def gather_data():
"""Return expert buffer data."""
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]
if args.reward_threshold is None:
default_reward_threshold = {"Pendulum-v0": -250, "Pendulum-v1": -250}
args.reward_threshold = default_reward_threshold.get(
args.task, env.spec.reward_threshold)
# 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,
unbounded=True,
).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
net_c1 = Net(
args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
concat=True,
device=args.device,
)
critic1 = Critic(net_c1, device=args.device).to(args.device)
critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
net_c2 = Net(
args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
concat=True,
device=args.device,
)
critic2 = Critic(net_c2, device=args.device).to(args.device)
critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
if args.auto_alpha:
target_entropy = -np.prod(env.action_space.shape)
log_alpha = torch.zeros(1, requires_grad=True, device=args.device)
alpha_optim = torch.optim.Adam([log_alpha], lr=args.alpha_lr)
args.alpha = (target_entropy, log_alpha, alpha_optim)
policy = SACPolicy(
actor,
actor_optim,
critic1,
critic1_optim,
critic2,
critic2_optim,
tau=args.tau,
gamma=args.gamma,
alpha=args.alpha,
reward_normalization=args.rew_norm,
estimation_step=args.n_step,
action_space=env.action_space,
)
# collector
buffer = VectorReplayBuffer(args.buffer_size, len(train_envs))
train_collector = Collector(policy,
train_envs,
buffer,
exploration_noise=True)
test_collector = Collector(policy, test_envs)
# train_collector.collect(n_step=args.buffer_size)
# log
log_path = os.path.join(args.logdir, args.task, 'sac')
writer = SummaryWriter(log_path)
logger = TensorboardLogger(writer)
def save_best_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(mean_rewards):
return mean_rewards >= args.reward_threshold
# trainer
offpolicy_trainer(
policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.step_per_collect,
args.test_num,
args.batch_size,
update_per_step=args.update_per_step,
save_best_fn=save_best_fn,
stop_fn=stop_fn,
logger=logger,
)
train_collector.reset()
result = train_collector.collect(n_step=args.buffer_size)
rews, lens = result["rews"], result["lens"]
print(f"Final reward: {rews.mean()}, length: {lens.mean()}")
if args.save_buffer_name.endswith(".hdf5"):
buffer.save_hdf5(args.save_buffer_name)
else:
pickle.dump(buffer, open(args.save_buffer_name, "wb"))
return buffer
def onpolicy_trainer(
policy: BasePolicy,
train_collector: Collector,
test_collector: Collector,
max_epoch: int,
step_per_epoch: int,
repeat_per_collect: int,
episode_per_test: int,
batch_size: int,
step_per_collect: Optional[int] = None,
episode_per_collect: Optional[int] = None,
train_fn: Optional[Callable[[int, int], None]] = None,
test_fn: Optional[Callable[[int, Optional[int]], None]] = None,
stop_fn: Optional[Callable[[float], bool]] = None,
save_fn: Optional[Callable[[BasePolicy], None]] = None,
reward_metric: Optional[Callable[[np.ndarray], np.ndarray]] = None,
logger: BaseLogger = LazyLogger(),
verbose: bool = True,
test_in_train: bool = True,
) -> Dict[str, Union[float, str]]:
"""A wrapper for on-policy trainer procedure.
The "step" in trainer means an environment step (a.k.a. transition).
:param policy: an instance of the :class:`~tianshou.policy.BasePolicy` class.
:param Collector train_collector: the collector used for training.
:param Collector test_collector: the collector used for testing.
:param int max_epoch: the maximum number of epochs for training. The training
process might be finished before reaching ``max_epoch`` if ``stop_fn`` is set.
:param int step_per_epoch: the number of transitions collected per epoch.
:param int repeat_per_collect: the number of repeat time for policy learning, for
example, set it to 2 means the policy needs to learn each given batch data
twice.
:param int episode_per_test: the number of episodes for one policy evaluation.
:param int batch_size: the batch size of sample data, which is going to feed in the
policy network.
:param int step_per_collect: the number of transitions the collector would collect
before the network update, i.e., trainer will collect "step_per_collect"
transitions and do some policy network update repeatly in each epoch.
:param int episode_per_collect: the number of episodes the collector would collect
before the network update, i.e., trainer will collect "episode_per_collect"
episodes and do some policy network update repeatly in each epoch.
:param function train_fn: a hook called at the beginning of training in each epoch.
It can be used to perform custom additional operations, with the signature ``f(
num_epoch: int, step_idx: int) -> None``.
:param function test_fn: a hook called at the beginning of testing in each epoch.
It can be used to perform custom additional operations, with the signature ``f(
num_epoch: int, step_idx: int) -> None``.
:param function save_fn: a hook called when the undiscounted average mean reward in
evaluation phase gets better, with the signature ``f(policy: BasePolicy) ->
None``.
:param function stop_fn: a function with signature ``f(mean_rewards: float) ->
bool``, receives the average undiscounted returns of the testing result,
returns a boolean which indicates whether reaching the goal.
:param function reward_metric: a function with signature ``f(rewards: np.ndarray
with shape (num_episode, agent_num)) -> np.ndarray with shape (num_episode,)``,
used in multi-agent RL. We need to return a single scalar for each episode's
result to monitor training in the multi-agent RL setting. This function
specifies what is the desired metric, e.g., the reward of agent 1 or the
average reward over all agents.
:param BaseLogger logger: A logger that logs statistics during
training/testing/updating. Default to a logger that doesn't log anything.
:param bool verbose: whether to print the information. Default to True.
:param bool test_in_train: whether to test in the training phase. Default to True.
:return: See :func:`~tianshou.trainer.gather_info`.
.. note::
Only either one of step_per_collect and episode_per_collect can be specified.
