Recurrent PPO

Implementation of recurrent policies for the Proximal Policy Optimization (PPO) algorithm. Other than adding support for recurrent policies (LSTM here), the behavior is the same as in SB3’s core PPO algorithm.

Available Policies

MlpLstmPolicy

alias of RecurrentActorCriticPolicy

CnnLstmPolicy

alias of RecurrentActorCriticCnnPolicy

MultiInputLstmPolicy

alias of RecurrentMultiInputActorCriticPolicy

Notes

Can I use?

  • Recurrent policies: ✔️

  • Multi processing: ✔️

  • Gym spaces:

Space

Action

Observation

Discrete

✔️

✔️

Box

✔️

✔️

MultiDiscrete

✔️

✔️

MultiBinary

✔️

✔️

Dict

❌

✔️

Example

Note

It is particularly important to pass the lstm_states and episode_start argument to the predict() method, so the cell and hidden states of the LSTM are correctly updated.

import numpy as np

from sb3_contrib import RecurrentPPO
from stable_baselines3.common.evaluation import evaluate_policy

model = RecurrentPPO("MlpLstmPolicy", "CartPole-v1", verbose=1)
model.learn(5000)

vec_env = model.get_env()
mean_reward, std_reward = evaluate_policy(model, vec_env, n_eval_episodes=20, warn=False)
print(mean_reward)

model.save("ppo_recurrent")
del model # remove to demonstrate saving and loading

model = RecurrentPPO.load("ppo_recurrent")

obs = vec_env.reset()
# cell and hidden state of the LSTM
lstm_states = None
num_envs = 1
# Episode start signals are used to reset the lstm states
episode_starts = np.ones((num_envs,), dtype=bool)
while True:
    action, lstm_states = model.predict(obs, state=lstm_states, episode_start=episode_starts, deterministic=True)
    obs, rewards, dones, info = vec_env.step(action)
    episode_starts = dones
    vec_env.render("human")

Results

Report on environments with masked velocity (with and without framestack) can be found here: https://wandb.ai/sb3/no-vel-envs/reports/PPO-vs-RecurrentPPO-aka-PPO-LSTM-on-environments-with-masked-velocity–VmlldzoxOTI4NjE4

RecurrentPPO was evaluated against PPO on:

  • PendulumNoVel-v1

  • LunarLanderNoVel-v2

  • CartPoleNoVel-v1

  • MountainCarContinuousNoVel-v0

  • CarRacing-v0

How to replicate the results?

Clone the repo for the experiment:

git clone https://github.com/DLR-RM/rl-baselines3-zoo
cd rl-baselines3-zoo
git checkout feat/recurrent-ppo

Run the benchmark (replace $ENV_ID by the envs mentioned above):

python train.py --algo ppo_lstm --env $ENV_ID --eval-episodes 10 --eval-freq 10000

Parameters

class sb3_contrib.ppo_recurrent.RecurrentPPO(policy, env, learning_rate=0.0003, n_steps=128, batch_size=128, n_epochs=10, gamma=0.99, gae_lambda=0.95, clip_range=0.2, clip_range_vf=None, normalize_advantage=True, ent_coef=0.0, vf_coef=0.5, max_grad_norm=0.5, use_sde=False, sde_sample_freq=-1, target_kl=None, stats_window_size=100, tensorboard_log=None, policy_kwargs=None, verbose=0, seed=None, device='auto', _init_setup_model=True)[source]

Proximal Policy Optimization algorithm (PPO) (clip version) with support for recurrent policies (LSTM).

Based on the original Stable Baselines 3 implementation.

