This PR adds strict typing to the output of `update` and `learn` in all policies. This will likely be the last large refactoring PR before the next release (0.6.0, not 1.0.0), so it requires some attention. Several difficulties were encountered on the path to that goal: 1. The policy hierarchy is actually "broken" in the sense that the keys of dicts that were output by `learn` did not follow the same enhancement (inheritance) pattern as the policies. This is a real problem and should be addressed in the near future. Generally, several aspects of the policy design and hierarchy might deserve a dedicated discussion. 2. Each policy needs to be generic in the stats return type, because one might want to extend it at some point and then also extend the stats. Even within the source code base this pattern is necessary in many places. 3. The interaction between learn and update is a bit quirky, we currently handle it by having update modify special field inside TrainingStats, whereas all other fields are handled by learn. 4. The IQM module is a policy wrapper and required a TrainingStatsWrapper. The latter relies on a bunch of black magic. They were addressed by: 1. Live with the broken hierarchy, which is now made visible by bounds in generics. We use type: ignore where appropriate. 2. Make all policies generic with bounds following the policy inheritance hierarchy (which is incorrect, see above). We experimented a bit with nested TrainingStats classes, but that seemed to add more complexity and be harder to understand. Unfortunately, mypy thinks that the code below is wrong, wherefore we have to add `type: ignore` to the return of each `learn` ```python T = TypeVar("T", bound=int) def f() -> T: return 3 ``` 3. See above 4. Write representative tests for the `TrainingStatsWrapper`. Still, the black magic might cause nasty surprises down the line (I am not proud of it)... Closes #933 --------- Co-authored-by: Maximilian Huettenrauch <m.huettenrauch@appliedai.de> Co-authored-by: Michael Panchenko <m.panchenko@appliedai.de>
253 lines
10 KiB
Python
253 lines
10 KiB
Python
from copy import deepcopy
|
|
from dataclasses import dataclass
|
|
from typing import Any, Generic, Literal, Self, TypeVar, cast
|
|
|
|
import gymnasium as gym
|
|
import numpy as np
|
|
import torch
|
|
from torch.distributions import Independent, Normal
|
|
|
|
from tianshou.data import Batch, ReplayBuffer
|
|
from tianshou.data.types import (
|
|
DistLogProbBatchProtocol,
|
|
ObsBatchProtocol,
|
|
RolloutBatchProtocol,
|
|
)
|
|
from tianshou.exploration import BaseNoise
|
|
from tianshou.policy import DDPGPolicy
|
|
from tianshou.policy.base import TLearningRateScheduler, TrainingStats
|
|
from tianshou.utils.conversion import to_optional_float
|
|
from tianshou.utils.optim import clone_optimizer
|
|
|
|
|
|
@dataclass(kw_only=True)
|
|
class SACTrainingStats(TrainingStats):
|
|
actor_loss: float
|
|
critic1_loss: float
|
|
critic2_loss: float
|
|
alpha: float | None = None
|
|
alpha_loss: float | None = None
|
|
|
|
|
|
TSACTrainingStats = TypeVar("TSACTrainingStats", bound=SACTrainingStats)
|
|
|
|
|
|
# TODO: the type ignore here is needed b/c the hierarchy is actually broken! Should reconsider the inheritance structure.
|
|
class SACPolicy(DDPGPolicy[TSACTrainingStats], Generic[TSACTrainingStats]): # type: ignore[type-var]
|
|
"""Implementation of Soft Actor-Critic. arXiv:1812.05905.
|
|
|
|
:param actor: the actor network following the rules in
|
|
:class:`~tianshou.policy.BasePolicy`. (s -> logits)
|
|
:param actor_optim: the optimizer for actor network.
|
|
:param critic: the first critic network. (s, a -> Q(s, a))
|
|
:param critic_optim: the optimizer for the first critic network.
|
|
:param action_space: Env's action space. Should be gym.spaces.Box.
|
|
:param critic2: the second critic network. (s, a -> Q(s, a)).
|
|
If None, use the same network as critic (via deepcopy).
|
|
:param critic2_optim: the optimizer for the second critic network.
|
|
If None, clone critic_optim to use for critic2.parameters().
|
|
:param tau: param for soft update of the target network.
|
|
:param gamma: discount factor, in [0, 1].
|
|
:param alpha: entropy regularization coefficient.
|
|
If a tuple (target_entropy, log_alpha, alpha_optim) is provided,
|
|
then alpha is automatically tuned.
|
|
:param estimation_step: The number of steps to look ahead.
|
|
:param exploration_noise: add noise to action for exploration.
|
|
This is useful when solving "hard exploration" problems.
