Language Language English 简体中文

unilab.algos.torch.him_ppo

class unilab.algos.torch.him_ppo.HIMActorCritic

Bases: Module

Parameters:

• num_actor_obs (int)

• num_critic_obs (int)

• num_one_step_obs (int)

• num_actions (int)

• actor_hidden_dims (list[int] | tuple[int, ...])

• critic_hidden_dims (list[int] | tuple[int, ...])

• activation (str)

• init_noise_std (float)

• estimator (dict | None)

is_recurrent = False

__init__(num_actor_obs, num_critic_obs, num_one_step_obs, num_actions, actor_hidden_dims=(512, 256, 128), critic_hidden_dims=(512, 256, 128), activation='elu', init_noise_std=1.0, estimator=None)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

Parameters:

• num_actor_obs (int)

• num_critic_obs (int)

• num_one_step_obs (int)

• num_actions (int)

• actor_hidden_dims (list[int] | tuple[int, ...])

• critic_hidden_dims (list[int] | tuple[int, ...])

• activation (str)

• init_noise_std (float)

• estimator (dict | None)

property action_mean: Tensor

property action_std: Tensor

property entropy: Tensor

reset(dones=None)

Parameters:

dones (Tensor | None)

Return type:

None

forward()

Define the computation performed at every call.

Should be overridden by all subclasses. :rtype: Tensor

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

update_distribution(obs_history)

Parameters:

obs_history (Tensor)

Return type:

None

act(obs_history, **kwargs)

Parameters:

obs_history (Tensor)

Return type:

Tensor

get_actions_log_prob(actions)

Parameters:

actions (Tensor)

Return type:

Tensor

act_inference(obs_history, observations=None)

Parameters:

obs_history (Tensor)

Return type:

Tensor

evaluate(critic_observations, **kwargs)

Parameters:

critic_observations (Tensor)

Return type:

Tensor

class unilab.algos.torch.him_ppo.HIMPPO

Bases: object

Parameters:

• num_learning_epochs (int)

• num_mini_batches (int)

• clip_param (float)

• gamma (float)

• lam (float)

• value_loss_coef (float)

• entropy_coef (float)

• learning_rate (float)

• max_grad_norm (float)

• use_clipped_value_loss (bool)

• schedule (str)

• desired_kl (float | None)

• device (str)

• kwargs (Any)

__init__(actor_critic, num_learning_epochs=1, num_mini_batches=1, clip_param=0.2, gamma=0.998, lam=0.95, value_loss_coef=1.0, entropy_coef=0.0, learning_rate=0.001, max_grad_norm=1.0, use_clipped_value_loss=True, schedule='fixed', desired_kl=0.01, device='cpu', **kwargs)

Parameters:

• num_learning_epochs (int)

• num_mini_batches (int)

• clip_param (float)

• gamma (float)

• lam (float)

• value_loss_coef (float)

• entropy_coef (float)

• learning_rate (float)

• max_grad_norm (float)

• use_clipped_value_loss (bool)

• schedule (str)

• desired_kl (float | None)

• device (str)

• kwargs (Any)

actor_critic: HIMActorCritic

init_storage(num_envs, num_transitions_per_env, actor_obs_shape, critic_obs_shape, action_shape)

Parameters:

• num_envs (int)

• num_transitions_per_env (int)

Return type:

None

test_mode()

Return type:

None

train_mode()

Return type:

None

act(obs, critic_obs)

Parameters:

• obs (Tensor)

• critic_obs (Tensor)

Return type:

Tensor

process_env_step(next_obs, rewards, dones, extras)

Parameters:

• next_obs (TensorDict | Tensor)

• rewards (Tensor)

• dones (Tensor)

• extras (dict[str, Tensor | TensorDict])

Return type:

None

compute_returns(last_critic_obs)

Parameters:

last_critic_obs (Tensor)

Return type:

None

update()

Return type:

tuple[float, float, float, float]

class unilab.algos.torch.him_ppo.HIMEstimator

Bases: Module

Parameters:

• temporal_steps (int)

• num_one_step_obs (int)

• enc_hidden_dims (list[int] | tuple[int, ...])

• tar_hidden_dims (list[int] | tuple[int, ...])

• activation (str)

• learning_rate (float)

• max_grad_norm (float)

• num_prototype (int)

• temperature (float)

• velocity_target_start (int | None)

• target_obs_start (int)

__init__(temporal_steps, num_one_step_obs, enc_hidden_dims=(128, 64, 16), tar_hidden_dims=(128, 64), activation='elu', learning_rate=0.001, max_grad_norm=10.0, num_prototype=32, temperature=3.0, velocity_target_start=None, target_obs_start=3)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

Parameters:

• temporal_steps (int)

• num_one_step_obs (int)

• enc_hidden_dims (list[int] | tuple[int, ...])

• tar_hidden_dims (list[int] | tuple[int, ...])

• activation (str)

• learning_rate (float)

• max_grad_norm (float)

• num_prototype (int)

• temperature (float)

• velocity_target_start (int | None)

• target_obs_start (int)

get_latent(obs_history)

Parameters:

obs_history (Tensor)

Return type:

tuple[Tensor, Tensor]

forward(obs_history)

Define the computation performed at every call.

Should be overridden by all subclasses.

Note

Although the recipe for forward pass needs to be defined within this function, one should call the Module instance afterwards instead of this since the former takes care of running the registered hooks while the latter silently ignores them.

Parameters:

obs_history (Tensor)

Return type:

tuple[Tensor, Tensor]

encode(obs_history)

Parameters:

obs_history (Tensor)

Return type:

tuple[Tensor, Tensor]

update(obs_history, next_critic_obs, lr=None)

Parameters:

• obs_history (Tensor)

• next_critic_obs (Tensor)

• lr (float | None)

Return type:

tuple[float, float]

class unilab.algos.torch.him_ppo.HIMRolloutStorage

Bases: object

Parameters:

• num_envs (int)

• num_transitions_per_env (int)

• obs_shape (Sequence[int])

• privileged_obs_shape (Sequence[int | None])

• actions_shape (Sequence[int])

• device (str)

class Transition

Bases: object

__init__()

clear()

Return type:

None

__init__(num_envs, num_transitions_per_env, obs_shape, privileged_obs_shape, actions_shape, device='cpu')

Parameters:

• num_envs (int)

• num_transitions_per_env (int)

• obs_shape (Sequence[int])

• privileged_obs_shape (Sequence[int | None])

• actions_shape (Sequence[int])

• device (str)

add_transition(transition)

Parameters:

transition (Transition)

Return type:

None

add_transitions(transition)

Parameters:

transition (Transition)

Return type:

None

clear()

Return type:

None

compute_returns(last_values, gamma, lam)

Parameters:

• last_values (Tensor)

• gamma (float)

• lam (float)

Return type:

None

mini_batch_generator(num_mini_batches, num_epochs=8)

Parameters:

• num_mini_batches (int)

• num_epochs (int)

Modules

actor_critic
algorithm
estimator
runner
storage