Introduction to PPO: https://spinningup.openai.com/en/latest/algorithms/ppo.html

Parameters:

  • policy (Union[str, Type[RecurrentActorCriticPolicy]]) – The policy model to use (MlpPolicy, CnnPolicy, …)

  • env (Union[Env, VecEnv, str]) – The environment to learn from (if registered in Gym, can be str)

  • learning_rate (Union[float, Callable[[float], float]]) – The learning rate, it can be a function of the current progress remaining (from 1 to 0)

  • n_steps (int) – 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)

  • batch_size (Optional[int]) – Minibatch size

  • n_epochs (int) – Number of epoch when optimizing the surrogate loss

  • gamma (float) – Discount factor

  • gae_lambda (float) – Factor for trade-off of bias vs variance for Generalized Advantage Estimator

  • clip_range (Union[float, Callable[[float], float]]) – Clipping parameter, it can be a function of the current progress remaining (from 1 to 0).

  • clip_range_vf (Union[None, float, Callable[[float], float]]) – Clipping parameter for the value function, it can be a function of the current progress remaining (from 1 to 0). This is a parameter specific to the OpenAI implementation. If None is passed (default), no clipping will be done on the value function. IMPORTANT: this clipping depends on the reward scaling.

  • normalize_advantage (bool) – Whether to normalize or not the advantage

  • ent_coef (float) – Entropy coefficient for the loss calculation

  • vf_coef (float) – Value function coefficient for the loss calculation

  • max_grad_norm (float) – The maximum value for the gradient clipping

  • target_kl (Optional[float]) – Limit the KL divergence between updates, because the clipping is not enough to prevent large update see issue #213 (cf https://github.com/hill-a/stable-baselines/issues/213) By default, there is no limit on the kl div.

  • stats_window_size (int) – Window size for the rollout logging, specifying the number of episodes to average the reported success rate, mean episode length, and mean reward over

  • tensorboard_log (Optional[str]) – the log location for tensorboard (if None, no logging)

  • policy_kwargs (Optional[Dict[str, Any]]) – additional arguments to be passed to the policy on creation

  • verbose (int) – the verbosity level: 0 no output, 1 info, 2 debug

  • seed (Optional[int]) – Seed for the pseudo random generators

  • device (Union[device, str]) – Device (cpu, cuda, …) on which the code should be run. Setting it to auto, the code will be run on the GPU if possible.

  • _init_setup_model (bool) – Whether or not to build the network at the creation of the instance

collect_rollouts(env, callback, rollout_buffer, n_rollout_steps)[source]

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.

Parameters:

  • env (VecEnv) – The training environment

  • callback (BaseCallback) – Callback that will be called at each step (and at the beginning and end of the rollout)

  • rollout_buffer (RolloutBuffer) – Buffer to fill with rollouts

  • n_steps – Number of experiences to collect per environment

Return type:

bool

Returns:

True if function returned with at least n_rollout_steps collected, False if callback terminated rollout prematurely.

get_env()

Returns the current environment (can be None if not defined).

Return type:

Optional[VecEnv]

Returns:

The current environment

get_parameters()

Return the parameters of the agent. This includes parameters from different networks, e.g. critics (value functions) and policies (pi functions).

Return type:

Dict[str, Dict]

Returns:

Mapping of from names of the objects to PyTorch state-dicts.

get_vec_normalize_env()

Return the VecNormalize wrapper of the training env if it exists.

Return type:

Optional[VecNormalize]

Returns:

The VecNormalize env.

learn(total_timesteps, callback=None, log_interval=1, tb_log_name='RecurrentPPO', reset_num_timesteps=True, progress_bar=False)[source]

Return a trained model.

Parameters:

  • total_timesteps (int) – The total number of samples (env steps) to train on

  • callback (Union[None, Callable, List[BaseCallback], BaseCallback]) – callback(s) called at every step with state of the algorithm.

  • log_interval (int) – The number of episodes before logging.

  • tb_log_name (str) – the name of the run for TensorBoard logging

  • reset_num_timesteps (bool) – whether or not to reset the current timestep number (used in logging)

  • progress_bar (bool) – Display a progress bar using tqdm and rich.

