env()
observation_space = env.observation_space['observation'] if isinstance(
env.observation_space, gym.spaces.Dict) else env.observation_space
args.state_shape = observation_space.shape or observation_space.n
args.action_shape = env.action_space.shape or env.action_space.n
args.max_action = env.action_space.high[0]
if agents is None:
agents = []
optims = []
for _ in range(args.n_pistons):
# model
net = DQN(observation_space.shape[2],
observation_space.shape[1],
observation_space.shape[0],
device=args.device).to(args.device)
actor = ActorProb(net,
args.action_shape,
max_action=args.max_action,
device=args.device).to(args.device)
net2 = DQN(observation_space.shape[2],
observation_space.shape[1],
observation_space.shape[0],
device=args.device).to(args.device)
critic = Critic(net2, device=args.device).to(args.device)
for m in set(actor.modules()).union(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(set(actor.parameters()).union(
critic.parameters()),
lr=args.lr)
def dist(*logits):
return Independent(Normal(*logits), 1)
agent = PPOPolicy(
actor,
critic,
optim,
dist,
discount_factor=args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
advantage_normalization=args.norm_adv,
recompute_advantage=args.recompute_adv,
# dual_clip=args.dual_clip,
# dual clip cause monotonically increasing log_std :)
value_clip=args.value_clip,
gae_lambda=args.gae_lambda,
action_space=env.action_space)
agents.append(agent)
optims.append(optim)
policy = MultiAgentPolicyManager(agents,
env,
action_scaling=True,
action_bound_method='clip')
return policy, optims, env.agents
def test_ppo(args=get_args()):
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
net = Net(args.layer_num, args.state_shape, device=args.device)
actor = ActorProb(net, args.action_shape, args.max_action,
args.device).to(args.device)
critic = Critic(Net(args.layer_num, args.state_shape, device=args.device),
device=args.device).to(args.device)
# orthogonal initialization
for m in list(actor.modules()) + list(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(list(actor.parameters()) +
list(critic.parameters()),
lr=args.lr)
dist = DiagGaussian
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,
reward_normalization=args.rew_norm,
# dual_clip=args.dual_clip,
# dual clip cause monotonically increasing log_std :)
value_clip=args.value_clip,
# action_range=[env.action_space.low[0], env.action_space.high[0]],)
# if clip the action, ppo would not converge :)
gae_lambda=args.gae_lambda)
# 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, '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()
def test_ppo(args=get_args()):
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 = 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.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device)
actor = ActorProb(net,
args.action_shape,
max_action=args.max_action,
device=args.device).to(args.device)
critic = Critic(Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device),
device=args.device).to(args.device)
# orthogonal initialization
for m in set(actor.modules()).union(critic.modules()):
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(set(actor.parameters()).union(
critic.parameters()),
lr=args.lr)
# replace DiagGuassian with Independent(Normal) which is equivalent
# pass *logits to be consistent with policy.forward
def dist(*logits):
return Independent(Normal(*logits), 1)
policy = PPOPolicy(
actor,
critic,
optim,
dist,
discount_factor=args.gamma,
max_grad_norm=args.max_grad_norm,
eps_clip=args.eps_clip,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
advantage_normalization=args.norm_adv,
recompute_advantage=args.recompute_adv,
# dual_clip=args.dual_clip,
# dual clip cause monotonically increasing log_std :)
value_clip=args.value_clip,
gae_lambda=args.gae_lambda,
action_space=env.action_space)
# collector
train_collector = Collector(
policy, train_envs,
VectorReplayBuffer(args.buffer_size, len(train_envs)))
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'ppo')
writer = SummaryWriter(log_path)
logger = TensorboardLogger(writer, save_interval=args.save_interval)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(mean_rewards):
return mean_rewards >= env.spec.reward_threshold
def save_checkpoint_fn(epoch, env_step, gradient_step):
# see also: https://pytorch.org/tutorials/beginner/saving_loading_models.html
torch.save(
{
'model': policy.state_dict(),
'optim': optim.state_dict(),
}, os.path.join(log_path, 'checkpoint.pth'))
if args.resume:
# load from existing checkpoint
print(f"Loading agent under {log_path}")
ckpt_path = os.path.join(log_path, 'checkpoint.pth')
if os.path.exists(ckpt_path):
checkpoint = torch.load(ckpt_path, map_location=args.device)
policy.load_state_dict(checkpoint['model'])
optim.load_state_dict(checkpoint['optim'])
print("Successfully restore policy and optim.")
else:
print("Fail to restore policy and optim.")
# trainer
result = onpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.repeat_per_collect,
args.test_num,
args.batch_size,
episode_per_collect=args.episode_per_collect,
stop_fn=stop_fn,
save_fn=save_fn,
logger=logger,
resume_from_log=args.resume,
save_checkpoint_fn=save_checkpoint_fn)
assert stop_fn(result['best_reward'])
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
policy.eval()
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
rews, lens = result["rews"], result["lens"]
print(f"Final reward: {rews.mean()}, length: {lens.mean()}")