mirror of
https://github.com/saymrwulf/stable-baselines3.git
synced 2026-07-25 19:48:33 +00:00
0c50d75ecb
* Removed unneeded overrides of feature_extractor and normalize_images in the TD3 Actor. * Add learning rate schedule example (#248) * Add learning rate schedule example * Update docs/guide/examples.rst Co-authored-by: Adam Gleave <adam@gleave.me> * Address comments Co-authored-by: Adam Gleave <adam@gleave.me> * Add supported action spaces checks (#254) * Add supported action spaces checks * Address comment * Use `pass` in an abstractmethod instead of deleting the arguments. * Remove the "deterministic" keyword from the forward method of the TD3 Actor since it always is deterministic anyways. * Rename _get_data to _get_data_to_reconstruct_model. _get_data was too generic and could have meant anything. * Remove the n_episodes_rollout parameter and allow passing tuples as train_freq instead. * Fix docstring of `train_freq` parameter. * Black fixes. * Fix TD3 delayed update + rename `_get_data()` * Fix TD3 test * Normalize `train_freq` to a tuple in the constructor and turn the warning into an assert. * Make one step the default train frequency. * Black fixes. * Change np.bool to bool. * Use the tuple format to specify an amount of steps in terms of steps or episodes in the collect_collouts of the off policy algorithm. * Use the tuple format to specify an amount of steps in terms of steps or episodes in the collect_collouts of HER. * Use named tuple for train freq * Rename train_freq to train_every and TrainFreq to ExperienceDuration. Also add some type annotations and documentation. * Black fixes. * Revert to train_freq * Fix terminal observation issues * Typo * Fix action noise bug in HER * Add assert when loading HER models * Update version Co-authored-by: Antonin RAFFIN <antonin.raffin@ensta.org> Co-authored-by: Adam Gleave <adam@gleave.me>
286 lines
11 KiB
Python
286 lines
11 KiB
Python
from typing import Any, Dict, List, Optional, Type, Union
|
|
|
|
import gym
|
|
import torch as th
|
|
from torch import nn
|
|
|
|
from stable_baselines3.common.policies import BasePolicy, ContinuousCritic, register_policy
|
|
from stable_baselines3.common.preprocessing import get_action_dim
|
|
from stable_baselines3.common.torch_layers import (
|
|
BaseFeaturesExtractor,
|
|
FlattenExtractor,
|
|
NatureCNN,
|
|
create_mlp,
|
|
get_actor_critic_arch,
|
|
)
|
|
from stable_baselines3.common.type_aliases import Schedule
|
|
|
|
|
|
class Actor(BasePolicy):
|
|
"""
|
|
Actor network (policy) for TD3.
|
|
|
|
:param observation_space: Obervation space
|
|
:param action_space: Action space
|
|
:param net_arch: Network architecture
|
|
:param features_extractor: Network to extract features
|
|
(a CNN when using images, a nn.Flatten() layer otherwise)
|
|
:param features_dim: Number of features
|
|
:param activation_fn: Activation function
|
|
:param normalize_images: Whether to normalize images or not,
|
|
dividing by 255.0 (True by default)
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
observation_space: gym.spaces.Space,
|
|
action_space: gym.spaces.Space,
|
|
net_arch: List[int],
|
|
features_extractor: nn.Module,
|
|
features_dim: int,
|
|
activation_fn: Type[nn.Module] = nn.ReLU,
|
|
normalize_images: bool = True,
|
|
):
|
|
super(Actor, self).__init__(
|
|
observation_space,
|
|
action_space,
|
|
features_extractor=features_extractor,
|
|
normalize_images=normalize_images,
|
|
squash_output=True,
|
|
)
|
|
|
|
self.net_arch = net_arch
|
|
self.features_dim = features_dim
|
|
self.activation_fn = activation_fn
|
|
|
|
action_dim = get_action_dim(self.action_space)
|
|
actor_net = create_mlp(features_dim, action_dim, net_arch, activation_fn, squash_output=True)
|
|
# Deterministic action
|
|
self.mu = nn.Sequential(*actor_net)
|
|
|
|
def _get_constructor_parameters(self) -> Dict[str, Any]:
|
|
data = super()._get_constructor_parameters()
|
|
|
|
data.update(
|
|
dict(
|
|
net_arch=self.net_arch,
|
|
features_dim=self.features_dim,
|
|
activation_fn=self.activation_fn,
|
|
features_extractor=self.features_extractor,
|
|
)
|
|
)
|
|
return data
|
|
|
|
def forward(self, obs: th.Tensor) -> th.Tensor:
|
|
# assert deterministic, 'The TD3 actor only outputs deterministic actions'
|
|
features = self.extract_features(obs)
|
|
return self.mu(features)
|
|
|
|
def _predict(self, observation: th.Tensor, deterministic: bool = False) -> th.Tensor:
|
|
# Note: the deterministic deterministic parameter is ignored in the case of TD3.
|
|
# Predictions are always deterministic.
|
|
return self.forward(observation)
|
|
|
|
|
|
class TD3Policy(BasePolicy):
|
|
"""
|
|
Policy class (with both actor and critic) for TD3.
