OscervaME
/
stable-baselines3-contrib-sacd
2
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
stable-baselines3-contrib-sacd/sb3_contrib/tqc/tqc.py

488 lines
22 KiB
Python
Raw Blame History

from typing import Any, Callable, Dict, List, Optional, Tuple, Type, Union
import numpy as np
import torch as th
from stable_baselines3.common import logger
from stable_baselines3.common.noise import ActionNoise
from stable_baselines3.common.off_policy_algorithm import OffPolicyAlgorithm
from stable_baselines3.common.type_aliases import GymEnv, MaybeCallback
from stable_baselines3.common.utils import polyak_update
from tqdm import tqdm
from sb3_contrib.tqc.policies import TQCPolicy
class TQC(OffPolicyAlgorithm):
"""
Controlling Overestimation Bias with Truncated Mixture of Continuous Distributional Quantile Critics
Paper: https://arxiv.org/abs/2005.04269
:param policy: (TQCPolicy or str) The policy model to use (MlpPolicy, CnnPolicy, ...)
:param env: (GymEnv or str) The environment to learn from (if registered in Gym, can be str)
:param learning_rate: (float or callable) learning rate for adam optimizer,
the same learning rate will be used for all networks (Q-Values, Actor and Value function)
it can be a function of the current progress remaining (from 1 to 0)
:param buffer_size: (int) size of the replay buffer
:param learning_starts: (int) how many steps of the model to collect transitions for before learning starts
:param batch_size: (int) Minibatch size for each gradient update
:param tau: (float) the soft update coefficient ("Polyak update", between 0 and 1)
:param gamma: (float) the discount factor
:param train_freq: (int) Update the model every ``train_freq`` steps.
:param gradient_steps: (int) How many gradient update after each step
:param n_episodes_rollout: (int) Update the model every ``n_episodes_rollout`` episodes.
Note that this cannot be used at the same time as ``train_freq``
:param action_noise: (ActionNoise) the action noise type (None by default), this can help
for hard exploration problem. Cf common.noise for the different action noise type.
:param optimize_memory_usage: (bool) Enable a memory efficient variant of the replay buffer
at a cost of more complexity.
See https://github.com/DLR-RM/stable-baselines3/issues/37#issuecomment-637501195
:param ent_coef: (str or float) Entropy regularization coefficient. (Equivalent to
inverse of reward scale in the original SAC paper.) Controlling exploration/exploitation trade-off.
Set it to 'auto' to learn it automatically (and 'auto_0.1' for using 0.1 as initial value)
:param target_update_interval: (int) update the target network every ``target_network_update_freq``
gradient steps.
:param target_entropy: (str or float) target entropy when learning ``ent_coef`` (``ent_coef = 'auto'``)
:param use_sde: (bool) Whether to use generalized State Dependent Exploration (gSDE)
instead of action noise exploration (default: False)
:param sde_sample_freq: (int) Sample a new noise matrix every n steps when using gSDE
Default: -1 (only sample at the beginning of the rollout)
:param use_sde_at_warmup: (bool) Whether to use gSDE instead of uniform sampling
during the warm up phase (before learning starts)
:param create_eval_env: (bool) Whether to create a second environment that will be
used for evaluating the agent periodically. (Only available when passing string for the environment)
:param policy_kwargs: (dict) additional arguments to be passed to the policy on creation
:param verbose: (int) the verbosity level: 0 no output, 1 info, 2 debug
:param seed: (int) Seed for the pseudo random generators
:param device: (str or th.device) Device (cpu, cuda, ...) on which the code should be run.
Setting it to auto, the code will be run on the GPU if possible.