|
|
"default" is equivalent to GaussianNoise(sigma=0.1).
|
|
:param deterministic_eval: whether to use deterministic action
|
|
(mean of Gaussian policy) in evaluation mode instead of stochastic
|
|
action sampled by the policy. Does not affect training.
|
|
:param action_scaling: whether to map actions from range [-1, 1]
|
|
to range[action_spaces.low, action_spaces.high].
|
|
:param action_bound_method: method to bound action to range [-1, 1],
|
|
can be either "clip" (for simply clipping the action)
|
|
or empty string for no bounding. Only used if the action_space is continuous.
|
|
:param observation_space: Env's observation space.
|
|
:param lr_scheduler: a learning rate scheduler that adjusts the learning rate
|
|
in optimizer in each policy.update()
|
|
|
|
.. seealso::
|
|
|
|
Please refer to :class:`~tianshou.policy.BasePolicy` for more detailed
|
|
explanation.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
*,
|
|
actor: torch.nn.Module,
|
|
actor_optim: torch.optim.Optimizer,
|
|
critic: torch.nn.Module,
|
|
critic_optim: torch.optim.Optimizer,
|
|
action_space: gym.Space,
|
|
critic2: torch.nn.Module | None = None,
|
|
critic2_optim: torch.optim.Optimizer | None = None,
|
|
tau: float = 0.005,
|
|
gamma: float = 0.99,
|
|
alpha: float | tuple[float, torch.Tensor, torch.optim.Optimizer] = 0.2,
|
|
estimation_step: int = 1,
|
|
exploration_noise: BaseNoise | Literal["default"] | None = None,
|
|
deterministic_eval: bool = True,
|
|
action_scaling: bool = True,
|
|
# TODO: some papers claim that tanh is crucial for SAC, yet DDPG will raise an
|
|
# error if tanh is used. Should be investigated.
|
|
action_bound_method: Literal["clip"] | None = "clip",
|
|
observation_space: gym.Space | None = None,
|
|
lr_scheduler: TLearningRateScheduler | None = None,
|
|
) -> None:
|
|
super().__init__(
|
|
actor=actor,
|
|
actor_optim=actor_optim,
|
|
critic=critic,
|
|
critic_optim=critic_optim,
|
|
action_space=action_space,
|
|
tau=tau,
|
|
gamma=gamma,
|
|
exploration_noise=exploration_noise,
|
|
estimation_step=estimation_step,
|
|
action_scaling=action_scaling,
|
|
action_bound_method=action_bound_method,
|
|
observation_space=observation_space,
|
|
lr_scheduler=lr_scheduler,
|
|
)
|
|
critic2 = critic2 or deepcopy(critic)
|
|
critic2_optim = critic2_optim or clone_optimizer(critic_optim, critic2.parameters())
|
|
self.critic2, self.critic2_old = critic2, deepcopy(critic2)
|
|
self.critic2_old.eval()
|
|
self.critic2_optim = critic2_optim
|
|
self.deterministic_eval = deterministic_eval
|
|
self.__eps = np.finfo(np.float32).eps.item()
|
|
|
|
self.alpha: float | torch.Tensor
|
|
self._is_auto_alpha = not isinstance(alpha, float)
|
|
if self._is_auto_alpha:
|
|
# TODO: why doesn't mypy understand that this must be a tuple?
|
|
alpha = cast(tuple[float, torch.Tensor, torch.optim.Optimizer], alpha)
|
|
if alpha[1].shape != torch.Size([1]):
|
|
raise ValueError(
|
|
f"Expected log_alpha to have shape torch.Size([1]), "
|
|
f"but got {alpha[1].shape} instead.",
|
|
)
|
|
if not alpha[1].requires_grad:
|
|
raise ValueError("Expected log_alpha to require gradient, but it doesn't.")
|
|
|
|
self.target_entropy, self.log_alpha, self.alpha_optim = alpha
|
|
self.alpha = self.log_alpha.detach().exp()
|
|
else:
|
|
alpha = cast(
|
|
float,
|
|
alpha,
|
|
) # can we convert alpha to a constant tensor here? then mypy wouldn't complain
|
|
self.alpha = alpha
|
|
|
|
# TODO or not TODO: add to BasePolicy?
|
|
self._check_field_validity()
|
|
|
|
def _check_field_validity(self) -> None:
|
|
if not isinstance(self.action_space, gym.spaces.Box):
|
|
raise ValueError(
|
|
f"SACPolicy only supports gym.spaces.Box, but got {self.action_space=}."
