metastable-baselines2/metastable_baselines2/common/on_policy_algorithm.py

344 lines
15 KiB
Python
Raw Normal View History

2023-11-19 18:34:15 +01:00
import sys
import time
from typing import Any, Dict, List, Optional, Tuple, Type, TypeVar, Union
2023-08-21 16:43:41 +02:00
2023-11-19 18:34:15 +01:00
import numpy as np
import torch as th
from gymnasium import spaces
from stable_baselines3.common.base_class import BaseAlgorithm
from stable_baselines3.common.buffers import DictRolloutBuffer, RolloutBuffer
from .buffers import BetterDictRolloutBuffer, BetterRolloutBuffer
2023-11-19 18:34:15 +01:00
from stable_baselines3.common.callbacks import BaseCallback
from stable_baselines3.common.policies import ActorCriticPolicy
from stable_baselines3.common.type_aliases import GymEnv, MaybeCallback, Schedule
from stable_baselines3.common.utils import obs_as_tensor, safe_mean
from stable_baselines3.common.vec_env import VecEnv
from stable_baselines3.common.on_policy_algorithm import OnPolicyAlgorithm
SelfOnPolicyAlgorithm = TypeVar("SelfOnPolicyAlgorithm", bound="BetterOnPolicyAlgorithm")
2023-08-21 16:43:41 +02:00
class BetterOnPolicyAlgorithm(OnPolicyAlgorithm):
"""
The base for On-Policy algorithms (ex: A2C/PPO).
:param policy: The policy model to use (MlpPolicy, CnnPolicy, ...)
:param env: The environment to learn from (if registered in Gym, can be str)
:param learning_rate: The learning rate, it can be a function
of the current progress remaining (from 1 to 0)
:param n_steps: The number of steps to run for each environment per update
(i.e. batch size is n_steps * n_env where n_env is number of environment copies running in parallel)
:param gamma: Discount factor
:param gae_lambda: Factor for trade-off of bias vs variance for Generalized Advantage Estimator.
Equivalent to classic advantage when set to 1.
:param ent_coef: Entropy coefficient for the loss calculation
:param vf_coef: Value function coefficient for the loss calculation
:param max_grad_norm: The maximum value for the gradient clipping
:param use_sde: Whether to use generalized State Dependent Exploration (gSDE)
instead of action noise exploration (default: False)
:param sde_sample_freq: Sample a new noise matrix every n steps when using gSDE
Default: -1 (only sample at the beginning of the rollout)
2023-11-19 18:34:15 +01:00
:param use_pca: Whether to use PCA.
:param rollout_buffer_class: Rollout buffer class to use. If ``None``, it will be automatically selected.
:param rollout_buffer_kwargs: Keyword arguments to pass to the rollout buffer on creation.
2023-08-21 16:43:41 +02:00
:param stats_window_size: Window size for the rollout logging, specifying the number of episodes to average
the reported success rate, mean episode length, and mean reward over
:param tensorboard_log: the log location for tensorboard (if None, no logging)
:param monitor_wrapper: When creating an environment, whether to wrap it
or not in a Monitor wrapper.
:param policy_kwargs: additional arguments to be passed to the policy on creation
:param verbose: Verbosity level: 0 for no output, 1 for info messages (such as device or wrappers used), 2 for
debug messages
:param seed: Seed for the pseudo random generators
:param 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: Whether or not to build the network at the creation of the instance
:param supported_action_spaces: The action spaces supported by the algorithm.