Return type:

TypeVar(SelfRecurrentPPO, bound= RecurrentPPO)

Returns:

the trained model

classmethod load(path, env=None, device='auto', custom_objects=None, print_system_info=False, force_reset=True, **kwargs)

Load the model from a zip-file. Warning: load re-creates the model from scratch, it does not update it in-place! For an in-place load use set_parameters instead.

Parameters:

  • path (Union[str, Path, BufferedIOBase]) – path to the file (or a file-like) where to load the agent from

  • env (Union[Env, VecEnv, None]) – the new environment to run the loaded model on (can be None if you only need prediction from a trained model) has priority over any saved environment

  • device (Union[device, str]) – Device on which the code should run.

  • custom_objects (Optional[Dict[str, Any]]) – Dictionary of objects to replace upon loading. If a variable is present in this dictionary as a key, it will not be deserialized and the corresponding item will be used instead. Similar to custom_objects in keras.models.load_model. Useful when you have an object in file that can not be deserialized.

  • print_system_info (bool) – Whether to print system info from the saved model and the current system info (useful to debug loading issues)

  • force_reset (bool) – Force call to reset() before training to avoid unexpected behavior. See https://github.com/DLR-RM/stable-baselines3/issues/597

  • kwargs – extra arguments to change the model when loading

Return type:

TypeVar(SelfBaseAlgorithm, bound= BaseAlgorithm)

Returns:

new model instance with loaded parameters

property logger: Logger

Getter for the logger object.

predict(observation, state=None, episode_start=None, deterministic=False)

Get the policy action from an observation (and optional hidden state). Includes sugar-coating to handle different observations (e.g. normalizing images).

Parameters:

  • observation (Union[ndarray, Dict[str, ndarray]]) – the input observation

  • state (Optional[Tuple[ndarray, ...]]) – The last hidden states (can be None, used in recurrent policies)

  • episode_start (Optional[ndarray]) – 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.

  • deterministic (bool) – Whether or not to return deterministic actions.

Return type:

Tuple[ndarray, Optional[Tuple[ndarray, ...]]]

Returns:

the model’s action and the next hidden state (used in recurrent policies)

save(path, exclude=None, include=None)

Save all the attributes of the object and the model parameters in a zip-file.

Parameters:

  • path (Union[str, Path, BufferedIOBase]) – path to the file where the rl agent should be saved

  • exclude (Optional[Iterable[str]]) – name of parameters that should be excluded in addition to the default ones

  • include (Optional[Iterable[str]]) – name of parameters that might be excluded but should be included anyway

Return type:

None

set_env(env, force_reset=True)

Checks the validity of the environment, and if it is coherent, set it as the current environment. Furthermore wrap any non vectorized env into a vectorized checked parameters: - observation_space - action_space

Parameters:
Return type:

None

set_logger(logger)

Setter for for logger object. :rtype: None

Warning

When passing a custom logger object, this will overwrite tensorboard_log and verbose settings passed to the constructor.

set_parameters(load_path_or_dict, exact_match=True, device='auto')

Load parameters from a given zip-file or a nested dictionary containing parameters for different modules (see get_parameters).

Parameters:

  • load_path_or_iter – Location of the saved data (path or file-like, see save), or a nested dictionary containing nn.Module parameters used by the policy. The dictionary maps object names to a state-dictionary returned by torch.nn.Module.state_dict().

  • exact_match (bool) – If True, the given parameters should include parameters for each module and each of their parameters, otherwise raises an Exception. If set to False, this can be used to update only specific parameters.

  • device (Union[device, str]) – Device on which the code should run.

Return type:

None

set_random_seed(seed=None)

Set the seed of the pseudo-random generators (python, numpy, pytorch, gym, action_space)

Parameters:

seed (Optional[int]) –

Return type:

None

train()[source]

Update policy using the currently gathered rollout buffer.