:param _init_setup_model: (bool) Whether or not to build the network at the creation of the instance
"""
def __init__(
self,
policy: Union[str, Type[TQCPolicy]],
env: Union[GymEnv, str],
learning_rate: Union[float, Callable] = 3e-4,
buffer_size: int = int(1e6),
learning_starts: int = 100,
batch_size: int = 256,
tau: float = 0.005,
gamma: float = 0.99,
train_freq: int = 1,
gradient_steps: int = 1,
n_episodes_rollout: int = -1,
action_noise: Optional[ActionNoise] = None,
optimize_memory_usage: bool = False,
ent_coef: Union[str, float] = "auto",
target_update_interval: int = 1,
target_entropy: Union[str, float] = "auto",
top_quantiles_to_drop_per_net: int = 2,
use_sde: bool = False,
sde_sample_freq: int = -1,
use_sde_at_warmup: bool = False,
tensorboard_log: Optional[str] = None,
create_eval_env: bool = False,
policy_kwargs: Dict[str, Any] = None,
verbose: int = 0,
seed: Optional[int] = None,
device: Union[th.device, str] = "auto",
_init_setup_model: bool = True,
):
super(TQC, self).__init__(
policy,
env,
TQCPolicy,
learning_rate,
buffer_size,
learning_starts,
batch_size,
tau,
gamma,
train_freq,
gradient_steps,
n_episodes_rollout,
action_noise,
policy_kwargs=policy_kwargs,
tensorboard_log=tensorboard_log,
verbose=verbose,
device=device,
create_eval_env=create_eval_env,
seed=seed,
use_sde=use_sde,
sde_sample_freq=sde_sample_freq,
use_sde_at_warmup=use_sde_at_warmup,
optimize_memory_usage=optimize_memory_usage,
)
self.target_entropy = target_entropy
self.log_ent_coef = None # type: Optional[th.Tensor]
# Entropy coefficient / Entropy temperature
# Inverse of the reward scale
self.ent_coef = ent_coef
self.target_update_interval = target_update_interval
self.ent_coef_optimizer = None
self.top_quantiles_to_drop_per_net = top_quantiles_to_drop_per_net
if _init_setup_model:
self._setup_model()
def _setup_model(self) -> None:
super(TQC, self)._setup_model()
self._create_aliases()
self.replay_buffer.actor = self.actor
self.replay_buffer.ent_coef = 0.0
# Target entropy is used when learning the entropy coefficient
if self.target_entropy == "auto":
# automatically set target entropy if needed
self.target_entropy = -np.prod(self.env.action_space.shape).astype(np.float32)
else:
# Force conversion
# this will also throw an error for unexpected string
self.target_entropy = float(self.target_entropy)
# The entropy coefficient or entropy can be learned automatically
# see Automating Entropy Adjustment for Maximum Entropy RL section
# of https://arxiv.org/abs/1812.05905
if isinstance(self.ent_coef, str) and self.ent_coef.startswith("auto"):
# Default initial value of ent_coef when learned
init_value = 1.0
if "_" in self.ent_coef:
init_value = float(self.ent_coef.split("_")[1])
assert init_value > 0.0, "The initial value of ent_coef must be greater than 0"
# Note: we optimize the log of the entropy coeff which is slightly different from the paper
# as discussed in https://github.com/rail-berkeley/softlearning/issues/37
self.log_ent_coef = th.log(th.ones(1, device=self.device) * init_value).requires_grad_(True)
self.ent_coef_optimizer = th.optim.Adam([self.log_ent_coef], lr=self.lr_schedule(1))
else:
# Force conversion to float
# this will throw an error if a malformed string (different from 'auto')
# is passed
self.ent_coef_tensor = th.tensor(float(self.ent_coef)).to(self.device)
def _create_aliases(self) -> None:
self.actor = self.policy.actor
self.critic = self.policy.critic
self.critic_target = self.policy.critic_target
@staticmethod
def quantile_huber_loss(quantiles: th.Tensor, samples: th.Tensor) -> th.Tensor:
# batch x nets x quantiles x samples
pairwise_delta = samples[:, None, None, :] - quantiles[:, :, :, None]
abs_pairwise_delta = th.abs(pairwise_delta)
huber_loss = th.where(abs_pairwise_delta > 1, abs_pairwise_delta - 0.5, pairwise_delta ** 2 * 0.5)
n_quantiles = quantiles.shape[2]
tau = th.arange(n_quantiles, device=quantiles.device).float() / n_quantiles + 1 / 2 / n_quantiles
loss = (th.abs(tau[None, None, :, None] - (pairwise_delta < 0).float()) * huber_loss).mean()
return loss
def train(self, gradient_steps: int, batch_size: int = 64) -> None:
# Update optimizers learning rate
optimizers = [self.actor.optimizer, self.critic.optimizer]
if self.ent_coef_optimizer is not None:
optimizers += [self.ent_coef_optimizer]
# Update learning rate according to lr schedule
self._update_learning_rate(optimizers)
ent_coef_losses, ent_coefs = [], []
actor_losses, critic_losses = [], []
for gradient_step in range(gradient_steps):
# Sample replay buffer
replay_data = self.replay_buffer.sample(batch_size, env=self._vec_normalize_env)
# We need to sample because `log_std` may have changed between two gradient steps
if self.use_sde:
self.actor.reset_noise()
# Action by the current actor for the sampled state
actions_pi, log_prob = self.actor.action_log_prob(replay_data.observations)
log_prob = log_prob.reshape(-1, 1)
ent_coef_loss = None
if self.ent_coef_optimizer is not None:
# Important: detach the variable from the graph
# so we don't change it with other losses
# see https://github.com/rail-berkeley/softlearning/issues/60
ent_coef = th.exp(self.log_ent_coef.detach())
ent_coef_loss = -(self.log_ent_coef * (log_prob + self.target_entropy).detach()).mean()
ent_coef_losses.append(ent_coef_loss.item())
else:
ent_coef = self.ent_coef_tensor
ent_coefs.append(ent_coef.item())
self.replay_buffer.ent_coef = ent_coef.item()
# Optimize entropy coefficient, also called
# entropy temperature or alpha in the paper
if ent_coef_loss is not None:
self.ent_coef_optimizer.zero_grad()
ent_coef_loss.backward()
self.ent_coef_optimizer.step()
with th.no_grad():
top_quantiles_to_drop = self.top_quantiles_to_drop_per_net * self.critic.n_critics
# Select action according to policy
next_actions, next_log_prob = self.actor.action_log_prob(replay_data.next_observations)
# Compute and cut quantiles at the next state
# batch x nets x quantiles
next_z = self.critic_target(replay_data.next_observations, next_actions)
sorted_z, _ = th.sort(next_z.reshape(batch_size, -1))
sorted_z_part = sorted_z[:, : self.critic.quantiles_total - top_quantiles_to_drop]
target_q = sorted_z_part - ent_coef * next_log_prob.reshape(-1, 1)
# td error + entropy term
q_backup = replay_data.rewards + (1 - replay_data.dones) * self.gamma * target_q
# Get current Q estimates
# using action from the replay buffer
current_z = self.critic(replay_data.observations, replay_data.actions)
# Compute critic loss
critic_loss = self.quantile_huber_loss(current_z, q_backup)
critic_losses.append(critic_loss.item())
# Optimize the critic
self.critic.optimizer.zero_grad()
critic_loss.backward()
self.critic.optimizer.step()
# Compute actor loss
qf_pi = self.critic(replay_data.observations, actions_pi).mean(2).mean(1, keepdim=True)
actor_loss = (ent_coef * log_prob - qf_pi).mean()
actor_losses.append(actor_loss.item())
# Optimize the actor
self.actor.optimizer.zero_grad()
actor_loss.backward()
self.actor.optimizer.step()
# Update target networks
if gradient_step % self.target_update_interval == 0:
polyak_update(self.critic.parameters(), self.critic_target.parameters(), self.tau)
self._n_updates += gradient_steps
logger.record("train/n_updates", self._n_updates, exclude="tensorboard")