|
|
f"Please use DiscreteSACPolicy for discrete action spaces.",
|
|
)
|
|
|
|
@property
|
|
def is_auto_alpha(self) -> bool:
|
|
return self._is_auto_alpha
|
|
|
|
def train(self, mode: bool = True) -> Self:
|
|
self.training = mode
|
|
self.actor.train(mode)
|
|
self.critic.train(mode)
|
|
self.critic2.train(mode)
|
|
return self
|
|
|
|
def sync_weight(self) -> None:
|
|
self.soft_update(self.critic_old, self.critic, self.tau)
|
|
self.soft_update(self.critic2_old, self.critic2, self.tau)
|
|
|
|
# TODO: violates Liskov substitution principle
|
|
def forward( # type: ignore
|
|
self,
|
|
batch: ObsBatchProtocol,
|
|
state: dict | Batch | np.ndarray | None = None,
|
|
**kwargs: Any,
|
|
) -> DistLogProbBatchProtocol:
|
|
logits, hidden = self.actor(batch.obs, state=state, info=batch.info)
|
|
assert isinstance(logits, tuple)
|
|
dist = Independent(Normal(*logits), 1)
|
|
if self.deterministic_eval and not self.training:
|
|
act = logits[0]
|
|
else:
|
|
act = dist.rsample()
|
|
log_prob = dist.log_prob(act).unsqueeze(-1)
|
|
# apply correction for Tanh squashing when computing logprob from Gaussian
|
|
# You can check out the original SAC paper (arXiv 1801.01290): Eq 21.
|
|
# in appendix C to get some understanding of this equation.
|
|
squashed_action = torch.tanh(act)
|
|
log_prob = log_prob - torch.log((1 - squashed_action.pow(2)) + self.__eps).sum(
|
|
-1,
|
|
keepdim=True,
|
|
)
|
|
result = Batch(
|
|
logits=logits,
|
|
act=squashed_action,
|
|
state=hidden,
|
|
dist=dist,
|
|
log_prob=log_prob,
|
|
)
|
|
return cast(DistLogProbBatchProtocol, result)
|
|
|
|
def _target_q(self, buffer: ReplayBuffer, indices: np.ndarray) -> torch.Tensor:
|
|
obs_next_batch = Batch(
|
|
obs=buffer[indices].obs_next,
|
|
info=[None] * len(indices),
|
|
) # obs_next: s_{t+n}
|
|
obs_next_result = self(obs_next_batch)
|
|
act_ = obs_next_result.act
|
|
return (
|
|
torch.min(
|
|
self.critic_old(obs_next_batch.obs, act_),
|
|
self.critic2_old(obs_next_batch.obs, act_),
|
|
)
|
|
- self.alpha * obs_next_result.log_prob
|
|
)
|
|
|
|
def learn(self, batch: RolloutBatchProtocol, *args: Any, **kwargs: Any) -> TSACTrainingStats: # type: ignore
|
|
# critic 1&2
|
|
td1, critic1_loss = self._mse_optimizer(batch, self.critic, self.critic_optim)
|
|
td2, critic2_loss = self._mse_optimizer(batch, self.critic2, self.critic2_optim)
|
|
batch.weight = (td1 + td2) / 2.0 # prio-buffer
|
|
|
|
# actor
|
|
obs_result = self(batch)
|
|
act = obs_result.act
|
|
current_q1a = self.critic(batch.obs, act).flatten()
|
|
current_q2a = self.critic2(batch.obs, act).flatten()
|
|
actor_loss = (
|
|
self.alpha * obs_result.log_prob.flatten() - torch.min(current_q1a, current_q2a)
|
|
).mean()
|
|
self.actor_optim.zero_grad()
|
|
actor_loss.backward()
|
|
self.actor_optim.step()
|
|
alpha_loss = None
|
|
|
|
if self.is_auto_alpha:
|
|
log_prob = obs_result.log_prob.detach() + self.target_entropy
|
|
# please take a look at issue #258 if you'd like to change this line
|
|
alpha_loss = -(self.log_alpha * log_prob).mean()
|
|
self.alpha_optim.zero_grad()
|
|
alpha_loss.backward()
|
|
self.alpha_optim.step()
|
|
self.alpha = self.log_alpha.detach().exp()
|
|
|
|
self.sync_weight()
|
|
|
|
return SACTrainingStats( # type: ignore[return-value]
|
|
actor_loss=actor_loss.item(),
|
|
critic1_loss=critic1_loss.item(),
|
|
critic2_loss=critic2_loss.item(),
|
|
alpha=to_optional_float(self.alpha),
|
|
alpha_loss=to_optional_float(alpha_loss),
|
|
)
|