"""
env_step, gradient_step = 0, 0
last_rew, last_len = 0.0, 0
stat: Dict[str, MovAvg] = defaultdict(MovAvg)
start_time = time.time()
train_collector.reset_stat()
test_collector.reset_stat()
test_in_train = test_in_train and train_collector.policy == policy
test_result = test_episode(policy, test_collector, test_fn, 0,
episode_per_test, logger, env_step,
reward_metric)
best_epoch = 0
best_reward, best_reward_std = test_result["rew"], test_result["rew_std"]
for epoch in range(1, 1 + max_epoch):
# train
policy.train()
with tqdm.tqdm(total=step_per_epoch,
desc=f"Epoch #{epoch}",
**tqdm_config) as t:
while t.n < t.total:
if train_fn:
train_fn(epoch, env_step)
result = train_collector.collect(n_step=step_per_collect,
n_episode=episode_per_collect)
if result["n/ep"] > 0 and reward_metric:
result["rews"] = reward_metric(result["rews"])
env_step += int(result["n/st"])
t.update(result["n/st"])
logger.log_train_data(result, env_step)
last_rew = result['rew'] if 'rew' in result else last_rew
last_len = result['len'] if 'len' in result else last_len
data = {
"env_step": str(env_step),
"rew": f"{last_rew:.2f}",
"len": str(int(last_len)),
"n/ep": str(int(result["n/ep"])),
"n/st": str(int(result["n/st"])),
}
if result["n/ep"] > 0:
if test_in_train and stop_fn and stop_fn(result["rew"]):
test_result = test_episode(policy, test_collector,
test_fn, epoch,
episode_per_test, logger,
env_step)
if stop_fn(test_result["rew"]):
if save_fn:
save_fn(policy)
t.set_postfix(**data)
return gather_info(start_time, train_collector,
test_collector,
test_result["rew"],
test_result["rew_std"])
else:
policy.train()
losses = policy.update(0,
train_collector.buffer,
batch_size=batch_size,
repeat=repeat_per_collect)
train_collector.reset_buffer()
step = max(
[1] +
[len(v) for v in losses.values() if isinstance(v, list)])
gradient_step += step
for k in losses.keys():
stat[k].add(losses[k])
losses[k] = stat[k].get()
data[k] = f"{losses[k]:.3f}"
logger.log_update_data(losses, gradient_step)
t.set_postfix(**data)
if t.n <= t.total:
t.update()
# test
test_result = test_episode(policy, test_collector, test_fn, epoch,
episode_per_test, logger, env_step,
reward_metric)
rew, rew_std = test_result["rew"], test_result["rew_std"]
if best_epoch == -1 or best_reward < rew:
best_reward, best_reward_std = rew, rew_std
best_epoch = epoch
if save_fn:
save_fn(policy)
if verbose:
print(
f"Epoch #{epoch}: test_reward: {rew:.6f} ± {rew_std:.6f}, best_reward:"
f" {best_reward:.6f} ± {best_reward_std:.6f} in #{best_epoch}")
if stop_fn and stop_fn(best_reward):
break
return gather_info(start_time, train_collector, test_collector,
best_reward, best_reward_std)
def test_sac(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]
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 = SubprocVectorEnv(
# [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,
unbounded=True).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
critic1 = Critic(args.layer_num, args.state_shape, args.action_shape,
args.device).to(args.device)
critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
critic2 = Critic(args.layer_num, args.state_shape, args.action_shape,
args.device).to(args.device)
critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
policy = SACPolicy(actor,
actor_optim,
critic1,
critic1_optim,
critic2,
critic2_optim,
args.tau,
args.gamma,
args.alpha,
[env.action_space.low[0], env.action_space.high[0]],
reward_normalization=args.rew_norm,
ignore_done=True)
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# train_collector.collect(n_step=args.buffer_size)
# log
log_path = os.path.join(args.logdir, args.task, 'sac')
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 = offpolicy_trainer(policy,
train_collector,
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'])
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_pg(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
# train_envs = gym.make(args.task)
# you can also use tianshou.env.SubprocVectorEnv
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,
args.action_shape,
device=args.device,
softmax=True)
net = net.to(args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.lr)
dist = torch.distributions.Categorical
policy = PGPolicy(net,
optim,
dist,
args.gamma,
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, 'pg')
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_td3(args=get_args()):
torch.set_num_threads(1) # we just need only one thread for NN
env = gym_make()
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.action_range = [env.action_space.low[0], env.action_space.high[0]]
# you can also use tianshou.env.SubprocVectorEnv
# train_envs = gym_make()
train_envs = VectorEnv(
[lambda: gym_make() for _ in range(args.training_num)])
# test_envs = gym_make()
test_envs = VectorEnv([lambda: gym_make() 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
policy = init_policy(args, env)
if args.train == 1:
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# train_collector.collect(n_step=args.buffer_size)
# log
log_path = os.path.join(args.logdir, args.task, 'td3')
writer = SummaryWriter(log_path)
def save_fn(policy):
torch.save(policy.state_dict(),
os.path.join(log_path, 'policy.pth'))
# trainer
result = offpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.collect_per_step,
args.test_num,
args.batch_size,
stop_fn=None,
save_fn=save_fn,
writer=writer)
train_collector.close()
test_collector.close()
else:
log_path = os.path.join(args.logdir, args.task, 'td3')
policy.load_state_dict(
torch.load(os.path.join(log_path, 'policy.pth'),
map_location=args.device))
# Let's watch its performance!
env = gym_make()
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_discrete_sac(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
train_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
test_envs = SubprocVectorEnv(
[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, softmax_output=False).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
net = Net(args.layer_num, args.state_shape, device=args.device)
critic1 = Critic(net, last_size=args.action_shape).to(args.device)
critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
net = Net(args.layer_num, args.state_shape, device=args.device)
critic2 = Critic(net, last_size=args.action_shape).to(args.device)
critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
# better not to use auto alpha in CartPole
if args.auto_alpha:
target_entropy = 0.98 * np.log(np.prod(args.action_shape))
log_alpha = torch.zeros(1, requires_grad=True, device=args.device)
alpha_optim = torch.optim.Adam([log_alpha], lr=args.alpha_lr)
args.alpha = (target_entropy, log_alpha, alpha_optim)
policy = DiscreteSACPolicy(actor,
actor_optim,
critic1,
critic1_optim,
critic2,
critic2_optim,
args.tau,
args.gamma,
args.alpha,
reward_normalization=args.rew_norm,
ignore_done=args.ignore_done)
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# train_collector.collect(n_step=args.buffer_size)
# log
log_path = os.path.join(args.logdir, args.task, 'discrete_sac')
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 = offpolicy_trainer(policy,
train_collector,
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,
test_in_train=False)
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)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
def test_trpo(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)
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:
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("-", "_")}_trpo'
log_path = os.path.join(args.logdir, args.task, 'trpo', log_file)
writer = SummaryWriter(log_path)
writer.add_text("args", str(args))
logger = TensorboardLogger(writer, update_interval=100, 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_collector_with_dict_state():
env = MyTestEnv(size=5, sleep=0, dict_state=True)
policy = MyPolicy(dict_state=True)
c0 = Collector(policy, env, ReplayBuffer(size=100),
Logger.single_preprocess_fn)
c0.collect(n_step=3)
c0.collect(n_episode=2)
assert len(c0.buffer) == 10
env_fns = [
lambda x=i: MyTestEnv(size=x, sleep=0, dict_state=True)
for i in [2, 3, 4, 5]
]
envs = DummyVectorEnv(env_fns)
envs.seed(666)
obs = envs.reset()
assert not np.isclose(obs[0]['rand'], obs[1]['rand'])
c1 = Collector(policy, envs,
VectorReplayBuffer(total_size=100, buffer_num=4),
Logger.single_preprocess_fn)
c1.collect(n_step=12)
result = c1.collect(n_episode=8)
assert result['n/ep'] == 8
lens = np.bincount(result['lens'])
assert result['n/st'] == 21 and np.all(lens == [0, 0, 2, 2, 2, 2]) or \
result['n/st'] == 20 and np.all(lens == [0, 0, 3, 1, 2, 2])
batch, _ = c1.buffer.sample(10)
c0.buffer.update(c1.buffer)
assert len(c0.buffer) in [42, 43]
if len(c0.buffer) == 42:
assert np.all(c0.buffer[:].obs.index[..., 0] == [
0,
1,
2,
3,
4,
0,
1,
2,
3,
4,
0,
1,
0,
1,
0,
1,
0,
1,
0,
1,
2,
0,
1,
2,
0,
1,
2,
3,
0,
1,
2,
3,
0,
1,
2,
3,
4,
0,
1,
2,
3,
4,
]), c0.buffer[:].obs.index[..., 0]
else:
assert np.all(c0.buffer[:].obs.index[..., 0] == [
0,
1,
2,
3,
4,
0,
1,
2,
3,
4,
0,
1,
0,
1,
0,
1,
0,
1,
2,
0,
1,
2,
0,
1,
2,
0,
1,
2,
3,
0,
1,
2,
3,
0,
1,
2,
3,
4,
0,
1,
2,
3,
4,
]), c0.buffer[:].obs.index[..., 0]
c2 = Collector(
policy, envs,
VectorReplayBuffer(total_size=100, buffer_num=4, stack_num=4),
Logger.single_preprocess_fn)
c2.collect(n_episode=10)
batch, _ = c2.buffer.sample(10)
