# Copyright 2023 OmniSafe Team. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Implementation of ActorQCritic."""
from __future__ import annotations
from copy import deepcopy
import torch
from torch import nn, optim
from torch.optim.lr_scheduler import ConstantLR, LinearLR
from omnisafe.models.actor import GaussianLearningActor, GaussianSACActor, MLPActor
from omnisafe.models.actor.actor_builder import ActorBuilder
from omnisafe.models.base import Critic
from omnisafe.models.critic.critic_builder import CriticBuilder
from omnisafe.typing import OmnisafeSpace
from omnisafe.utils.config import ModelConfig
[docs]class ActorQCritic(nn.Module):
"""Class for ActorQCritic.
In OmniSafe, we combine the actor and critic into one this class.
+-----------------+---------------------------------------------------+
| Model | Description |
+=================+===================================================+
| Actor | Input is observation. Output is action. |
+-----------------+---------------------------------------------------+
| Reward Q Critic | Input is obs-action pair. Output is reward value. |
+-----------------+---------------------------------------------------+
Args:
obs_space (OmnisafeSpace): The observation space.
act_space (OmnisafeSpace): The action space.
model_cfgs (ModelConfig): The model configurations.
epochs (int): The number of epochs.
Attributes:
actor (Actor): The actor network.
target_actor (Actor): The target actor network.
reward_critic (Critic): The critic network.
target_reward_critic (Critic): The target critic network.
actor_optimizer (Optimizer): The optimizer for the actor network.
reward_critic_optimizer (Optimizer): The optimizer for the critic network.
std_schedule (Schedule): The schedule for the standard deviation of the Gaussian distribution.
"""
# pylint: disable-next=too-many-arguments
def __init__(
self,
obs_space: OmnisafeSpace,
act_space: OmnisafeSpace,
model_cfgs: ModelConfig,
epochs: int,
) -> None:
"""Initialize an instance of :class:`ActorQCritic`."""
super().__init__()
self.actor: GaussianLearningActor | GaussianSACActor | MLPActor = ActorBuilder(
obs_space=obs_space,
act_space=act_space,
hidden_sizes=model_cfgs.actor.hidden_sizes,
activation=model_cfgs.actor.activation,
weight_initialization_mode=model_cfgs.weight_initialization_mode,
).build_actor(actor_type=model_cfgs.actor_type)
self.reward_critic: Critic = CriticBuilder(
obs_space=obs_space,
act_space=act_space,
hidden_sizes=model_cfgs.critic.hidden_sizes,
activation=model_cfgs.critic.activation,
weight_initialization_mode=model_cfgs.weight_initialization_mode,
num_critics=model_cfgs.critic.num_critics,
use_obs_encoder=False,
).build_critic(critic_type='q')
self.target_reward_critic: Critic = deepcopy(self.reward_critic)
for param in self.target_reward_critic.parameters():
param.requires_grad = False
self.target_actor: GaussianLearningActor | GaussianSACActor | MLPActor = deepcopy(
self.actor,
)
for param in self.target_actor.parameters():
param.requires_grad = False
self.add_module('actor', self.actor)
self.add_module('reward_critic', self.reward_critic)
if model_cfgs.actor.lr is not None:
self.actor_optimizer: optim.Optimizer
self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=model_cfgs.actor.lr)
if model_cfgs.critic.lr is not None:
self.reward_critic_optimizer: optim.Optimizer
self.reward_critic_optimizer = optim.Adam(
self.reward_critic.parameters(),
lr=model_cfgs.critic.lr,
)
self.actor_scheduler: LinearLR | ConstantLR
if model_cfgs.linear_lr_decay:
self.actor_scheduler = LinearLR(
self.actor_optimizer,
start_factor=1.0,
end_factor=0.0,
total_iters=epochs,
verbose=True,
)
else:
self.actor_scheduler = ConstantLR(
self.actor_optimizer,
factor=1.0,
total_iters=epochs,
verbose=True,
)
[docs] def step(self, obs: torch.Tensor, deterministic: bool = False) -> torch.Tensor:
"""Choose the action based on the observation. used in rollout without gradient.
Args:
obs (torch.tensor): The observation.
deterministic (bool, optional): Whether to use deterministic action. Defaults to False.
Returns:
The deterministic action if ``deterministic`` is True, otherwise the action with
Gaussian noise.
"""
with torch.no_grad():
return self.actor.predict(obs, deterministic=deterministic)
[docs] def forward(self, obs: torch.Tensor, deterministic: bool = False) -> torch.Tensor:
"""Choose the action based on the observation. used in training with gradient.
Args:
obs (torch.tensor): The observation.
deterministic (bool, optional): Whether to use deterministic action. Defaults to False.
Returns:
The deterministic action if ``deterministic`` is True, otherwise the action with
Gaussian noise.
"""
return self.step(obs, deterministic=deterministic)
[docs] def polyak_update(self, tau: float) -> None:
"""Update the target network with polyak averaging.
Args:
tau (float): The polyak averaging factor.
"""
for param, target_param in zip(
self.reward_critic.parameters(),
self.target_reward_critic.parameters(),
):
target_param.data.copy_(tau * param.data + (1 - tau) * target_param.data)
for param, target_param in zip(self.actor.parameters(), self.target_actor.parameters()):
target_param.data.copy_(tau * param.data + (1 - tau) * target_param.data)