Return type:

None

RecurrentPPO Policies

sb3_contrib.ppo_recurrent.MlpLstmPolicy

alias of RecurrentActorCriticPolicy

class sb3_contrib.common.recurrent.policies.RecurrentActorCriticPolicy(observation_space, action_space, lr_schedule, net_arch=None, activation_fn=<class 'torch.nn.modules.activation.Tanh'>, ortho_init=True, use_sde=False, log_std_init=0.0, full_std=True, use_expln=False, squash_output=False, features_extractor_class=<class 'stable_baselines3.common.torch_layers.FlattenExtractor'>, features_extractor_kwargs=None, share_features_extractor=True, normalize_images=True, optimizer_class=<class 'torch.optim.adam.Adam'>, optimizer_kwargs=None, lstm_hidden_size=256, n_lstm_layers=1, shared_lstm=False, enable_critic_lstm=True, lstm_kwargs=None)[source]

Recurrent policy class for actor-critic algorithms (has both policy and value prediction). To be used with A2C, PPO and the likes. It assumes that both the actor and the critic LSTM have the same architecture.

Parameters:

  • observation_space (Space) – Observation space

  • action_space (Space) – Action space

  • lr_schedule (Callable[[float], float]) – Learning rate schedule (could be constant)

  • net_arch (Union[List[int], Dict[str, List[int]], None]) – The specification of the policy and value networks.

  • activation_fn (Type[Module]) – Activation function

  • ortho_init (bool) – Whether to use or not orthogonal initialization

  • use_sde (bool) – Whether to use State Dependent Exploration or not

  • log_std_init (float) – Initial value for the log standard deviation

  • full_std (bool) – Whether to use (n_features x n_actions) parameters for the std instead of only (n_features,) when using gSDE

  • use_expln (bool) – Use expln() function instead of exp() to ensure a positive standard deviation (cf paper). It allows to keep variance above zero and prevent it from growing too fast. In practice, exp() is usually enough.

  • squash_output (bool) – Whether to squash the output using a tanh function, this allows to ensure boundaries when using gSDE.

  • features_extractor_class (Type[BaseFeaturesExtractor]) – Features extractor to use.

  • features_extractor_kwargs (Optional[Dict[str, Any]]) – Keyword arguments to pass to the features extractor.

  • share_features_extractor (bool) – If True, the features extractor is shared between the policy and value networks.

  • normalize_images (bool) – Whether to normalize images or not, dividing by 255.0 (True by default)

  • optimizer_class (Type[Optimizer]) – The optimizer to use, th.optim.Adam by default

  • optimizer_kwargs (Optional[Dict[str, Any]]) – Additional keyword arguments, excluding the learning rate, to pass to the optimizer

  • lstm_hidden_size (int) – Number of hidden units for each LSTM layer.

  • n_lstm_layers (int) – Number of LSTM layers.

  • shared_lstm (bool) – Whether the LSTM is shared between the actor and the critic (in that case, only the actor gradient is used) By default, the actor and the critic have two separate LSTM.

  • enable_critic_lstm (bool) – Use a seperate LSTM for the critic.

  • lstm_kwargs (Optional[Dict[str, Any]]) – Additional keyword arguments to pass the the LSTM constructor.

evaluate_actions(obs, actions, lstm_states, episode_starts)[source]

Evaluate actions according to the current policy, given the observations.

Parameters:

  • obs (Tensor) – Observation.

  • actions (Tensor) –

  • lstm_states (RNNStates) – The last hidden and memory states for the LSTM.

  • episode_starts (Tensor) – Whether the observations correspond to new episodes or not (we reset the lstm states in that case).

Return type:

Tuple[Tensor, Tensor, Tensor]

Returns:

estimated value, log likelihood of taking those actions and entropy of the action distribution.

forward(obs, lstm_states, episode_starts, deterministic=False)[source]

Forward pass in all the networks (actor and critic)

Parameters:

  • obs (Tensor) – Observation. Observation

  • lstm_states (RNNStates) – The last hidden and memory states for the LSTM.

  • episode_starts (Tensor) – Whether the observations correspond to new episodes or not (we reset the lstm states in that case).

