from typing import Union, Dict, List, Tuple, Callable, Any
from copy import deepcopy
import torch as t
import torch.nn as nn
import numpy as np
from machin.frame.buffers.buffer import Buffer
from machin.frame.transition import Transition
from machin.model.nets.base import NeuralNetworkModule
from .base import TorchFramework, Config
from .utils import (
hard_update,
soft_update,
safe_call,
safe_return,
assert_and_get_valid_models,
assert_and_get_valid_optimizer,
assert_and_get_valid_criterion,
assert_and_get_valid_lr_scheduler,
)
[docs]class SAC(TorchFramework):
"""
SAC framework.
"""
_is_top = ["actor", "critic", "critic2", "critic_target", "critic2_target"]
_is_restorable = ["actor", "critic_target", "critic2_target"]
def __init__(
self,
actor: Union[NeuralNetworkModule, nn.Module],
critic: Union[NeuralNetworkModule, nn.Module],
critic_target: Union[NeuralNetworkModule, nn.Module],
critic2: Union[NeuralNetworkModule, nn.Module],
critic2_target: Union[NeuralNetworkModule, nn.Module],
optimizer: Callable,
criterion: Callable,
*_,
lr_scheduler: Callable = None,
lr_scheduler_args: Tuple[Tuple, Tuple, Tuple] = None,
lr_scheduler_kwargs: Tuple[Dict, Dict, Dict] = None,
target_entropy: float = None,
initial_entropy_alpha: float = 1.0,
batch_size: int = 100,
update_rate: float = 0.005,
update_steps: Union[int, None] = None,
actor_learning_rate: float = 0.0005,
critic_learning_rate: float = 0.001,
alpha_learning_rate: float = 0.001,
discount: float = 0.99,
gradient_max: float = np.inf,
replay_size: int = 500000,
replay_device: Union[str, t.device] = "cpu",
replay_buffer: Buffer = None,
visualize: bool = False,
visualize_dir: str = "",
**__
):
"""
See Also:
:class:`.A2C`
:class:`.DDPG`
Important:
When given a state, and an optional action, actor must
at least return two values, similar to the actor structure
described in :class:`.A2C`. However, when actor is asked to
select an action based on the current state, you must make
sure that the sampling process is **differentiable**. E.g.
use the ``rsample`` method of torch distributions instead
of the ``sample`` method.
Compared to other actor-critic methods, SAC embeds the
entropy term into its reward function directly, rather than adding
the entropy term to actor's loss function. Therefore, we do not use
the entropy output of your actor network.
The SAC algorithm uses Q network as critics, so please reference
:class:`.DDPG` for the requirements and the definition of
``action_trans_func``.
Args:
actor: Actor network module.
critic: Critic network module.
critic_target: Target critic network module.
critic2: The second critic network module.
critic2_target: The second target critic network module.
optimizer: Optimizer used to optimize ``actor``, ``critic`` and
``critic2``.
criterion: Criterion used to evaluate the value loss.
*_:
lr_scheduler: Learning rate scheduler of ``optimizer``.
lr_scheduler_args: Arguments of the learning rate scheduler.
lr_scheduler_kwargs: Keyword arguments of the learning
rate scheduler.
target_entropy: Target entropy weight :math:`\\alpha` used in
the SAC soft value function:
:math:`V_{soft}(s_t) = \\mathbb{E}_{q_t\\sim\\pi}[\
Q_{soft}(s_t,a_t) - \
\\alpha log\\pi(a_t|s_t)]`
initial_entropy_alpha: Initial entropy weight :math:`\\alpha`
gradient_max: Maximum gradient.
batch_size: Batch size used during training.
update_rate: :math:`\\tau` used to update target networks.
Target parameters are updated as:
:math:`\\theta_t = \\theta * \\tau + \\theta_t * (1 - \\tau)`
update_steps: Training step number used to update target networks.
actor_learning_rate: Learning rate of the actor optimizer,
not compatible with ``lr_scheduler``.
critic_learning_rate: Learning rate of the critic optimizer,
not compatible with ``lr_scheduler``.
discount: :math:`\\gamma` used in the bellman function.
replay_size: Replay buffer size. Not compatible with
``replay_buffer``.
replay_device: Device where the replay buffer locates on, Not
compatible with ``replay_buffer``.
replay_buffer: Custom replay buffer.
visualize: Whether visualize the network flow in the first pass.
visualize_dir: Visualized graph save directory.
"""
self.batch_size = batch_size
self.update_rate = update_rate
self.update_steps = update_steps
self.discount = discount
self.visualize = visualize
self.visualize_dir = visualize_dir
self.entropy_alpha = t.tensor([initial_entropy_alpha], requires_grad=True)
self.grad_max = gradient_max
self.target_entropy = target_entropy
self._update_counter = 0
if update_rate is not None and update_steps is not None:
raise ValueError(
"You can only specify one target network update"
" scheme, either by update_rate or update_steps,"
" but not both."