|
|
|
|
:param observation_space: Observation space
|
|
:param action_space: Action space
|
|
:param lr_schedule: Learning rate schedule (could be constant)
|
|
:param net_arch: The specification of the policy and value networks.
|
|
:param activation_fn: Activation function
|
|
:param features_extractor_class: Features extractor to use.
|
|
:param features_extractor_kwargs: Keyword arguments
|
|
to pass to the features extractor.
|
|
:param normalize_images: Whether to normalize images or not,
|
|
dividing by 255.0 (True by default)
|
|
:param optimizer_class: The optimizer to use,
|
|
``th.optim.Adam`` by default
|
|
:param optimizer_kwargs: Additional keyword arguments,
|
|
excluding the learning rate, to pass to the optimizer
|
|
:param n_critics: Number of critic networks to create.
|
|
:param share_features_extractor: Whether to share or not the features extractor
|
|
between the actor and the critic (this saves computation time)
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
observation_space: gym.spaces.Space,
|
|
action_space: gym.spaces.Space,
|
|
lr_schedule: Schedule,
|
|
net_arch: Optional[Union[List[int], Dict[str, List[int]]]] = None,
|
|
activation_fn: Type[nn.Module] = nn.ReLU,
|
|
features_extractor_class: Type[BaseFeaturesExtractor] = FlattenExtractor,
|
|
features_extractor_kwargs: Optional[Dict[str, Any]] = None,
|
|
normalize_images: bool = True,
|
|
optimizer_class: Type[th.optim.Optimizer] = th.optim.Adam,
|
|
optimizer_kwargs: Optional[Dict[str, Any]] = None,
|
|
n_critics: int = 2,
|
|
share_features_extractor: bool = True,
|
|
):
|
|
super(TD3Policy, self).__init__(
|
|
observation_space,
|
|
action_space,
|
|
features_extractor_class,
|
|
features_extractor_kwargs,
|
|
optimizer_class=optimizer_class,
|
|
optimizer_kwargs=optimizer_kwargs,
|
|
squash_output=True,
|
|
)
|
|
|
|
# Default network architecture, from the original paper
|
|
if net_arch is None:
|
|
if features_extractor_class == FlattenExtractor:
|
|
net_arch = [400, 300]
|
|
else:
|
|
net_arch = []
|
|
|
|
actor_arch, critic_arch = get_actor_critic_arch(net_arch)
|
|
|
|
self.net_arch = net_arch
|
|
self.activation_fn = activation_fn
|
|
self.net_args = {
|
|
"observation_space": self.observation_space,
|
|
"action_space": self.action_space,
|
|
"net_arch": actor_arch,
|
|
"activation_fn": self.activation_fn,
|
|
"normalize_images": normalize_images,
|
|
}
|
|
self.actor_kwargs = self.net_args.copy()
|
|
self.critic_kwargs = self.net_args.copy()
|
|
self.critic_kwargs.update(
|
|
{
|
|
"n_critics": n_critics,
|
|
"net_arch": critic_arch,
|
|
"share_features_extractor": share_features_extractor,
|
|
}
|
|
)
|
|
|
|
self.actor, self.actor_target = None, None
|
|
self.critic, self.critic_target = None, None
|
|
self.share_features_extractor = share_features_extractor
|
|
|
|
self._build(lr_schedule)
|
|
|
|
def _build(self, lr_schedule: Schedule) -> None:
|
|
# Create actor and target
|
|
# the features extractor should not be shared
|
|
self.actor = self.make_actor(features_extractor=None)
|
|
self.actor_target = self.make_actor(features_extractor=None)
|
|
# Initialize the target to have the same weights as the actor
|
|
self.actor_target.load_state_dict(self.actor.state_dict())
|
|
|
|
self.actor.optimizer = self.optimizer_class(self.actor.parameters(), lr=lr_schedule(1), **self.optimizer_kwargs)
|
|
|
|
if self.share_features_extractor:
|
|
self.critic = self.make_critic(features_extractor=self.actor.features_extractor)
|
|
# Critic target should not share the features extactor with critic
|
|
# but it can share it with the actor target as actor and critic are sharing
|
|
# the same features_extractor too
|
|
# NOTE: as a result the effective poliak (soft-copy) coefficient for the features extractor
|
|
# will be 2 * tau instead of tau (updated one time with the actor, a second time with the critic)
|
|
self.critic_target = self.make_critic(features_extractor=self.actor_target.features_extractor)
|
|
else:
|
|
# Create new features extractor for each network
|
