Yi Su dd16818ce4
implement REDQ based on original contribution by @Jimenius (#623)
Co-authored-by: Minhui Li
 <limh@lamda.nju.edu.cn>
2022-05-01 00:06:00 +08:00

201 lines
8.3 KiB
Python

from copy import deepcopy
from typing import Any, Dict, Optional, Tuple, Union
import numpy as np
import torch
from torch.distributions import Independent, Normal
from tianshou.data import Batch, ReplayBuffer
from tianshou.exploration import BaseNoise
from tianshou.policy import DDPGPolicy
class REDQPolicy(DDPGPolicy):
"""Implementation of REDQ. arXiv:2101.05982.
:param torch.nn.Module actor: the actor network following the rules in
:class:`~tianshou.policy.BasePolicy`. (s -> logits)
:param torch.optim.Optimizer actor_optim: the optimizer for actor network.
:param torch.nn.Module critics: critic ensemble networks.
:param torch.optim.Optimizer critics_optim: the optimizer for the critic networks.
:param int ensemble_size: Number of sub-networks in the critic ensemble.
Default to 10.
:param int subset_size: Number of networks in the subset. Default to 2.
:param float tau: param for soft update of the target network. Default to 0.005.
:param float gamma: discount factor, in [0, 1]. Default to 0.99.
:param (float, torch.Tensor, torch.optim.Optimizer) or float alpha: entropy
regularization coefficient. Default to 0.2.
If a tuple (target_entropy, log_alpha, alpha_optim) is provided, then
alpha is automatically tuned.
:param bool reward_normalization: normalize the reward to Normal(0, 1).
Default to False.
:param int actor_delay: Number of critic updates before an actor update.
Default to 20.
:param BaseNoise exploration_noise: add a noise to action for exploration.
Default to None. This is useful when solving hard-exploration problem.
:param bool deterministic_eval: whether to use deterministic action (mean
of Gaussian policy) instead of stochastic action sampled by the policy.
Default to True.
:param str target_mode: methods to integrate critic values in the subset,
currently support minimum and average. Default to min.
:param bool action_scaling: whether to map actions from range [-1, 1] to range
[action_spaces.low, action_spaces.high]. Default to True.
:param str 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.
Default to "clip".
:param Optional[gym.Space] action_space: env's action space, mandatory if you want
to use option "action_scaling" or "action_bound_method". Default to None.
.. seealso::
Please refer to :class:`~tianshou.policy.BasePolicy` for more detailed
explanation.
"""
def __init__(
self,
actor: torch.nn.Module,
actor_optim: torch.optim.Optimizer,
critics: torch.nn.Module,
critics_optim: torch.optim.Optimizer,
ensemble_size: int = 10,
subset_size: int = 2,
tau: float = 0.005,
gamma: float = 0.99,
alpha: Union[float, Tuple[float, torch.Tensor, torch.optim.Optimizer]] = 0.2,
reward_normalization: bool = False,
estimation_step: int = 1,
actor_delay: int = 20,
exploration_noise: Optional[BaseNoise] = None,
deterministic_eval: bool = True,
target_mode: str = "min",
**kwargs: Any,
) -> None:
super().__init__(
None, None, None, None, tau, gamma, exploration_noise,
reward_normalization, estimation_step, **kwargs
)
self.actor, self.actor_optim = actor, actor_optim
self.critics, self.critics_old = critics, deepcopy(critics)
self.critics_old.eval()
self.critics_optim = critics_optim
assert 0 < subset_size <= ensemble_size, \
"Invalid choice of ensemble size or subset size."
self.ensemble_size = ensemble_size
self.subset_size = subset_size
self._is_auto_alpha = False
self._alpha: Union[float, torch.Tensor]
if isinstance(alpha, tuple):
self._is_auto_alpha = True
self._target_entropy, self._log_alpha, self._alpha_optim = alpha
assert alpha[1].shape == torch.Size([1]) and alpha[1].requires_grad
self._alpha = self._log_alpha.detach().exp()
else:
self._alpha = alpha
if target_mode in ("min", "mean"):
self.target_mode = target_mode
else:
raise ValueError("Unsupported mode of Q target computing.")
self.critic_gradient_step = 0
self.actor_delay = actor_delay
self._deterministic_eval = deterministic_eval
self.__eps = np.finfo(np.float32).eps.item()
def train(self, mode: bool = True) -> "REDQPolicy":
self.training = mode
self.actor.train(mode)
self.critics.train(mode)
return self
def sync_weight(self) -> None:
for o, n in zip(self.critics_old.parameters(), self.critics.parameters()):
o.data.copy_(o.data * (1.0 - self.tau) + n.data * self.tau)
def forward( # type: ignore
self,
batch: Batch,
state: Optional[Union[dict, Batch, np.ndarray]] = None,
input: str = "obs",
**kwargs: Any,
) -> Batch:
obs = batch[input]
logits, h = self.actor(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)
return Batch(
logits=logits, act=squashed_action, state=h, dist=dist, log_prob=log_prob
)
def _target_q(self, buffer: ReplayBuffer, indices: np.ndarray) -> torch.Tensor:
batch = buffer[indices] # batch.obs: s_{t+n}
obs_next_result = self(batch, input="obs_next")
a_ = obs_next_result.act
sample_ensemble_idx = np.random.choice(
self.ensemble_size, self.subset_size, replace=False
)
qs = self.critics_old(batch.obs_next, a_)[sample_ensemble_idx, ...]
if self.target_mode == "min":
target_q, _ = torch.min(qs, dim=0)
elif self.target_mode == "mean":
target_q = torch.mean(qs, dim=0)
target_q -= self._alpha * obs_next_result.log_prob
return target_q
def learn(self, batch: Batch, **kwargs: Any) -> Dict[str, float]:
# critic ensemble
weight = getattr(batch, "weight", 1.0)
current_qs = self.critics(batch.obs, batch.act).flatten(1)
target_q = batch.returns.flatten()
td = current_qs - target_q
critic_loss = (td.pow(2) * weight).mean()
self.critics_optim.zero_grad()
critic_loss.backward()
self.critics_optim.step()
batch.weight = torch.mean(td, dim=0) # prio-buffer
self.critic_gradient_step += 1
# actor
if self.critic_gradient_step % self.actor_delay == 0:
obs_result = self(batch)
a = obs_result.act
current_qa = self.critics(batch.obs, a).mean(dim=0).flatten()
actor_loss = (self._alpha * obs_result.log_prob.flatten() -
current_qa).mean()
self.actor_optim.zero_grad()
actor_loss.backward()
self.actor_optim.step()
if self._is_auto_alpha:
log_prob = obs_result.log_prob.detach() + self._target_entropy
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()
result = {"loss/critics": critic_loss.item()}
if self.critic_gradient_step % self.actor_delay == 0:
result["loss/actor"] = actor_loss.item(),
if self._is_auto_alpha:
result["loss/alpha"] = alpha_loss.item()
result["alpha"] = self._alpha.item() # type: ignore
return result