"""
def __init__(
self,
policy: Union[str, Type[ActorCriticPolicy]],
env: Union[GymEnv, str],
learning_rate: Union[float, Schedule],
n_steps: int,
gamma: float,
gae_lambda: float,
ent_coef: float,
vf_coef: float,
max_grad_norm: float,
use_sde: bool,
sde_sample_freq: int,
use_pca: bool,
pca_is: bool,
2023-11-19 18:34:15 +01:00
rollout_buffer_class: Optional[Type[RolloutBuffer]] = None,
rollout_buffer_kwargs: Optional[Dict[str, Any]] = None,
2023-08-21 16:43:41 +02:00
stats_window_size: int = 100,
tensorboard_log: Optional[str] = None,
monitor_wrapper: bool = True,
policy_kwargs: Optional[Dict[str, Any]] = None,
verbose: int = 0,
seed: Optional[int] = None,
device: Union[th.device, str] = "auto",
_init_setup_model: bool = True,
supported_action_spaces: Optional[Tuple[spaces.Space, ...]] = None,
):
assert not (use_sde and use_pca)
self.use_pca = use_pca
2023-11-19 18:34:15 +01:00
assert not pca_is or use_pca
self.pca_is = pca_is
2023-11-19 18:34:15 +01:00
assert not rollout_buffer_class and not rollout_buffer_kwargs
2023-08-21 16:43:41 +02:00
super().__init__(
policy=policy,
env=env,
learning_rate=learning_rate,
n_steps=n_steps,
gamma=gamma,
gae_lambda=gae_lambda,
ent_coef=ent_coef,
vf_coef=vf_coef,
max_grad_norm=max_grad_norm,
policy_kwargs=policy_kwargs,
verbose=verbose,
device=device,
use_sde=use_sde,
sde_sample_freq=sde_sample_freq,
2023-11-19 18:34:15 +01:00
#rollout_buffer_class = rollout_buffer_class,
#rollout_buffer_kwargs = rollout_buffer_kwargs,
2023-08-21 16:43:41 +02:00
# support_multi_env=True,
seed=seed,
stats_window_size=stats_window_size,
tensorboard_log=tensorboard_log,
supported_action_spaces=supported_action_spaces,
monitor_wrapper=monitor_wrapper,
_init_setup_model=_init_setup_model
)
if 'dist_kwargs' not in self.policy_kwargs:
self.policy_kwargs['dist_kwargs'] = {}
self.policy_kwargs['dist_kwargs']['n_envs'] = len(self.env.envs)
2023-11-19 18:34:15 +01:00
self.rollout_buffer_class = None
self.rollout_buffer_kwargs = {}
2023-08-21 16:43:41 +02:00
def _setup_model(self) -> None:
self._setup_lr_schedule()
self.set_random_seed(self.seed)
2023-11-19 18:34:15 +01:00
if self.rollout_buffer_class is None:
if isinstance(self.observation_space, spaces.Dict):
self.rollout_buffer_class = BetterDictRolloutBuffer
2023-11-19 18:34:15 +01:00
else:
self.rollout_buffer_class = BetterRolloutBuffer
2023-08-21 16:43:41 +02:00
2023-11-19 18:34:15 +01:00
self.rollout_buffer = self.rollout_buffer_class(
2023-08-21 16:43:41 +02:00
self.n_steps,
self.observation_space,
self.action_space,
device=self.device,
gamma=self.gamma,
gae_lambda=self.gae_lambda,
n_envs=self.n_envs,
2023-11-19 18:34:15 +01:00
**self.rollout_buffer_kwargs,
2023-08-21 16:43:41 +02:00
)
self.policy = self.policy_class( # pytype:disable=not-instantiable
self.observation_space,
self.action_space,
self.lr_schedule,
use_sde=self.use_sde,
use_pca=self.use_pca,
**self.policy_kwargs # pytype:disable=not-instantiable
)
self.policy = self.policy.to(self.device)
def collect_rollouts(
self,
env: VecEnv,
callback: BaseCallback,
rollout_buffer: RolloutBuffer,
n_rollout_steps: int,
) -> bool:
"""
Collect experiences using the current policy and fill a ``RolloutBuffer``.
The term rollout here refers to the model-free notion and should not
be used with the concept of rollout used in model-based RL or planning.
:param env: The training environment
:param callback: Callback that will be called at each step
(and at the beginning and end of the rollout)
:param rollout_buffer: Buffer to fill with rollouts
:param n_rollout_steps: Number of experiences to collect per environment
:return: True if function returned with at least `n_rollout_steps`
collected, False if callback terminated rollout prematurely.
"""
assert self._last_obs is not None, "No previous observation was provided"
# Switch to eval mode (this affects batch norm / dropout)
self.policy.set_training_mode(False)
n_steps = 0
rollout_buffer.reset()
# Sample new weights for the state dependent exploration
if self.use_sde:
self.policy.reset_noise(env.num_envs)
callback.on_rollout_start()
while n_steps < n_rollout_steps:
if self.use_sde and self.sde_sample_freq > 0 and n_steps % self.sde_sample_freq == 0:
# Sample a new noise matrix
self.policy.reset_noise(env.num_envs)
with th.no_grad():
# Convert to pytorch tensor or to TensorDict
obs_tensor = obs_as_tensor(self._last_obs, self.device)
actions, values, log_probs, distributions = self.policy(obs_tensor, conditioned_log_probs=self.pca_is)
2023-08-21 16:43:41 +02:00
actions = actions.cpu().numpy()
# Rescale and perform action
clipped_actions = actions
# Clip the actions to avoid out of bound error
if isinstance(self.action_space, spaces.Box):
2023-11-19 18:34:15 +01:00
if self.policy.squash_output:
# Unscale the actions to match env bounds
# if they were previously squashed (scaled in [-1, 1])
clipped_actions = self.policy.unscale_action(clipped_actions)
