522f7fbf98
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>
195 lines
8.6 KiB
Python
195 lines
8.6 KiB
Python
import warnings
|
|
from dataclasses import dataclass
|
|
from typing import Any, Literal, TypeVar
|
|
|
|
import gymnasium as gym
|
|
import torch
|
|
import torch.nn.functional as F
|
|
from torch.distributions import kl_divergence
|
|
|
|
from tianshou.data import Batch, SequenceSummaryStats
|
|
from tianshou.policy import NPGPolicy
|
|
from tianshou.policy.base import TLearningRateScheduler
|
|
from tianshou.policy.modelfree.npg import NPGTrainingStats
|
|
from tianshou.policy.modelfree.pg import TDistributionFunction
|
|
|
|
|
|
@dataclass(kw_only=True)
|
|
class TRPOTrainingStats(NPGTrainingStats):
|
|
step_size: SequenceSummaryStats
|
|
|
|
|
|
TTRPOTrainingStats = TypeVar("TTRPOTrainingStats", bound=TRPOTrainingStats)
|
|
|
|
|
|
class TRPOPolicy(NPGPolicy[TTRPOTrainingStats]):
|
|
"""Implementation of Trust Region Policy Optimization. arXiv:1502.05477.
|
|
|
|
:param actor: the actor network following the rules in BasePolicy. (s -> logits)
|
|
:param critic: the critic network. (s -> V(s))
|
|
:param optim: the optimizer for actor and critic network.
|
|
:param dist_fn: distribution class for computing the action.
|
|
:param action_space: env's action space
|
|
:param max_kl: max kl-divergence used to constrain each actor network update.
|
|
:param backtrack_coeff: Coefficient to be multiplied by step size when
|
|
constraints are not met.
|
|
:param max_backtracks: Max number of backtracking times in linesearch.
|
|
:param optim_critic_iters: Number of times to optimize critic network per update.
|
|
:param actor_step_size: step size for actor update in natural gradient direction.
|
|
:param advantage_normalization: whether to do per mini-batch advantage
|
|
normalization.
|
|
:param gae_lambda: in [0, 1], param for Generalized Advantage Estimation.
|
|
:param max_batchsize: the maximum size of the batch when computing GAE.
|
|
:param discount_factor: in [0, 1].
|
|
:param reward_normalization: normalize estimated values to have std close to 1.
|
|
:param deterministic_eval: if True, use deterministic evaluation.
|
|
:param observation_space: the space of the observation.
|
|
:param action_scaling: if True, scale the action from [-1, 1] to the range of
|
|
action_space. Only used if the action_space is continuous.
|
|
:param action_bound_method: method to bound action to range [-1, 1].
|
|
:param lr_scheduler: if not None, will be called in `policy.update()`.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