def test_collector_with_ma():
env = MyTestEnv(size=5, sleep=0, ma_rew=4)
policy = MyPolicy()
c0 = Collector(policy, env, ReplayBuffer(size=100),
Logger.single_preprocess_fn)
# n_step=3 will collect a full episode
r = c0.collect(n_step=3)['rews']
assert len(r) == 0
r = c0.collect(n_episode=2)['rews']
assert r.shape == (2, 4) and np.all(r == 1)
env_fns = [
lambda x=i: MyTestEnv(size=x, sleep=0, ma_rew=4) for i in [2, 3, 4, 5]
]
envs = DummyVectorEnv(env_fns)
c1 = Collector(policy, envs,
VectorReplayBuffer(total_size=100, buffer_num=4),
Logger.single_preprocess_fn)
r = c1.collect(n_step=12)['rews']
assert r.shape == (2, 4) and np.all(r == 1), r
r = c1.collect(n_episode=8)['rews']
assert r.shape == (8, 4) and np.all(r == 1)
batch, _ = c1.buffer.sample(10)
print(batch)
c0.buffer.update(c1.buffer)
assert len(c0.buffer) in [42, 43]
if len(c0.buffer) == 42:
rew = [
0,
0,
0,
0,
1,
0,
0,
0,
0,
1,
0,
1,
0,
1,
0,
1,
0,
1,
0,
0,
1,
0,
0,
1,
0,
0,
0,
1,
0,
0,
0,
1,
0,
0,
0,
0,
1,
0,
0,
0,
0,
1,
]
else:
rew = [
0,
0,
0,
0,
1,
0,
0,
0,
0,
1,
0,
1,
0,
1,
0,
1,
0,
0,
1,
0,
0,
1,
0,
0,
1,
0,
0,
0,
1,
0,
0,
0,
1,
0,
0,
0,
0,
1,
0,
0,
0,
0,
1,
]
assert np.all(c0.buffer[:].rew == [[x] * 4 for x in rew])
assert np.all(c0.buffer[:].done == rew)
c2 = Collector(
policy, envs,
VectorReplayBuffer(total_size=100, buffer_num=4, stack_num=4),
Logger.single_preprocess_fn)
r = c2.collect(n_episode=10)['rews']
assert r.shape == (10, 4) and np.all(r == 1)
batch, _ = c2.buffer.sample(10)
def test_fqf(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
# train_envs = gym.make(args.task)
# you can also use tianshou.env.SubprocVectorEnv
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
feature_net = Net(
args.state_shape,
args.hidden_sizes[-1],
hidden_sizes=args.hidden_sizes[:-1],
device=args.device,
softmax=False
)
net = FullQuantileFunction(
feature_net,
args.action_shape,
args.hidden_sizes,
num_cosines=args.num_cosines,
device=args.device
)
optim = torch.optim.Adam(net.parameters(), lr=args.lr)
fraction_net = FractionProposalNetwork(args.num_fractions, net.input_dim)
fraction_optim = torch.optim.RMSprop(
fraction_net.parameters(), lr=args.fraction_lr
)
policy = FQFPolicy(
net,
optim,
fraction_net,
fraction_optim,
args.gamma,
args.num_fractions,
args.ent_coef,
args.n_step,
target_update_freq=args.target_update_freq
).to(args.device)
# buffer
if args.prioritized_replay:
buf = PrioritizedVectorReplayBuffer(
args.buffer_size,
buffer_num=len(train_envs),
alpha=args.alpha,
beta=args.beta
)
else:
buf = VectorReplayBuffer(args.buffer_size, buffer_num=len(train_envs))
# collector
train_collector = Collector(policy, train_envs, buf, exploration_noise=True)
test_collector = Collector(policy, test_envs, exploration_noise=True)
# policy.set_eps(1)
train_collector.collect(n_step=args.batch_size * args.training_num)
# log
log_path = os.path.join(args.logdir, args.task, 'fqf')
writer = SummaryWriter(log_path)
logger = TensorboardLogger(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
def train_fn(epoch, env_step):
# eps annnealing, just a demo
if env_step <= 10000:
policy.set_eps(args.eps_train)
elif env_step <= 50000:
eps = args.eps_train - (env_step - 10000) / \
40000 * (0.9 * args.eps_train)
policy.set_eps(eps)
else:
policy.set_eps(0.1 * args.eps_train)
def test_fn(epoch, env_step):
policy.set_eps(args.eps_test)
# trainer
result = offpolicy_trainer(
policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.step_per_collect,
args.test_num,
args.batch_size,
train_fn=train_fn,
test_fn=test_fn,
stop_fn=stop_fn,
save_fn=save_fn,
logger=logger,
update_per_step=args.update_per_step
)
assert stop_fn(result['best_reward'])
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
env = gym.make(args.task)
policy.eval()
policy.set_eps(args.eps_test)
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_td3(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]
# 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 = SubprocVectorEnv(
[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 = Actor(
args.layer_num, args.state_shape, args.action_shape,
args.max_action, args.device
).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
critic1 = Critic(
args.layer_num, args.state_shape, args.action_shape, args.device
).to(args.device)
critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
critic2 = Critic(
args.layer_num, args.state_shape, args.action_shape, args.device
).to(args.device)
critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
policy = TD3Policy(
actor, actor_optim, critic1, critic1_optim, critic2, critic2_optim,
args.tau, args.gamma, args.exploration_noise, args.policy_noise,
args.update_actor_freq, args.noise_clip,
[env.action_space.low[0], env.action_space.high[0]],
reward_normalization=True, ignore_done=True)
# collector
train_collector = Collector(
policy, train_envs, ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# train_collector.collect(n_step=args.buffer_size)
# log
writer = SummaryWriter(args.logdir + '/' + 'td3')
def stop_fn(x):
return x >= env.spec.reward_threshold
# trainer
result = offpolicy_trainer(
policy, train_collector, test_collector, args.epoch,
args.step_per_epoch, args.collect_per_step, args.test_num,
args.batch_size, stop_fn=stop_fn, writer=writer, task=args.task)
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 save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
>>>>>>> 4fd826761c9884457928da9dac52d7ee1c51443a
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_fn()
test_pg()
def test_ddpg(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]
args.exploration_noise = args.exploration_noise * args.max_action
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)
if args.training_num > 1:
train_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
else:
train_envs = gym.make(args.task)
# test_envs = gym.make(args.task)
test_envs = SubprocVectorEnv(
[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_a = Net(args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device)
actor = Actor(net_a,
args.action_shape,
max_action=args.max_action,
device=args.device).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
net_c = Net(args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
concat=True,
device=args.device)
critic = Critic(net_c, device=args.device).to(args.device)
critic_optim = torch.optim.Adam(critic.parameters(), lr=args.critic_lr)
policy = DDPGPolicy(
actor,
actor_optim,
critic,
critic_optim,
tau=args.tau,
gamma=args.gamma,
exploration_noise=GaussianNoise(sigma=args.exploration_noise),
estimation_step=args.n_step,
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)
train_collector.collect(n_step=args.start_timesteps, random=True)
# log
t0 = datetime.datetime.now().strftime("%m%d_%H%M%S")
log_file = f'seed_{args.seed}_{t0}-{args.task.replace("-", "_")}_ddpg'
log_path = os.path.join(args.logdir, args.task, 'ddpg', log_file)
writer = SummaryWriter(log_path)
writer.add_text("args", str(args))
logger = TensorboardLogger(writer)
def save_best_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
if not args.watch:
# trainer
result = offpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.step_per_collect,
args.test_num,
args.batch_size,
save_best_fn=save_best_fn,
logger=logger,
update_per_step=args.update_per_step,
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_c51(args=get_args()):
env = make_atari_env(args)
args.state_shape = env.observation_space.shape or env.observation_space.n
args.action_shape = env.env.action_space.shape or env.env.action_space.n
# should be N_FRAMES x H x W
print("Observations shape:", args.state_shape)
print("Actions shape:", args.action_shape)
# make environments
train_envs = SubprocVectorEnv(
[lambda: make_atari_env(args) for _ in range(args.training_num)])
test_envs = SubprocVectorEnv(
[lambda: make_atari_env_watch(args) 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)
# define model
net = C51(*args.state_shape, args.action_shape, args.num_atoms,
args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.lr)
# define policy
policy = C51Policy(net,
optim,
args.gamma,
args.num_atoms,
args.v_min,
args.v_max,
args.n_step,
target_update_freq=args.target_update_freq).to(
args.device)
# 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)
# replay buffer: `save_last_obs` and `stack_num` can be removed together
# when you have enough RAM
buffer = ReplayBuffer(args.buffer_size,
ignore_obs_next=True,
save_only_last_obs=True,
stack_num=args.frames_stack)
# collector
train_collector = Collector(policy, train_envs, buffer)
test_collector = Collector(policy, test_envs)
# log
log_path = os.path.join(args.logdir, args.task, 'c51')
writer = SummaryWriter(log_path)
def save_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(mean_rewards):
if env.env.spec.reward_threshold:
return mean_rewards >= env.spec.reward_threshold
elif 'Pong' in args.task:
return mean_rewards >= 20
else:
return False
def train_fn(epoch, env_step):
# nature DQN setting, linear decay in the first 1M steps
if env_step <= 1e6:
eps = args.eps_train - env_step / 1e6 * \
(args.eps_train - args.eps_train_final)
else:
eps = args.eps_train_final
policy.set_eps(eps)
writer.add_scalar('train/eps', eps, global_step=env_step)
def test_fn(epoch, env_step):
policy.set_eps(args.eps_test)
# watch agent's performance
def watch():
print("Testing agent ...")