  • deterministic (bool) – Whether to sample or use deterministic actions

Return type:

Tuple[Tensor, Tensor, Tensor, RNNStates]

Returns:

action, value and log probability of the action

get_distribution(obs, lstm_states, episode_starts)[source]

Get the current policy distribution given the observations.

Parameters:

  • obs (Tensor) – Observation.

  • lstm_states (Tuple[Tensor, Tensor]) – The last hidden and memory states for the LSTM.

  • episode_starts (Tensor) – Whether the observations correspond to new episodes or not (we reset the lstm states in that case).

Return type:

Tuple[Distribution, Tuple[Tensor, ...]]

Returns:

the action distribution and new hidden states.

predict(observation, state=None, episode_start=None, deterministic=False)[source]

Get the policy action from an observation (and optional hidden state). Includes sugar-coating to handle different observations (e.g. normalizing images).

Parameters:

  • observation (Union[ndarray, Dict[str, ndarray]]) – the input observation

  • lstm_states – The last hidden and memory states for the LSTM.

  • episode_starts – Whether the observations correspond to new episodes or not (we reset the lstm states in that case).

  • deterministic (bool) – Whether or not to return deterministic actions.

Return type:

Tuple[ndarray, Optional[Tuple[ndarray, ...]]]

Returns:

the model’s action and the next hidden state (used in recurrent policies)

predict_values(obs, lstm_states, episode_starts)[source]

Get the estimated values according to the current policy given the observations.

Parameters:

  • obs (Tensor) – Observation.

  • lstm_states (Tuple[Tensor, Tensor]) – The last hidden and memory states for the LSTM.

  • episode_starts (Tensor) – Whether the observations correspond to new episodes or not (we reset the lstm states in that case).

Return type:

Tensor

Returns:

the estimated values.

sb3_contrib.ppo_recurrent.CnnLstmPolicy

alias of RecurrentActorCriticCnnPolicy

class sb3_contrib.common.recurrent.policies.RecurrentActorCriticCnnPolicy(observation_space, action_space, lr_schedule, net_arch=None, activation_fn=<class 'torch.nn.modules.activation.Tanh'>, ortho_init=True, use_sde=False, log_std_init=0.0, full_std=True, use_expln=False, squash_output=False, features_extractor_class=<class 'stable_baselines3.common.torch_layers.NatureCNN'>, features_extractor_kwargs=None, share_features_extractor=True, normalize_images=True, optimizer_class=<class 'torch.optim.adam.Adam'>, optimizer_kwargs=None, lstm_hidden_size=256, n_lstm_layers=1, shared_lstm=False, enable_critic_lstm=True, lstm_kwargs=None)[source]

CNN recurrent policy class for actor-critic algorithms (has both policy and value prediction). Used by A2C, PPO and the likes.

Parameters:

  • observation_space (Space) – Observation space

  • action_space (Space) – Action space

  • lr_schedule (Callable[[float], float]) – Learning rate schedule (could be constant)

  • net_arch (Union[List[int], Dict[str, List[int]], None]) – The specification of the policy and value networks.

  • activation_fn (Type[Module]) – Activation function

  • ortho_init (bool) – Whether to use or not orthogonal initialization

  • use_sde (bool) – Whether to use State Dependent Exploration or not

  • log_std_init (float) – Initial value for the log standard deviation

  • full_std (bool) – Whether to use (n_features x n_actions) parameters for the std instead of only (n_features,) when using gSDE

  • use_expln (bool) – Use expln() function instead of exp() to ensure a positive standard deviation (cf paper). It allows to keep variance above zero and prevent it from growing too fast. In practice, exp() is usually enough.

  • squash_output (bool) – Whether to squash the output using a tanh function, this allows to ensure boundaries when using gSDE.

  • features_extractor_class (Type[BaseFeaturesExtractor]) – Features extractor to use.

  • features_extractor_kwargs (Optional[Dict[str, Any]]) – Keyword arguments to pass to the features extractor.

  • share_features_extractor (bool) – If True, the features extractor is shared between the policy and value networks.