logger.record("train/ent_coef", np.mean(ent_coefs))
logger.record("train/actor_loss", np.mean(actor_losses))
logger.record("train/critic_loss", np.mean(critic_losses))
if len(ent_coef_losses) > 0:
logger.record("train/ent_coef_loss", np.mean(ent_coef_losses))
def pretrain(
self,
gradient_steps: int,
batch_size: int = 64,
n_action_samples: int = -1,
target_update_interval: int = 1,
tau: float = 0.005,
strategy: str = "exp",
reduce: str = "mean",
exp_temperature: float = 1.0,
off_policy_update_freq: int = -1,
) -> None:
"""
Pretrain with Critic Regularized Regression (CRR)
Paper: https://arxiv.org/abs/2006.15134
"""
# Update optimizers learning rate
optimizers = [self.actor.optimizer, self.critic.optimizer]
if self.ent_coef_optimizer is not None:
optimizers += [self.ent_coef_optimizer]
# Update learning rate according to lr schedule
self._update_learning_rate(optimizers)
actor_losses, critic_losses = [], []
for gradient_step in tqdm(range(gradient_steps)):
if off_policy_update_freq > 0 and gradient_step % off_policy_update_freq == 0:
self.train(gradient_steps=1, batch_size=batch_size)
continue
# Sample replay buffer
replay_data = self.replay_buffer.sample(batch_size, env=self._vec_normalize_env)
# We need to sample because `log_std` may have changed between two gradient steps
if self.use_sde:
self.actor.reset_noise()
# Action by the current actor for the sampled state
_, log_prob = self.actor.action_log_prob(replay_data.observations)
log_prob = log_prob.reshape(-1, 1)
ent_coef_loss = None
if self.ent_coef_optimizer is not None:
# Important: detach the variable from the graph
# so we don't change it with other losses
# see https://github.com/rail-berkeley/softlearning/issues/60
ent_coef = th.exp(self.log_ent_coef.detach())
ent_coef_loss = -(self.log_ent_coef * (log_prob + self.target_entropy).detach()).mean()
else:
ent_coef = self.ent_coef_tensor
self.replay_buffer.ent_coef = ent_coef.item()
# Optimize entropy coefficient, also called
# entropy temperature or alpha in the paper
if ent_coef_loss is not None:
self.ent_coef_optimizer.zero_grad()
ent_coef_loss.backward()
self.ent_coef_optimizer.step()
with th.no_grad():
top_quantiles_to_drop = self.top_quantiles_to_drop_per_net * self.critic.n_critics
# Select action according to policy
next_actions, next_log_prob = self.actor.action_log_prob(replay_data.next_observations)
# Compute and cut quantiles at the next state
# batch x nets x quantiles
next_z = self.critic_target(replay_data.next_observations, next_actions)
sorted_z, _ = th.sort(next_z.reshape(batch_size, -1))
sorted_z_part = sorted_z[:, : self.critic.quantiles_total - top_quantiles_to_drop]
target_q = sorted_z_part - ent_coef * next_log_prob.reshape(-1, 1)
# td error + entropy term
q_backup = replay_data.rewards + (1 - replay_data.dones) * self.gamma * target_q
# Get current Q estimates
# using action from the replay buffer
current_z = self.critic(replay_data.observations, replay_data.actions)
# Compute critic loss
critic_loss = self.quantile_huber_loss(current_z, q_backup)
critic_losses.append(critic_loss.item())
# Optimize the critic
self.critic.optimizer.zero_grad()
critic_loss.backward()
self.critic.optimizer.step()
if strategy == "bc":
# Behavior cloning
weight = 1
else:
# Tensor version: TODO: check that the reshape works as expected
# cleaner but not faster on cpu for large batch size
# with th.no_grad():
# # Q-value for the action in the buffer
# qf_buffer = self.critic(replay_data.observations, replay_data.actions).mean(2).mean(1, keepdim=True)
# # Create tensor to avoid loop
# # Note: For SDE, we need to sample several matrices
# obs_ = replay_data.observations.repeat(n_action_samples, 1)
# if self.use_sde:
# self.actor.reset_noise(batch_size * n_action_samples)
# actions_pi, _ = self.actor.action_log_prob(obs_)
# qf_pi = self.critic(obs_, actions_pi.detach()).mean(2).mean(1, keepdim=True)
# # Agregate: reduce mean or reduce max
# if reduce == "max":
# _, qf_agg = qf_pi.reshape(n_action_samples, batch_size, 1).max(axis=0)
# else:
# qf_agg = qf_pi.reshape(n_action_samples, batch_size, 1).mean(axis=0)
with th.no_grad():
qf_buffer = self.critic(replay_data.observations, replay_data.actions).mean(2).mean(1, keepdim=True)
# Use the mean (as done in AWAC, cf rlkit)
if n_action_samples == -1:
actions_pi = self.actor.forward(replay_data.observations, deterministic=True)
qf_agg = self.critic(replay_data.observations, actions_pi).mean(2).mean(1, keepdim=True)
else:
qf_agg = None
for _ in range(n_action_samples):
if self.use_sde:
self.actor.reset_noise()
actions_pi, _ = self.actor.action_log_prob(replay_data.observations)
qf_pi = self.critic(replay_data.observations, actions_pi.detach()).mean(2).mean(1, keepdim=True)
if qf_agg is None:
if reduce == "max":
qf_agg = qf_pi
else:
qf_agg = qf_pi / n_action_samples
else:
if reduce == "max":
qf_agg = th.max(qf_pi, qf_agg)
else:
qf_agg += qf_pi / n_action_samples
advantage = qf_buffer - qf_agg
if strategy == "binary":
# binary advantage
weight = advantage > 0
else:
# exp advantage
exp_clip = 20.0
weight = th.clamp(th.exp(advantage / exp_temperature), 0.0, exp_clip)
# Log prob by the current actor for the sampled state and action
log_prob = self.actor.evaluate_actions(replay_data.observations, replay_data.actions)
log_prob = log_prob.reshape(-1, 1)
# weigthed regression loss (close to policy gradient loss)
actor_loss = (-log_prob * weight).mean()
# actor_loss = ((actions_pi - replay_data.actions * weight) ** 2).mean()
actor_losses.append(actor_loss.item())
# Optimize the actor
self.actor.optimizer.zero_grad()
actor_loss.backward()
self.actor.optimizer.step()
# Update target networks
if gradient_step % target_update_interval == 0:
polyak_update(self.critic.parameters(), self.critic_target.parameters(), tau)
if self.use_sde:
print(f"std={(self.actor.get_std()).mean().item()}")
def learn(
self,
total_timesteps: int,
callback: MaybeCallback = None,
log_interval: int = 4,
eval_env: Optional[GymEnv] = None,
eval_freq: int = -1,
n_eval_episodes: int = 5,
tb_log_name: str = "TQC",
eval_log_path: Optional[str] = None,
reset_num_timesteps: bool = True,
) -> OffPolicyAlgorithm:
return super(TQC, self).learn(
total_timesteps=total_timesteps,
callback=callback,
log_interval=log_interval,
eval_env=eval_env,
eval_freq=eval_freq,
n_eval_episodes=n_eval_episodes,
tb_log_name=tb_log_name,
eval_log_path=eval_log_path,
reset_num_timesteps=reset_num_timesteps,
)
def _excluded_save_params(self) -> List[str]:
"""
Returns the names of the parameters that should be excluded by default
when saving the model.
:return: (List[str]) List of parameters that should be excluded from save
"""
# Exclude aliases
return super(TQC, self)._excluded_save_params() + ["actor", "critic", "critic_target"]
def _get_torch_save_params(self) -> Tuple[List[str], List[str]]:
"""
cf base class
"""
state_dicts = ["policy", "actor.optimizer", "critic.optimizer"]
saved_pytorch_variables = ["log_ent_coef"]
if self.ent_coef_optimizer is not None:
state_dicts.append("ent_coef_optimizer")
else:
saved_pytorch_variables.append("ent_coef_tensor")
return state_dicts, saved_pytorch_variables
Reference in New Issue View Git Blame Copy Permalink