policy.eval()
policy.set_eps(args.eps_test)
test_envs.seed(args.seed)
test_collector.reset()
result = test_collector.collect(n_episode=[1] * args.test_num,
render=args.render)
pprint.pprint(result)
if args.watch:
watch()
exit(0)
# test train_collector and start filling replay buffer
train_collector.collect(n_step=args.batch_size * 4)
# trainer
result = offpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.collect_per_step,
args.test_num,
args.batch_size,
train_fn=train_fn,
test_fn=test_fn,
stop_fn=stop_fn,
save_fn=save_fn,
writer=writer,
test_in_train=False)
pprint.pprint(result)
watch()
def test_ppo(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
if args.reward_threshold is None:
default_reward_threshold = {"CartPole-v0": 195}
args.reward_threshold = default_reward_threshold.get(
args.task, env.spec.reward_threshold)
# train_envs = gym.make(args.task)
# you can also use tianshou.env.SubprocVectorEnv
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 = Actor(net, args.action_shape, device=args.device).to(args.device)
critic = Critic(net, device=args.device).to(args.device)
actor_critic = ActorCritic(actor, critic)
# orthogonal initialization
for m in actor_critic.modules():
if isinstance(m, torch.nn.Linear):
torch.nn.init.orthogonal_(m.weight)
torch.nn.init.zeros_(m.bias)
optim = torch.optim.Adam(actor_critic.parameters(), lr=args.lr)
dist = torch.distributions.Categorical
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,
gae_lambda=args.gae_lambda,
reward_normalization=args.rew_norm,
dual_clip=args.dual_clip,
value_clip=args.value_clip,
action_space=env.action_space,
deterministic_eval=True,
advantage_normalization=args.norm_adv,
recompute_advantage=args.recompute_adv)
feature_dim = args.hidden_sizes[-1]
feature_net = MLP(np.prod(args.state_shape),
output_dim=feature_dim,
hidden_sizes=args.hidden_sizes[:-1],
device=args.device)
action_dim = np.prod(args.action_shape)
icm_net = IntrinsicCuriosityModule(feature_net,
feature_dim,
action_dim,
hidden_sizes=args.hidden_sizes[-1:],
device=args.device).to(args.device)
icm_optim = torch.optim.Adam(icm_net.parameters(), lr=args.lr)
policy = ICMPolicy(policy, icm_net, icm_optim, args.lr_scale,
args.reward_scale, args.forward_loss_weight)
# 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_icm')
writer = SummaryWriter(log_path)
logger = TensorboardLogger(writer)
def save_best_fn(policy):
torch.save(policy.state_dict(), os.path.join(log_path, 'policy.pth'))
def stop_fn(mean_rewards):
return mean_rewards >= args.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_best_fn=save_best_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_collector_with_atari_setting():
reference_obs = np.zeros([6, 4, 84, 84])
for i in range(6):
reference_obs[i, 3, np.arange(84), np.arange(84)] = i
reference_obs[i, 2, np.arange(84)] = i
reference_obs[i, 1, :, np.arange(84)] = i
reference_obs[i, 0] = i
# atari single buffer
env = MyTestEnv(size=5, sleep=0, array_state=True)
policy = MyPolicy()
c0 = Collector(policy, env, ReplayBuffer(size=100))
c0.collect(n_step=6)
c0.collect(n_episode=2)
assert c0.buffer.obs.shape == (100, 4, 84, 84)
assert c0.buffer.obs_next.shape == (100, 4, 84, 84)
assert len(c0.buffer) == 15
obs = np.zeros_like(c0.buffer.obs)
obs[np.arange(15)] = reference_obs[np.arange(15) % 5]
assert np.all(obs == c0.buffer.obs)
c1 = Collector(policy, env, ReplayBuffer(size=100, ignore_obs_next=True))
c1.collect(n_episode=3)
assert np.allclose(c0.buffer.obs, c1.buffer.obs)
with pytest.raises(AttributeError):
c1.buffer.obs_next
assert np.all(reference_obs[[1, 2, 3, 4, 4] * 3] == c1.buffer[:].obs_next)
c2 = Collector(
policy, env,
ReplayBuffer(size=100, ignore_obs_next=True, save_only_last_obs=True))
c2.collect(n_step=8)
assert c2.buffer.obs.shape == (100, 84, 84)
obs = np.zeros_like(c2.buffer.obs)
obs[np.arange(8)] = reference_obs[[0, 1, 2, 3, 4, 0, 1, 2], -1]
assert np.all(c2.buffer.obs == obs)
assert np.allclose(c2.buffer[:].obs_next,
reference_obs[[1, 2, 3, 4, 4, 1, 2, 2], -1])
# atari multi buffer
env_fns = [
lambda x=i: MyTestEnv(size=x, sleep=0, array_state=True)
for i in [2, 3, 4, 5]
]
envs = DummyVectorEnv(env_fns)
c3 = Collector(policy, envs,
VectorReplayBuffer(total_size=100, buffer_num=4))
c3.collect(n_step=12)
result = c3.collect(n_episode=9)
assert result["n/ep"] == 9 and result["n/st"] == 23
assert c3.buffer.obs.shape == (100, 4, 84, 84)
obs = np.zeros_like(c3.buffer.obs)
obs[np.arange(8)] = reference_obs[[0, 1, 0, 1, 0, 1, 0, 1]]
obs[np.arange(25, 34)] = reference_obs[[0, 1, 2, 0, 1, 2, 0, 1, 2]]
obs[np.arange(50, 58)] = reference_obs[[0, 1, 2, 3, 0, 1, 2, 3]]
obs[np.arange(75, 85)] = reference_obs[[0, 1, 2, 3, 4, 0, 1, 2, 3, 4]]
assert np.all(obs == c3.buffer.obs)
obs_next = np.zeros_like(c3.buffer.obs_next)
obs_next[np.arange(8)] = reference_obs[[1, 2, 1, 2, 1, 2, 1, 2]]
obs_next[np.arange(25, 34)] = reference_obs[[1, 2, 3, 1, 2, 3, 1, 2, 3]]
obs_next[np.arange(50, 58)] = reference_obs[[1, 2, 3, 4, 1, 2, 3, 4]]
obs_next[np.arange(75, 85)] = reference_obs[[1, 2, 3, 4, 5, 1, 2, 3, 4, 5]]
assert np.all(obs_next == c3.buffer.obs_next)
c4 = Collector(
policy, envs,
VectorReplayBuffer(total_size=100,
buffer_num=4,
stack_num=4,
ignore_obs_next=True,
save_only_last_obs=True))
c4.collect(n_step=12)
result = c4.collect(n_episode=9)
assert result["n/ep"] == 9 and result["n/st"] == 23
assert c4.buffer.obs.shape == (100, 84, 84)
obs = np.zeros_like(c4.buffer.obs)
slice_obs = reference_obs[:, -1]
obs[np.arange(8)] = slice_obs[[0, 1, 0, 1, 0, 1, 0, 1]]
obs[np.arange(25, 34)] = slice_obs[[0, 1, 2, 0, 1, 2, 0, 1, 2]]
obs[np.arange(50, 58)] = slice_obs[[0, 1, 2, 3, 0, 1, 2, 3]]
obs[np.arange(75, 85)] = slice_obs[[0, 1, 2, 3, 4, 0, 1, 2, 3, 4]]
assert np.all(c4.buffer.obs == obs)
obs_next = np.zeros([len(c4.buffer), 4, 84, 84])
ref_index = np.array([
1,
1,
1,
1,
1,
1,
1,
1,
1,
2,
2,
1,
2,
2,
1,
2,
2,
1,
2,
3,
3,
1,
2,
3,
3,
1,
2,
3,
4,
4,
1,
2,
3,
4,
4,
])
obs_next[:, -1] = slice_obs[ref_index]
ref_index -= 1
ref_index[ref_index < 0] = 0
obs_next[:, -2] = slice_obs[ref_index]
ref_index -= 1
ref_index[ref_index < 0] = 0
obs_next[:, -3] = slice_obs[ref_index]
ref_index -= 1
ref_index[ref_index < 0] = 0
obs_next[:, -4] = slice_obs[ref_index]
assert np.all(obs_next == c4.buffer[:].obs_next)
buf = ReplayBuffer(100,
stack_num=4,
ignore_obs_next=True,
save_only_last_obs=True)
c5 = Collector(policy, envs, CachedReplayBuffer(buf, 4, 10))
result_ = c5.collect(n_step=12)
assert len(buf) == 5 and len(c5.buffer) == 12
result = c5.collect(n_episode=9)
assert result["n/ep"] == 9 and result["n/st"] == 23
assert len(buf) == 35