  • normalize_images (bool) – Whether to normalize images or not, dividing by 255.0 (True by default)

  • optimizer_class (Type[Optimizer]) – The optimizer to use, th.optim.Adam by default

  • optimizer_kwargs (Optional[Dict[str, Any]]) – Additional keyword arguments, excluding the learning rate, to pass to the optimizer

  • lstm_hidden_size (int) – Number of hidden units for each LSTM layer.

  • n_lstm_layers (int) – Number of LSTM layers.

  • shared_lstm (bool) – Whether the LSTM is shared between the actor and the critic. By default, only the actor has a recurrent network.

  • enable_critic_lstm (bool) – Use a seperate LSTM for the critic.

  • lstm_kwargs (Optional[Dict[str, Any]]) – Additional keyword arguments to pass the the LSTM constructor.

sb3_contrib.ppo_recurrent.MultiInputLstmPolicy

alias of RecurrentMultiInputActorCriticPolicy

class sb3_contrib.common.recurrent.policies.RecurrentMultiInputActorCriticPolicy(observation_space, action_space, lr_schedule, net_arch=None, activation_fn=<class 'torch.nn.modules.activation.Tanh'>, ortho_init=True, use_sde=False, log_std_init=0.0, full_std=True, use_expln=False, squash_output=False, features_extractor_class=<class 'stable_baselines3.common.torch_layers.CombinedExtractor'>, features_extractor_kwargs=None, share_features_extractor=True, normalize_images=True, optimizer_class=<class 'torch.optim.adam.Adam'>, optimizer_kwargs=None, lstm_hidden_size=256, n_lstm_layers=1, shared_lstm=False, enable_critic_lstm=True, lstm_kwargs=None)[source]

MultiInputActorClass policy class for actor-critic algorithms (has both policy and value prediction). Used by A2C, PPO and the likes.

Parameters:

  • observation_space (Space) – Observation space

  • action_space (Space) – Action space

  • lr_schedule (Callable[[float], float]) – Learning rate schedule (could be constant)

  • net_arch (Union[List[int], Dict[str, List[int]], None]) – The specification of the policy and value networks.

  • activation_fn (Type[Module]) – Activation function

  • ortho_init (bool) – Whether to use or not orthogonal initialization

  • use_sde (bool) – Whether to use State Dependent Exploration or not

  • log_std_init (float) – Initial value for the log standard deviation

  • full_std (bool) – Whether to use (n_features x n_actions) parameters for the std instead of only (n_features,) when using gSDE

  • use_expln (bool) – Use expln() function instead of exp() to ensure a positive standard deviation (cf paper). It allows to keep variance above zero and prevent it from growing too fast. In practice, exp() is usually enough.

  • squash_output (bool) – Whether to squash the output using a tanh function, this allows to ensure boundaries when using gSDE.

  • features_extractor_class (Type[BaseFeaturesExtractor]) – Features extractor to use.

  • features_extractor_kwargs (Optional[Dict[str, Any]]) – Keyword arguments to pass to the features extractor.

  • share_features_extractor (bool) – If True, the features extractor is shared between the policy and value networks.

  • normalize_images (bool) – Whether to normalize images or not, dividing by 255.0 (True by default)

  • optimizer_class (Type[Optimizer]) – The optimizer to use, th.optim.Adam by default

  • optimizer_kwargs (Optional[Dict[str, Any]]) – Additional keyword arguments, excluding the learning rate, to pass to the optimizer

  • lstm_hidden_size (int) – Number of hidden units for each LSTM layer.

  • n_lstm_layers (int) – Number of LSTM layers.

  • shared_lstm (bool) – Whether the LSTM is shared between the actor and the critic. By default, only the actor has a recurrent network.

  • enable_critic_lstm (bool) – Use a seperate LSTM for the critic.

  • lstm_kwargs (Optional[Dict[str, Any]]) – Additional keyword arguments to pass the the LSTM constructor.