else:
# Otherwise, clip the actions to avoid out of bound error
# as we are sampling from an unbounded Gaussian distribution
clipped_actions = np.clip(actions, self.action_space.low, self.action_space.high)
2023-08-21 16:43:41 +02:00
new_obs, rewards, dones, infos = env.step(clipped_actions)
self.num_timesteps += env.num_envs
# Give access to local variables
callback.update_locals(locals())
2023-11-19 18:34:15 +01:00
if not callback.on_step():
2023-08-21 16:43:41 +02:00
return False
self._update_info_buffer(infos)
n_steps += 1
if isinstance(self.action_space, spaces.Discrete):
# Reshape in case of discrete action
actions = actions.reshape(-1, 1)
# Handle timeout by bootstraping with value function
# see GitHub issue #633
for idx, done in enumerate(dones):
if (
done
and infos[idx].get("terminal_observation") is not None
and infos[idx].get("TimeLimit.truncated", False)
):
terminal_obs = self.policy.obs_to_tensor(infos[idx]["terminal_observation"])[0]
with th.no_grad():
terminal_value = self.policy.predict_values(terminal_obs)[0]
rewards[idx] += self.gamma * terminal_value
2024-01-22 19:58:08 +01:00
rollout_buffer.add(self._last_obs, actions, rewards, self._last_episode_starts, values, log_probs, distributions.distribution.mean, distributions.distribution.scale)
2023-08-21 16:43:41 +02:00
self._last_obs = new_obs
self._last_episode_starts = dones
with th.no_grad():
# Compute value for the last timestep
values = self.policy.predict_values(obs_as_tensor(new_obs, self.device))
rollout_buffer.compute_returns_and_advantage(last_values=values, dones=dones)
2023-11-19 18:34:15 +01:00
callback.update_locals(locals())
2023-08-21 16:43:41 +02:00
callback.on_rollout_end()
return True
def train(self) -> None:
"""
Consume current rollout data and update policy parameters.
Implemented by individual algorithms.
"""
raise NotImplementedError
def predict(
self,
observation: Union[np.ndarray, Dict[str, np.ndarray]],
state: Optional[Tuple[np.ndarray, ...]] = None,
episode_start: Optional[np.ndarray] = None,
deterministic: bool = False,
trajectory: th.Tensor = None,
) -> Tuple[np.ndarray, Optional[Tuple[np.ndarray, ...]]]:
"""
Get the policy action from an observation (and optional hidden state).
Includes sugar-coating to handle different observations (e.g. normalizing images).
:param observation: the input observation
:param state: The last hidden states (can be None, used in recurrent policies)
:param episode_start: The last masks (can be None, used in recurrent policies)
this correspond to beginning of episodes,
where the hidden states of the RNN must be reset.
:param deterministic: Whether or not to return deterministic actions.
:param trajectory: Past trajectory. Only required when using PCA.
:return: the model's action and the next hidden state
(used in recurrent policies)
"""
return self.policy.predict(observation, state, episode_start, deterministic, trajectory=trajectory)
2023-08-21 16:43:41 +02:00
def learn(
self: SelfOnPolicyAlgorithm,
total_timesteps: int,
callback: MaybeCallback = None,
log_interval: int = 1,
tb_log_name: str = "OnPolicyAlgorithm",
reset_num_timesteps: bool = True,
progress_bar: bool = False,
) -> SelfOnPolicyAlgorithm:
iteration = 0
total_timesteps, callback = self._setup_learn(
total_timesteps,
callback,
reset_num_timesteps,
tb_log_name,
progress_bar,
)
callback.on_training_start(locals(), globals())
while self.num_timesteps < total_timesteps:
continue_training = self.collect_rollouts(self.env, callback, self.rollout_buffer, n_rollout_steps=self.n_steps)
2023-11-19 18:34:15 +01:00
if not continue_training:
2023-08-21 16:43:41 +02:00
break
iteration += 1
self._update_current_progress_remaining(self.num_timesteps, total_timesteps)
# Display training infos
if log_interval is not None and iteration % log_interval == 0:
2023-11-19 18:34:15 +01:00
assert self.ep_info_buffer is not None
2023-08-21 16:43:41 +02:00
time_elapsed = max((time.time_ns() - self.start_time) / 1e9, sys.float_info.epsilon)
fps = int((self.num_timesteps - self._num_timesteps_at_start) / time_elapsed)
self.logger.record("time/iterations", iteration, exclude="tensorboard")
if len(self.ep_info_buffer) > 0 and len(self.ep_info_buffer[0]) > 0:
self.logger.record("rollout/ep_rew_mean", safe_mean([ep_info["r"] for ep_info in self.ep_info_buffer]))
self.logger.record("rollout/ep_len_mean", safe_mean([ep_info["l"] for ep_info in self.ep_info_buffer]))
self.logger.record("time/fps", fps)
self.logger.record("time/time_elapsed", int(time_elapsed), exclude="tensorboard")
self.logger.record("time/total_timesteps", self.num_timesteps, exclude="tensorboard")
self.logger.dump(step=self.num_timesteps)
self.train()
callback.on_training_end()
return self
def _get_torch_save_params(self) -> Tuple[List[str], List[str]]:
state_dicts = ["policy", "policy.optimizer"]
return state_dicts, []