assert np.all(buf.obs[:len(buf)] == slice_obs[[
0, 1, 0, 1, 2, 0, 1, 0, 1, 2, 3, 0, 1, 2, 3, 4, 0, 1, 0, 1, 2, 0, 1, 0,
1, 2, 3, 0, 1, 2, 0, 1, 2, 3, 4
]])
assert np.all(buf[:].obs_next[:, -1] == slice_obs[[
1, 1, 1, 2, 2, 1, 1, 1, 2, 3, 3, 1, 2, 3, 4, 4, 1, 1, 1, 2, 2, 1, 1, 1,
2, 3, 3, 1, 2, 2, 1, 2, 3, 4, 4
]])
assert len(buf) == len(c5.buffer)
# test buffer=None
c6 = Collector(policy, envs)
result1 = c6.collect(n_step=12)
for key in ["n/ep", "n/st", "rews", "lens"]:
assert np.allclose(result1[key], result_[key])
result2 = c6.collect(n_episode=9)
for key in ["n/ep", "n/st", "rews", "lens"]:
assert np.allclose(result2[key], result[key])
def test_dqn(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
# train_envs = gym.make(args.task)
# you can also use tianshou.env.SubprocVectorEnv
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)
# Q_param = V_param = {"hidden_sizes": [128]}
# model
net = Net(
args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
device=args.device,
# dueling=(Q_param, V_param),
).to(args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.lr)
policy = DQNPolicy(net,
optim,
args.gamma,
args.n_step,
target_update_freq=args.target_update_freq)
# buffer
if args.prioritized_replay:
buf = PrioritizedReplayBuffer(args.buffer_size,
alpha=args.alpha,
beta=args.beta)
else:
buf = ReplayBuffer(args.buffer_size)
# collector
train_collector = Collector(policy, train_envs, buf)
test_collector = Collector(policy, test_envs)
# policy.set_eps(1)
train_collector.collect(n_step=args.batch_size)
# log
log_path = os.path.join(args.logdir, args.task, 'dqn')
writer = SummaryWriter(log_path)
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 train_fn(epoch, env_step):
# eps annnealing, just a demo
if env_step <= 10000:
policy.set_eps(args.eps_train)
elif env_step <= 50000:
eps = args.eps_train - (env_step - 10000) / \
40000 * (0.9 * args.eps_train)
policy.set_eps(eps)
else:
policy.set_eps(0.1 * args.eps_train)
def test_fn(epoch, env_step):
policy.set_eps(args.eps_test)
# trainer
result = offpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.collect_per_step,
args.test_num,
args.batch_size,
train_fn=train_fn,
test_fn=test_fn,
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)
policy.eval()
policy.set_eps(args.eps_test)
collector = Collector(policy, env)
result = collector.collect(n_episode=1, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
def test_collector():
writer = SummaryWriter('log/collector')
logger = Logger(writer)
env_fns = [lambda x=i: MyTestEnv(size=x, sleep=0) for i in [2, 3, 4, 5]]
venv = SubprocVectorEnv(env_fns)
dum = DummyVectorEnv(env_fns)
policy = MyPolicy()
env = env_fns[0]()
c0 = Collector(policy, env, ReplayBuffer(size=100), logger.preprocess_fn)
c0.collect(n_step=3)
assert len(c0.buffer) == 3
assert np.allclose(c0.buffer.obs[:4, 0], [0, 1, 0, 0])
assert np.allclose(c0.buffer[:].obs_next[..., 0], [1, 2, 1])
c0.collect(n_episode=3)
assert len(c0.buffer) == 8
assert np.allclose(c0.buffer.obs[:10, 0], [0, 1, 0, 1, 0, 1, 0, 1, 0, 0])
assert np.allclose(c0.buffer[:].obs_next[..., 0], [1, 2, 1, 2, 1, 2, 1, 2])
c0.collect(n_step=3, random=True)
c1 = Collector(policy, venv,
VectorReplayBuffer(total_size=100, buffer_num=4),
logger.preprocess_fn)
c1.collect(n_step=8)
obs = np.zeros(100)
obs[[0, 1, 25, 26, 50, 51, 75, 76]] = [0, 1, 0, 1, 0, 1, 0, 1]
assert np.allclose(c1.buffer.obs[:, 0], obs)
assert np.allclose(c1.buffer[:].obs_next[..., 0], [1, 2, 1, 2, 1, 2, 1, 2])
c1.collect(n_episode=4)
assert len(c1.buffer) == 16
obs[[2, 3, 27, 52, 53, 77, 78, 79]] = [0, 1, 2, 2, 3, 2, 3, 4]
assert np.allclose(c1.buffer.obs[:, 0], obs)
assert np.allclose(c1.buffer[:].obs_next[..., 0],
[1, 2, 1, 2, 1, 2, 3, 1, 2, 3, 4, 1, 2, 3, 4, 5])
c1.collect(n_episode=4, random=True)
c2 = Collector(policy, dum, VectorReplayBuffer(total_size=100,
buffer_num=4),
logger.preprocess_fn)
c2.collect(n_episode=7)
obs1 = obs.copy()
obs1[[4, 5, 28, 29, 30]] = [0, 1, 0, 1, 2]
obs2 = obs.copy()
obs2[[28, 29, 30, 54, 55, 56, 57]] = [0, 1, 2, 0, 1, 2, 3]
c2obs = c2.buffer.obs[:, 0]
assert np.all(c2obs == obs1) or np.all(c2obs == obs2)
c2.reset_env()
c2.reset_buffer()
assert c2.collect(n_episode=8)['n/ep'] == 8
obs[[4, 5, 28, 29, 30, 54, 55, 56, 57]] = [0, 1, 0, 1, 2, 0, 1, 2, 3]
assert np.all(c2.buffer.obs[:, 0] == obs)
c2.collect(n_episode=4, random=True)
# test corner case
with pytest.raises(TypeError):
Collector(policy, dum, ReplayBuffer(10))
with pytest.raises(TypeError):
Collector(policy, dum, PrioritizedReplayBuffer(10, 0.5, 0.5))
with pytest.raises(TypeError):
c2.collect()
# test NXEnv
for obs_type in ["array", "object"]:
envs = SubprocVectorEnv(
[lambda i=x: NXEnv(i, obs_type) for x in [5, 10, 15, 20]])
c3 = Collector(policy, envs,
VectorReplayBuffer(total_size=100, buffer_num=4))
c3.collect(n_step=6)
assert c3.buffer.obs.dtype == object
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 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)))
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 train_agent(
args: argparse.Namespace = get_args(),
agent_learn: Optional[BasePolicy] = None,
agent_opponent: Optional[BasePolicy] = None,
optim: Optional[torch.optim.Optimizer] = None,
) -> Tuple[dict, BasePolicy]:
def env_func():
return TicTacToeEnv(args.board_size, args.win_size)
train_envs = DummyVectorEnv([env_func for _ in range(args.training_num)])
test_envs = DummyVectorEnv([env_func 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)
policy, optim = get_agents(args,
agent_learn=agent_learn,
agent_opponent=agent_opponent,
optim=optim)
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# policy.set_eps(1)
train_collector.collect(n_step=args.batch_size)
# log
if not hasattr(args, 'writer'):
log_path = os.path.join(args.logdir, 'tic_tac_toe', 'dqn')
writer = SummaryWriter(log_path)
args.writer = writer
else:
writer = args.writer
def save_fn(policy):
if hasattr(args, 'model_save_path'):
model_save_path = args.model_save_path
else:
model_save_path = os.path.join(args.logdir, 'tic_tac_toe', 'dqn',
'policy.pth')
torch.save(policy.policies[args.agent_id - 1].state_dict(),
model_save_path)
def stop_fn(x):
return x >= args.win_rate
def train_fn(x):
policy.policies[args.agent_id - 1].set_eps(args.eps_train)
def test_fn(x):
policy.policies[args.agent_id - 1].set_eps(args.eps_test)
# trainer
result = offpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.collect_per_step,
args.test_num,
args.batch_size,
train_fn=train_fn,
test_fn=test_fn,
stop_fn=stop_fn,
save_fn=save_fn,
writer=writer,
test_in_train=False)
return result, policy.policies[args.agent_id - 1]
def test_ppo(args=get_args()):