|
self.critic = self.make_critic(features_extractor=None)
|
|
self.critic_target = self.make_critic(features_extractor=None)
|
|
|
|
self.critic_target.load_state_dict(self.critic.state_dict())
|
|
self.critic.optimizer = self.optimizer_class(self.critic.parameters(), lr=lr_schedule(1), **self.optimizer_kwargs)
|
|
|
|
def _get_constructor_parameters(self) -> Dict[str, Any]:
|
|
data = super()._get_constructor_parameters()
|
|
|
|
data.update(
|
|
dict(
|
|
net_arch=self.net_arch,
|
|
activation_fn=self.net_args["activation_fn"],
|
|
n_critics=self.critic_kwargs["n_critics"],
|
|
lr_schedule=self._dummy_schedule, # dummy lr schedule, not needed for loading policy alone
|
|
optimizer_class=self.optimizer_class,
|
|
optimizer_kwargs=self.optimizer_kwargs,
|
|
features_extractor_class=self.features_extractor_class,
|
|
features_extractor_kwargs=self.features_extractor_kwargs,
|
|
share_features_extractor=self.share_features_extractor,
|
|
)
|
|
)
|
|
return data
|
|
|
|
def make_actor(self, features_extractor: Optional[BaseFeaturesExtractor] = None) -> Actor:
|
|
actor_kwargs = self._update_features_extractor(self.actor_kwargs, features_extractor)
|
|
return Actor(**actor_kwargs).to(self.device)
|
|
|
|
def make_critic(self, features_extractor: Optional[BaseFeaturesExtractor] = None) -> ContinuousCritic:
|
|
critic_kwargs = self._update_features_extractor(self.critic_kwargs, features_extractor)
|
|
return ContinuousCritic(**critic_kwargs).to(self.device)
|
|
|
|
def forward(self, observation: th.Tensor, deterministic: bool = False) -> th.Tensor:
|
|
return self._predict(observation, deterministic=deterministic)
|
|
|
|
def _predict(self, observation: th.Tensor, deterministic: bool = False) -> th.Tensor:
|
|
# Note: the deterministic deterministic parameter is ignored in the case of TD3.
|
|
# Predictions are always deterministic.
|
|
return self.actor(observation)
|
|
|
|
|
|
MlpPolicy = TD3Policy
|
|
|
|
|
|
class CnnPolicy(TD3Policy):
|
|
"""
|
|
Policy class (with both actor and critic) for TD3.
|
|
|
|
:param observation_space: Observation space
|
|
:param action_space: Action space
|
|
:param lr_schedule: Learning rate schedule (could be constant)
|
|
:param net_arch: The specification of the policy and value networks.
|
|
:param activation_fn: Activation function
|
|
:param features_extractor_class: Features extractor to use.
|
|
:param features_extractor_kwargs: Keyword arguments
|
|
to pass to the features extractor.
|
|
:param normalize_images: Whether to normalize images or not,
|
|
dividing by 255.0 (True by default)
|
|
:param optimizer_class: The optimizer to use,
|
|
``th.optim.Adam`` by default
|
|
:param optimizer_kwargs: Additional keyword arguments,
|
|
excluding the learning rate, to pass to the optimizer
|
|
:param n_critics: Number of critic networks to create.
|
|
:param share_features_extractor: Whether to share or not the features extractor
|
|
between the actor and the critic (this saves computation time)
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
observation_space: gym.spaces.Space,
|
|
action_space: gym.spaces.Space,
|
|
lr_schedule: Schedule,
|
|
net_arch: Optional[Union[List[int], Dict[str, List[int]]]] = None,
|
|
activation_fn: Type[nn.Module] = nn.ReLU,
|
|
features_extractor_class: Type[BaseFeaturesExtractor] = NatureCNN,
|
|
features_extractor_kwargs: Optional[Dict[str, Any]] = None,
|
|
normalize_images: bool = True,
|
|
optimizer_class: Type[th.optim.Optimizer] = th.optim.Adam,
|
|
optimizer_kwargs: Optional[Dict[str, Any]] = None,
|
|
n_critics: int = 2,
|
|
share_features_extractor: bool = True,
|
|
):
|
|
super(CnnPolicy, self).__init__(
|
|
observation_space,
|
|
action_space,
|
|
lr_schedule,
|
|
net_arch,
|
|
activation_fn,
|
|
features_extractor_class,
|
|
features_extractor_kwargs,
|
|
normalize_images,
|
|
optimizer_class,
|
|
optimizer_kwargs,
|
|
n_critics,
|
|
share_features_extractor,
|
|
)
|
|
|
|
|
|
register_policy("MlpPolicy", MlpPolicy)
|
|
register_policy("CnnPolicy", CnnPolicy)
|