)
self.actor = actor
self.critic = critic
self.critic_target = critic_target
self.critic2 = critic2
self.critic2_target = critic2_target
self.actor_optim = optimizer(self.actor.parameters(), lr=actor_learning_rate)
self.critic_optim = optimizer(self.critic.parameters(), lr=critic_learning_rate)
self.critic2_optim = optimizer(
self.critic2.parameters(), lr=critic_learning_rate
)
self.alpha_optim = optimizer([self.entropy_alpha], lr=alpha_learning_rate)
self.replay_buffer = (
Buffer(replay_size, replay_device)
if replay_buffer is None
else replay_buffer
)
# Make sure target and online networks have the same weight
with t.no_grad():
hard_update(self.critic, self.critic_target)
hard_update(self.critic2, self.critic2_target)
if lr_scheduler is not None:
if lr_scheduler_args is None:
lr_scheduler_args = ((), (), ())
if lr_scheduler_kwargs is None:
lr_scheduler_kwargs = ({}, {}, {})
self.actor_lr_sch = lr_scheduler(
self.actor_optim, *lr_scheduler_args[0], **lr_scheduler_kwargs[0],
)
self.critic_lr_sch = lr_scheduler(
self.critic_optim, *lr_scheduler_args[1], **lr_scheduler_kwargs[1]
)
self.critic2_lr_sch = lr_scheduler(
self.critic2_optim, *lr_scheduler_args[1], **lr_scheduler_kwargs[1]
)
self.alpha_lr_sch = lr_scheduler(
self.alpha_optim, *lr_scheduler_args[2], **lr_scheduler_kwargs[2]
)
self.criterion = criterion
super().__init__()
@property
def optimizers(self):
return [
self.actor_optim,
self.critic_optim,
self.critic2_optim,
self.alpha_optim,
]
@optimizers.setter
def optimizers(self, optimizers):
(
self.actor_optim,
self.critic_optim,
self.critic2_optim,
self.alpha_optim,
) = optimizers
@property
def lr_schedulers(self):
if (
hasattr(self, "actor_lr_sch")
and hasattr(self, "critic_lr_sch")
and hasattr(self, "critic2_lr_sch")
and hasattr(self, "alpha_lr_sch")
):
return [
self.actor_lr_sch,
self.critic_lr_sch,
self.critic2_lr_sch,
self.alpha_lr_sch,
]
return []
[docs] def act(self, state: Dict[str, Any], **__):
"""
Use actor network to produce an action for the current state.
Returns:
Anything produced by actor.
"""
return safe_return(safe_call(self.actor, state))
def _criticize(
self,
state: Dict[str, Any],
action: Dict[str, Any],
use_target: bool = False,
**__
):
"""
Use critic network to evaluate current value.
Args:
state: Current state.
action: Current action.
use_target: Whether to use the target network.
Returns:
Q Value of shape ``[batch_size, 1]``.
"""
if use_target:
return safe_call(self.critic_target, state, action)[0]
else:
return safe_call(self.critic, state, action)[0]
def _criticize2(
self, state: Dict[str, Any], action: Dict[str, Any], use_target=False, **__
):
"""
Use the second critic network to evaluate current value.
Args:
state: Current state.
action: Current action.
use_target: Whether to use the target network.
Returns:
Q Value of shape ``[batch_size, 1]``.
"""
if use_target:
return safe_call(self.critic2_target, state, action)[0]
else:
return safe_call(self.critic2, state, action)[0]
[docs] def store_episode(self, episode: List[Union[Transition, Dict]]):
"""
Add a full episode of transition samples to the replay buffer.
"""
self.replay_buffer.store_episode(
episode,
required_attrs=("state", "action", "next_state", "reward", "terminal"),
)
[docs] def update(
self,
update_value=True,
update_policy=True,
update_target=True,
update_entropy_alpha=True,
concatenate_samples=True,
**__
):
"""
Update network weights by sampling from replay buffer.
Args:
update_value: Whether to update the Q network.
update_policy: Whether to update the actor network.
update_target: Whether to update targets.
update_entropy_alpha: Whether to update :math:`alpha` of entropy.
concatenate_samples: Whether to concatenate the samples.