env = gym.make(args.task)
train_envs = gym.make(args.task)
args.state_shape = get_observation_size(train_envs)
args.action_shape = get_action_size(train_envs)
test_envs = gym.make(args.task)
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
train_envs.seed(args.seed)
test_envs.seed(args.seed)
# model
net = Net_Dis(args.layer_num, args.state_shape, device=args.device)
actor = Actor_Dis(net, args.action_shape).to(args.device)
critic = Critic_Dis(net).to(args.device)
optim = torch.optim.Adam(list(actor.parameters()) +
list(critic.parameters()),
lr=args.lr)
dist = torch.distributions.Categorical
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=None)
# collector
train_collector = Collector(policy, train_envs,
ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# log
writer = SummaryWriter(args.logdir + '/' + 'ppo')
def stop_fn(x):
if env.env.spec.reward_threshold:
return x >= env.spec.reward_threshold
else:
return False
# 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,
writer=writer,
task=args.task)
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_step=2000, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
collector.close()
def offpolicy_trainer(
policy: BasePolicy,
train_collector: Collector,
max_epoch: int,
step_per_epoch: int,
collect_per_step: int,
batch_size: int,
update_per_step: int = 1,
train_fn: Optional[Callable[[int], None]] = None,
writer: Optional[SummaryWriter] = None,
log_interval: int = 1,
) -> int:
"""A wrapper for off-policy trainer procedure. The ``step`` in trainer
means a policy network update.
:param policy: an instance of the :class:`~tianshou.policy.BasePolicy`
class.
:param train_collector: the collector used for training.
:type train_collector: :class:`~tianshou.data.Collector`
:param test_collector: the collector used for testing.
:type test_collector: :class:`~tianshou.data.Collector`
:param int max_epoch: the maximum of epochs for training. The training
process might be finished before reaching the ``max_epoch``.
:param int step_per_epoch: the number of step for updating policy network
in one epoch.
:param int collect_per_step: the number of frames the collector would
collect before the network update. In other words, collect some frames
and do some policy network update.
:param episode_per_test: the number of episodes for one policy evaluation.
:param int batch_size: the batch size of sample data, which is going to
feed in the policy network.
:param int update_per_step: the number of times the policy network would
be updated after frames are collected, for example, set it to 256 means
it updates policy 256 times once after ``collect_per_step`` frames are
collected.
:param function train_fn: a function receives the current number of epoch
index and performs some operations at the beginning of training in this
epoch.
:param torch.utils.tensorboard.SummaryWriter writer: a TensorBoard
SummaryWriter.
:param int log_interval: the log interval of the writer.
:return: See :func:`~tianshou.trainer.gather_info`.
"""
global_step = 0
best_epoch, best_reward = -1, -1.
stat = {}
start_time = time.time()
for epoch in range(1, 1 + max_epoch):
# train
policy.train()
if train_fn:
train_fn(epoch)
with tqdm.tqdm(total=step_per_epoch,
desc=f'Epoch #{epoch}',
**tqdm_config) as t:
while t.n < t.total:
assert train_collector.policy == policy
result = train_collector.collect(n_step=collect_per_step)
data = {}
for i in range(update_per_step * min(
result['n/st'] // collect_per_step, t.total - t.n)):
global_step += collect_per_step
losses = policy.update(batch_size, train_collector.buffer)
for k in result.keys():
data[k] = f'{result[k]:.2f}'
if writer and global_step % log_interval == 0:
writer.add_scalar('train/' + k,
result[k],
global_step=global_step)
for k in losses.keys():
if stat.get(k) is None:
stat[k] = MovAvg()
stat[k].add(losses[k])
data[k] = f'{stat[k].get():.6f}'
if writer and global_step % log_interval == 0:
writer.add_scalar(k,
stat[k].get(),
global_step=global_step)
data['eps'] = policy.eps
t.update(1)
t.set_postfix(**data)
if t.n <= t.total:
t.update()
return global_step
def onpolicy_trainer(
policy: BasePolicy,
train_collector: Collector,
test_collector: Collector,
max_epoch: int,
step_per_epoch: int,
collect_per_step: int,
repeat_per_collect: int,
episode_per_test: Union[int, List[int]],
batch_size: int,
train_fn: Optional[Callable[[int], None]] = None,
test_fn: Optional[Callable[[int], None]] = None,
stop_fn: Optional[Callable[[float], bool]] = None,
save_fn: Optional[Callable[[BasePolicy], None]] = None,
log_fn: Optional[Callable[[dict], None]] = None,
writer: Optional[SummaryWriter] = None,
log_interval: int = 1,
verbose: bool = True,
test_in_train: bool = True,
) -> Dict[str, Union[float, str]]:
"""A wrapper for on-policy trainer procedure.
:param policy: an instance of the :class:`~tianshou.policy.BasePolicy`
class.
:param train_collector: the collector used for training.
:type train_collector: :class:`~tianshou.data.Collector`
:param test_collector: the collector used for testing.
:type test_collector: :class:`~tianshou.data.Collector`
:param int max_epoch: the maximum of epochs for training. The training
process might be finished before reaching the ``max_epoch``.
:param int step_per_epoch: the number of step for updating policy network
in one epoch.
:param int collect_per_step: the number of frames the collector would
collect before the network update. In other words, collect some frames
and do one policy network update.
:param int repeat_per_collect: the number of repeat time for policy
learning, for example, set it to 2 means the policy needs to learn each
given batch data twice.
:param episode_per_test: the number of episodes for one policy evaluation.
:type episode_per_test: int or list of ints
:param int batch_size: the batch size of sample data, which is going to
feed in the policy network.
:param function train_fn: a function receives the current number of epoch
index and performs some operations at the beginning of training in this
epoch.
:param function test_fn: a function receives the current number of epoch
index and performs some operations at the beginning of testing in this
epoch.
:param function save_fn: a function for saving policy when the undiscounted
average mean reward in evaluation phase gets better.
:param function stop_fn: a function receives the average undiscounted
returns of the testing result, return a boolean which indicates whether
reaching the goal.
:param function log_fn: a function receives env info for logging.
:param torch.utils.tensorboard.SummaryWriter writer: a TensorBoard
SummaryWriter.