*,
|
|
actor: torch.nn.Module,
|
|
critic: torch.nn.Module,
|
|
optim: torch.optim.Optimizer,
|
|
dist_fn: TDistributionFunction,
|
|
action_space: gym.Space,
|
|
max_kl: float = 0.01,
|
|
backtrack_coeff: float = 0.8,
|
|
max_backtracks: int = 10,
|
|
optim_critic_iters: int = 5,
|
|
actor_step_size: float = 0.5,
|
|
advantage_normalization: bool = True,
|
|
gae_lambda: float = 0.95,
|
|
max_batchsize: int = 256,
|
|
discount_factor: float = 0.99,
|
|
# TODO: rename to return_normalization?
|
|
reward_normalization: bool = False,
|
|
deterministic_eval: bool = False,
|
|
observation_space: gym.Space | None = None,
|
|
action_scaling: bool = True,
|
|
action_bound_method: Literal["clip", "tanh"] | None = "clip",
|
|
lr_scheduler: TLearningRateScheduler | None = None,
|
|
) -> None:
|
|
super().__init__(
|
|
actor=actor,
|
|
critic=critic,
|
|
optim=optim,
|
|
dist_fn=dist_fn,
|
|
action_space=action_space,
|
|
optim_critic_iters=optim_critic_iters,
|
|
actor_step_size=actor_step_size,
|
|
advantage_normalization=advantage_normalization,
|
|
gae_lambda=gae_lambda,
|
|
max_batchsize=max_batchsize,
|
|
discount_factor=discount_factor,
|
|
reward_normalization=reward_normalization,
|
|
deterministic_eval=deterministic_eval,
|
|
observation_space=observation_space,
|
|
action_scaling=action_scaling,
|
|
action_bound_method=action_bound_method,
|
|
lr_scheduler=lr_scheduler,
|
|
)
|
|
self.max_backtracks = max_backtracks
|
|
self.max_kl = max_kl
|
|
self.backtrack_coeff = backtrack_coeff
|
|
|
|
def learn( # type: ignore
|
|
self,
|
|
batch: Batch,
|
|
batch_size: int | None,
|
|
repeat: int,
|
|
**kwargs: Any,
|
|
) -> TTRPOTrainingStats:
|
|
actor_losses, vf_losses, step_sizes, kls = [], [], [], []
|
|
split_batch_size = batch_size or -1
|
|
for _ in range(repeat):
|
|
for minibatch in batch.split(split_batch_size, merge_last=True):
|
|
# optimize actor
|
|
# direction: calculate villia gradient
|
|
dist = self(minibatch).dist # TODO could come from batch
|
|
ratio = (dist.log_prob(minibatch.act) - minibatch.logp_old).exp().float()
|
|
ratio = ratio.reshape(ratio.size(0), -1).transpose(0, 1)
|
|
actor_loss = -(ratio * minibatch.adv).mean()
|
|
flat_grads = self._get_flat_grad(actor_loss, self.actor, retain_graph=True).detach()
|
|
|
|
# direction: calculate natural gradient
|
|
with torch.no_grad():
|
|
old_dist = self(minibatch).dist
|
|
|
|
kl = kl_divergence(old_dist, dist).mean()
|
|
# calculate first order gradient of kl with respect to theta
|
|
flat_kl_grad = self._get_flat_grad(kl, self.actor, create_graph=True)
|
|
search_direction = -self._conjugate_gradients(flat_grads, flat_kl_grad, nsteps=10)
|
|
|
|
# stepsize: calculate max stepsize constrained by kl bound
|
|
step_size = torch.sqrt(
|
|
2
|
|
* self.max_kl
|
|
/ (search_direction * self._MVP(search_direction, flat_kl_grad)).sum(
|
|
0,
|
|
keepdim=True,
|
|
),
|
|
)
|
|
|
|
# stepsize: linesearch stepsize
|
|
with torch.no_grad():
|
|
flat_params = torch.cat(
|
|
[param.data.view(-1) for param in self.actor.parameters()],
|
|
)
|
|
for i in range(self.max_backtracks):
|
|
new_flat_params = flat_params + step_size * search_direction
|
|
self._set_from_flat_params(self.actor, new_flat_params)
|
|
# calculate kl and if in bound, loss actually down
|
|
new_dist = self(minibatch).dist
|
|
new_dratio = (
|
|
(new_dist.log_prob(minibatch.act) - minibatch.logp_old).exp().float()
|
|
)
|
|
new_dratio = new_dratio.reshape(new_dratio.size(0), -1).transpose(0, 1)
|
|
new_actor_loss = -(new_dratio * minibatch.adv).mean()
|
|
kl = kl_divergence(old_dist, new_dist).mean()
|
|
|
|
if kl < self.max_kl and new_actor_loss < actor_loss:
|
|
if i > 0:
|
|
warnings.warn(f"Backtracking to step {i}.")
|
|
break
|
|
if i < self.max_backtracks - 1:
|
|
step_size = step_size * self.backtrack_coeff
|
|
else:
|
|
self._set_from_flat_params(self.actor, new_flat_params)
|
|
step_size = torch.tensor([0.0])
|
|
warnings.warn(
|
|
"Line search failed! It seems hyperparamters"
|
|
" are poor and need to be changed.",
|
|
)
|
|
|
|
# optimize critic
|
|
# TODO: remove type-ignore once the top-level type-ignore is removed
|
|
for _ in range(self.optim_critic_iters): # type: ignore
|
|
value = self.critic(minibatch.obs).flatten()
|
|
vf_loss = F.mse_loss(minibatch.returns, value)
|
|
self.optim.zero_grad()
|
|
vf_loss.backward()
|
|
self.optim.step()
|
|
|
|
actor_losses.append(actor_loss.item())
|
|
vf_losses.append(vf_loss.item())
|
|
step_sizes.append(step_size.item())
|
|
kls.append(kl.item())
|
|
|
|
actor_loss_summary_stat = SequenceSummaryStats.from_sequence(actor_losses)
|
|
vf_loss_summary_stat = SequenceSummaryStats.from_sequence(vf_losses)
|
|
kl_summary_stat = SequenceSummaryStats.from_sequence(kls)
|
|
step_size_stat = SequenceSummaryStats.from_sequence(step_sizes)
|
|
|
|
return TRPOTrainingStats( # type: ignore[return-value]
|
|
actor_loss=actor_loss_summary_stat,
|
|
vf_loss=vf_loss_summary_stat,
|
|
kl=kl_summary_stat,
|
|
step_size=step_size_stat,
|
|
)
|