Returns:
mean value of estimated policy value, value loss
"""
self.actor.train()
self.critic.train()
self.critic2.train()
(
batch_size,
(state, action, reward, next_state, terminal, others,),
) = self.replay_buffer.sample_batch(
self.batch_size,
concatenate_samples,
sample_method="random_unique",
sample_attrs=["state", "action", "reward", "next_state", "terminal", "*"],
)
# Update critic network first
with t.no_grad():
next_action, next_action_log_prob, *_ = self.act(next_state)
next_action = self.action_transform_function(
next_action, next_state, others
)
next_value = self._criticize(next_state, next_action, True)
next_value2 = self._criticize2(next_state, next_action, True)
next_value = t.min(next_value, next_value2)
next_value = next_value.view(
batch_size, -1
) - self.entropy_alpha.item() * next_action_log_prob.view(batch_size, -1)
y_i = self.reward_function(
reward, self.discount, next_value, terminal, others
)
cur_value = self._criticize(state, action)
cur_value2 = self._criticize2(state, action)
value_loss = self.criterion(cur_value, y_i.type_as(cur_value))
value_loss2 = self.criterion(cur_value2, y_i.type_as(cur_value))
if self.visualize:
self.visualize_model(value_loss, "critic", self.visualize_dir)
if update_value:
self.critic.zero_grad()
value_loss.backward()
nn.utils.clip_grad_norm_(self.critic.parameters(), self.grad_max)
self.critic_optim.step()
self.critic2.zero_grad()
value_loss2.backward()
nn.utils.clip_grad_norm_(self.critic2.parameters(), self.grad_max)
self.critic2_optim.step()
# Update actor network
cur_action, cur_action_log_prob, *_ = self.act(state)
cur_action_log_prob = cur_action_log_prob.view(batch_size, 1)
cur_action = self.action_transform_function(cur_action, state, others)
act_value = self._criticize(state, cur_action)
act_value2 = self._criticize2(state, cur_action)
act_value = t.min(act_value, act_value2)
act_policy_loss = (
self.entropy_alpha.item() * cur_action_log_prob - act_value
).mean()
if self.visualize:
self.visualize_model(act_policy_loss, "actor", self.visualize_dir)
if update_policy:
self.actor.zero_grad()
act_policy_loss.backward()
nn.utils.clip_grad_norm_(self.actor.parameters(), self.grad_max)
self.actor_optim.step()
# Update target networks
if update_target:
if self.update_rate is not None:
soft_update(self.critic_target, self.critic, self.update_rate)
soft_update(self.critic2_target, self.critic2, self.update_rate)
else:
self._update_counter += 1
if self._update_counter % self.update_steps == 0:
hard_update(self.critic_target, self.critic)
hard_update(self.critic2_target, self.critic2)
if update_entropy_alpha and self.target_entropy is not None:
alpha_loss = -(
t.log(self.entropy_alpha)
* (cur_action_log_prob + self.target_entropy).cpu().detach()
).mean()
self.alpha_optim.zero_grad()
alpha_loss.backward()
self.alpha_optim.step()
# prevent nan
with t.no_grad():
self.entropy_alpha.clamp_(1e-6, 1e6)
self.actor.eval()
self.critic.eval()
self.critic2.eval()
# use .item() to prevent memory leakage
return -act_policy_loss.item(), (value_loss.item() + value_loss2.item()) / 2
[docs] def update_lr_scheduler(self):
"""
Update learning rate schedulers.
"""
if hasattr(self, "actor_lr_sch"):
self.actor_lr_sch.step()
if hasattr(self, "critic_lr_sch"):
self.critic_lr_sch.step()
[docs] def load(self, model_dir, network_map=None, version=-1):
# DOC INHERITED
super().load(model_dir, network_map, version)
with t.no_grad():
hard_update(self.critic, self.critic_target)
hard_update(self.critic, self.critic2_target)
[docs] @staticmethod
def reward_function(reward, discount, next_value, terminal, _):
next_value = next_value.to(reward.device)
terminal = terminal.to(reward.device)
return reward + discount * ~terminal * next_value
[docs] @classmethod
def generate_config(cls, config: Union[Dict[str, Any], Config]):
default_values = {
"models": ["Actor", "Critic", "Critic", "Critic", "Critic"],
"model_args": ((), (), (), (), ()),
"model_kwargs": ({}, {}, {}, {}, {}),
"optimizer": "Adam",
"criterion": "MSELoss",
"criterion_args": (),
"criterion_kwargs": {},
"lr_scheduler": None,
"lr_scheduler_args": None,
"lr_scheduler_kwargs": None,
"target_entropy": None,
"initial_entropy_alpha": 1.0,
"batch_size": 100,
"update_rate": 0.001,
"update_steps": None,
"actor_learning_rate": 0.0005,
"critic_learning_rate": 0.001,
"alpha_learning_rate": 0.001,
"discount": 0.99,
"gradient_max": np.inf,
"replay_size": 500000,
"replay_device": "cpu",
"replay_buffer": None,
"visualize": False,
"visualize_dir": "",
}
config = deepcopy(config)
config["frame"] = "SAC"
if "frame_config" not in config:
config["frame_config"] = default_values
else:
config["frame_config"] = {**config["frame_config"], **default_values}
return config
[docs] @classmethod
def init_from_config(
cls,
config: Union[Dict[str, Any], Config],
model_device: Union[str, t.device] = "cpu",
):
f_config = config["frame_config"]
models = assert_and_get_valid_models(f_config["models"])
model_args = f_config["model_args"]
model_kwargs = f_config["model_kwargs"]
models = [
m(*arg, **kwarg).to(model_device)
for m, arg, kwarg in zip(models, model_args, model_kwargs)
]
optimizer = assert_and_get_valid_optimizer(f_config["optimizer"])
criterion = assert_and_get_valid_criterion(f_config["criterion"])(
*f_config["criterion_args"], **f_config["criterion_kwargs"]
)
lr_scheduler = f_config["lr_scheduler"] and assert_and_get_valid_lr_scheduler(
f_config["lr_scheduler"]
)
f_config["optimizer"] = optimizer
f_config["criterion"] = criterion
f_config["lr_scheduler"] = lr_scheduler
frame = cls(*models, **f_config)
return frame