:param int log_interval: the log interval of the writer.
:param bool verbose: whether to print the information.
:param bool test_in_train: whether to test in the training phase.
:return: See :func:`~tianshou.trainer.gather_info`.
"""
global_step = 0
best_epoch, best_reward = -1, -1
stat = {}
start_time = time.time()
test_in_train = test_in_train and train_collector.policy == policy
for epoch in range(1, 1 + max_epoch):
# train
policy.train()
if train_fn:
train_fn(epoch)
with tqdm.tqdm(total=step_per_epoch,
desc=f'Epoch #{epoch}',
**tqdm_config) as t:
while t.n < t.total:
result = train_collector.collect(n_episode=collect_per_step,
log_fn=log_fn)
data = {}
if test_in_train and stop_fn and stop_fn(result['rew']):
test_result = test_episode(policy, test_collector, test_fn,
epoch, episode_per_test)
if stop_fn and stop_fn(test_result['rew']):
if save_fn:
save_fn(policy)
for k in result.keys():
data[k] = f'{result[k]:.2f}'
t.set_postfix(**data)
return gather_info(start_time, train_collector,
test_collector, test_result['rew'])
else:
policy.train()
if train_fn:
train_fn(epoch)
losses = policy.learn(train_collector.sample(0), batch_size,
repeat_per_collect)
train_collector.reset_buffer()
step = 1
for k in losses.keys():
if isinstance(losses[k], list):
step = max(step, len(losses[k]))
global_step += step
for k in result.keys():
data[k] = f'{result[k]:.2f}'
if writer and global_step % log_interval == 0:
writer.add_scalar(k,
result[k],
global_step=global_step)
for k in losses.keys():
if stat.get(k) is None:
stat[k] = MovAvg()
stat[k].add(losses[k])
data[k] = f'{stat[k].get():.6f}'
if writer and global_step % log_interval == 0:
writer.add_scalar(k,
stat[k].get(),
global_step=global_step)
t.update(step)
t.set_postfix(**data)
if t.n <= t.total:
t.update()
# test
result = test_episode(policy, test_collector, test_fn, epoch,
episode_per_test)
if best_epoch == -1 or best_reward < result['rew']:
best_reward = result['rew']
best_epoch = epoch
if save_fn:
save_fn(policy)
if verbose:
print(f'Epoch #{epoch}: test_reward: {result["rew"]:.6f}, '
f'best_reward: {best_reward:.6f} in #{best_epoch}')
if stop_fn and stop_fn(best_reward):
break
return gather_info(start_time, train_collector, test_collector,
best_reward)
def test_ddpg(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
actor = Actor(args.layer_num, args.state_shape, args.action_shape,
args.max_action, args.device).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
critic = Critic(args.layer_num, args.state_shape, args.action_shape,
args.device).to(args.device)
critic_optim = torch.optim.Adam(critic.parameters(), lr=args.critic_lr)
policy = DDPGPolicy(actor,
actor_optim,
critic,
critic_optim,
args.tau,
args.gamma,
args.exploration_noise,
[env.action_space.low[0], env.action_space.high[0]],
reward_normalization=True,
ignore_done=True)
# 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, 'ddpg')
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 = offpolicy_trainer(policy,
train_collector,
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'])
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 train_agent(
args: argparse.Namespace = get_args(),
agent_learn: Optional[BasePolicy] = None,
agent_opponent: Optional[BasePolicy] = None,
optim: Optional[torch.optim.Optimizer] = None,
) -> Tuple[dict, BasePolicy]:
def env_func():
return TicTacToeEnv(args.board_size, args.win_size)
train_envs = DummyVectorEnv([env_func for _ in range(args.training_num)])
test_envs = DummyVectorEnv([env_func 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)
policy, optim = get_agents(args,
agent_learn=agent_learn,
agent_opponent=agent_opponent,
optim=optim)
# collector
train_collector = Collector(policy,
train_envs,
VectorReplayBuffer(args.buffer_size,
len(train_envs)),
exploration_noise=True)
test_collector = Collector(policy, test_envs, exploration_noise=True)
# policy.set_eps(1)
train_collector.collect(n_step=args.batch_size * args.training_num)
# log
log_path = os.path.join(args.logdir, 'tic_tac_toe', 'dqn')
writer = SummaryWriter(log_path)
writer.add_text("args", str(args))
logger = TensorboardLogger(writer)
def save_fn(policy):
if hasattr(args, 'model_save_path'):
model_save_path = args.model_save_path
else:
model_save_path = os.path.join(args.logdir, 'tic_tac_toe', 'dqn',
'policy.pth')
torch.save(policy.policies[args.agent_id - 1].state_dict(),
model_save_path)
def stop_fn(mean_rewards):
return mean_rewards >= args.win_rate
def train_fn(epoch, env_step):
policy.policies[args.agent_id - 1].set_eps(args.eps_train)
def test_fn(epoch, env_step):
policy.policies[args.agent_id - 1].set_eps(args.eps_test)
def reward_metric(rews):
return rews[:, args.agent_id - 1]
# trainer
result = offpolicy_trainer(policy,
train_collector,
test_collector,
args.epoch,
args.step_per_epoch,
args.step_per_collect,
args.test_num,
args.batch_size,
train_fn=train_fn,
test_fn=test_fn,
stop_fn=stop_fn,
save_fn=save_fn,
update_per_step=args.update_per_step,
logger=logger,
test_in_train=False,
reward_metric=reward_metric)
return result, policy.policies[args.agent_id - 1]
def test_ppo(args=get_args()):
env, train_envs, test_envs = make_atari_env(
args.task,
args.seed,
args.training_num,
args.test_num,
scale=args.scale_obs,
frame_stack=args.frames_stack,
)
args.state_shape = env.observation_space.shape or env.observation_space.n
args.action_shape = env.action_space.shape or env.action_space.n
# should be N_FRAMES x H x W
print("Observations shape:", args.state_shape)
print("Actions shape:", args.action_shape)
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
# define model
net = DQN(*args.state_shape,
args.action_shape,
device=args.device,
features_only=True,
output_dim=args.hidden_size)
actor = Actor(net,
args.action_shape,
device=args.device,
softmax_output=False)
critic = Critic(net, device=args.device)
optim = torch.optim.Adam(ActorCritic(actor, 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)
# define policy
def dist(p):
return torch.distributions.Categorical(logits=p)
policy = PPOPolicy(
actor,
critic,
optim,
dist,
discount_factor=args.gamma,
gae_lambda=args.gae_lambda,
max_grad_norm=args.max_grad_norm,
vf_coef=args.vf_coef,
ent_coef=args.ent_coef,
reward_normalization=args.rew_norm,
action_scaling=False,
lr_scheduler=lr_scheduler,
action_space=env.action_space,
eps_clip=args.eps_clip,
value_clip=args.value_clip,
dual_clip=args.dual_clip,
advantage_normalization=args.norm_adv,
recompute_advantage=args.recompute_adv,
).to(args.device)
if args.icm_lr_scale > 0:
feature_net = DQN(*args.state_shape,
args.action_shape,
args.device,
features_only=True)
action_dim = np.prod(args.action_shape)
feature_dim = feature_net.output_dim
icm_net = IntrinsicCuriosityModule(
feature_net.net,
feature_dim,
action_dim,
hidden_sizes=args.hidden_sizes,
device=args.device,
)
icm_optim = torch.optim.Adam(icm_net.parameters(), lr=args.lr)
policy = ICMPolicy(policy, icm_net, icm_optim, args.icm_lr_scale,
args.icm_reward_scale,
args.icm_forward_loss_weight).to(args.device)
# 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)
# replay buffer: `save_last_obs` and `stack_num` can be removed together
# when you have enough RAM
buffer = VectorReplayBuffer(
args.buffer_size,
buffer_num=len(train_envs),
ignore_obs_next=True,
save_only_last_obs=True,
stack_num=args.frames_stack,
)
# collector
train_collector = Collector(policy,
train_envs,
buffer,
exploration_noise=True)
test_collector = Collector(policy, test_envs, exploration_noise=True)
# log
now = datetime.datetime.now().strftime("%y%m%d-%H%M%S")
args.algo_name = "ppo_icm" if args.icm_lr_scale > 0 else "ppo"
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):
torch.save(policy.state_dict(), os.path.join(log_path, "policy.pth"))
def stop_fn(mean_rewards):
if env.spec.reward_threshold:
return mean_rewards >= env.spec.reward_threshold
elif "Pong" in args.task:
return mean_rewards >= 20
else:
return False
def save_checkpoint_fn(epoch, env_step, gradient_step):
# see also: https://pytorch.org/tutorials/beginner/saving_loading_models.html
ckpt_path = os.path.join(log_path, "checkpoint.pth")
torch.save({"model": policy.state_dict()}, ckpt_path)
return ckpt_path
# watch agent's performance
def watch():
print("Setup test envs ...")
policy.eval()
test_envs.seed(args.seed)
if args.save_buffer_name:
print(f"Generate buffer with size {args.buffer_size}")
buffer = VectorReplayBuffer(
args.buffer_size,
buffer_num=len(test_envs),
ignore_obs_next=True,
save_only_last_obs=True,
stack_num=args.frames_stack,
)
collector = Collector(policy,
test_envs,
buffer,
exploration_noise=True)
result = collector.collect(n_step=args.buffer_size)
print(f"Save buffer into {args.save_buffer_name}")
# Unfortunately, pickle will cause oom with 1M buffer size
buffer.save_hdf5(args.save_buffer_name)
else:
print("Testing agent ...")
test_collector.reset()
result = test_collector.collect(n_episode=args.test_num,
render=args.render)
rew = result["rews"].mean()
print(f"Mean reward (over {result['n/ep']} episodes): {rew}")
if args.watch:
watch()
exit(0)
# test train_collector and start filling replay buffer
train_collector.collect(n_step=args.batch_size * args.training_num)
# 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_best_fn=save_best_fn,
logger=logger,
test_in_train=False,
resume_from_log=args.resume_id is not None,
save_checkpoint_fn=save_checkpoint_fn,
)
pprint.pprint(result)
watch()
def test_td3_bc():
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] # float
print("device:", args.device)
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))
args.state_dim = args.state_shape[0]
args.action_dim = args.action_shape[0]
print("Max_action", args.max_action)
test_envs = SubprocVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.test_num)])
if args.norm_obs:
test_envs = VectorEnvNormObs(test_envs, update_obs_rms=False)
# seed
np.random.seed(args.seed)
torch.manual_seed(args.seed)
test_envs.seed(args.seed)
# model
# actor network
net_a = Net(
args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device,
)
actor = Actor(
net_a,
action_shape=args.action_shape,
max_action=args.max_action,
device=args.device,
).to(args.device)
actor_optim = torch.optim.Adam(actor.parameters(), lr=args.actor_lr)
# critic network
net_c1 = Net(
args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
concat=True,
device=args.device,
)
net_c2 = Net(
args.state_shape,
args.action_shape,
hidden_sizes=args.hidden_sizes,
concat=True,
device=args.device,
)
critic1 = Critic(net_c1, device=args.device).to(args.device)
critic1_optim = torch.optim.Adam(critic1.parameters(), lr=args.critic_lr)
critic2 = Critic(net_c2, device=args.device).to(args.device)
critic2_optim = torch.optim.Adam(critic2.parameters(), lr=args.critic_lr)
policy = TD3BCPolicy(
actor,
actor_optim,
critic1,
critic1_optim,
critic2,
critic2_optim,
tau=args.tau,
gamma=args.gamma,
exploration_noise=GaussianNoise(sigma=args.exploration_noise),
policy_noise=args.policy_noise,
update_actor_freq=args.update_actor_freq,
noise_clip=args.noise_clip,
alpha=args.alpha,
estimation_step=args.n_step,
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
test_collector = Collector(policy, test_envs)
# log
now = datetime.datetime.now().strftime("%y%m%d-%H%M%S")
args.algo_name = "td3_bc"
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):
torch.save(policy.state_dict(), os.path.join(log_path, "policy.pth"))
def watch():
if args.resume_path is None:
args.resume_path = os.path.join(log_path, "policy.pth")
policy.load_state_dict(
torch.load(args.resume_path, map_location=torch.device("cpu")))
policy.eval()
collector = Collector(policy, env)
collector.collect(n_episode=1, render=1 / 35)
if not args.watch:
replay_buffer = load_buffer_d4rl(args.expert_data_task)
if args.norm_obs:
replay_buffer, obs_rms = normalize_all_obs_in_replay_buffer(
replay_buffer)
test_envs.set_obs_rms(obs_rms)
# trainer
result = offline_trainer(
policy,
replay_buffer,
test_collector,
args.epoch,
args.step_per_epoch,
args.test_num,
args.batch_size,
save_best_fn=save_best_fn,
logger=logger,
)
pprint.pprint(result)
else:
watch()
# 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()}"
)
writer = SummaryWriter(log_path)
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
######################################### Train #############################
# trainer
result = offline_trainer(
policy, offline_buffer, test_collector, args.epoch,
args.step_per_epoch, 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!
policy.eval()
test_envs.seed(args.seed)
test_collector.reset()
result = test_collector.collect(n_episode=[1] * args.test_num, render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
if __name__ == '__main__':
main()
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.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device)
actor = Actor(net, args.action_shape, device=args.device).to(args.device)
critic = Critic(net, device=args.device).to(args.device)
optim = torch.optim.Adam(set(actor.parameters()).union(
critic.parameters()),
lr=args.lr)
dist = torch.distributions.Categorical
policy = A2CPolicy(actor,
critic,
optim,
dist,
discount_factor=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,
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, 'a2c')
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()}")
policy.eval()
# 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(args.state_shape,
hidden_sizes=args.hidden_sizes,
device=args.device)
net = Actor(net, args.action_shape, device=args.device).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.il_step_per_epoch,
args.step_per_collect,
args.test_num,
args.batch_size,
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)
il_policy.eval()
collector = Collector(il_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_dqn(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
# train_envs = gym.make(args.task)
# you can also use tianshou.env.SubprocVectorEnv
train_envs = DummyVectorEnv(
[lambda: gym.make(args.task) for _ in range(args.training_num)])
# test_envs = gym.make(args.task)
test_envs = SubprocVectorEnv(
[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,
args.action_shape, args.device,
dueling=(2, 2)).to(args.device)
optim = torch.optim.Adam(net.parameters(), lr=args.lr)
policy = DQNPolicy(
net, optim, args.gamma, args.n_step,
target_update_freq=args.target_update_freq)
# collector
train_collector = Collector(
policy, train_envs, ReplayBuffer(args.buffer_size))
test_collector = Collector(policy, test_envs)
# policy.set_eps(1)
train_collector.collect(n_step=args.batch_size)
# log
log_path = os.path.join(args.logdir, args.task, 'dqn')
writer = SummaryWriter(log_path)
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 train_fn(epoch, env_step): # exp decay
eps = max(args.eps_train * (1 - 5e-6) ** env_step, args.eps_test)
policy.set_eps(eps)
def test_fn(epoch, env_step):
policy.set_eps(args.eps_test)
# trainer
result = offpolicy_trainer(
policy, train_collector, test_collector, args.epoch,
args.step_per_epoch, args.collect_per_step, args.test_num,
args.batch_size, train_fn=train_fn, test_fn=test_fn,
stop_fn=stop_fn, save_fn=save_fn, writer=writer,
test_in_train=False)
assert stop_fn(result['best_reward'])
if __name__ == '__main__':
pprint.pprint(result)
# Let's watch its performance!
policy.eval()
policy.set_eps(args.eps_test)
test_envs.seed(args.seed)
test_collector.reset()
result = test_collector.collect(n_episode=[1] * args.test_num,
render=args.render)
print(f'Final reward: {result["rew"]}, length: {result["len"]}')
