types

Source code: tianshou/data/types.py

---

class ActBatchProtocol(*args, **kwargs)

Simplest batch, just containing the action. Useful e.g., for random policy.

act: Tensor | ndarray

class ActStateBatchProtocol(*args, **kwargs)

Contains action and state (which can be None), useful for policies that can support RNNs.

state: dict | BatchProtocol | ndarray | None

class BatchWithAdvantagesProtocol(*args, **kwargs)

Contains estimated advantages and values.

Returns are usually computed from GAE of advantages by adding the value.

adv: Tensor

v_s: Tensor

class BatchWithReturnsProtocol(*args, **kwargs)

With added returns, usually computed with GAE.

returns: Tensor | ndarray

class DistBatchProtocol(*args, **kwargs)

Contains dist instances for actions (created by dist_fn).

Usually categorical or normal.

dist: Distribution

class DistLogProbBatchProtocol(*args, **kwargs)

Contains dist objects that can be sampled from and log_prob of taken action.

log_prob: Tensor

class FQFBatchProtocol(*args, **kwargs)

Model outputs, fractions and quantiles_tau - specific to the FQF model.

fractions: Tensor

quantiles_tau: Tensor

class ImitationBatchProtocol(*args, **kwargs)

Similar to other batches, but contains imitation_logits and q_value fields.

imitation_logits: Tensor

q_value: Tensor

state: dict | Batch | ndarray | None

class LogpOldProtocol(*args, **kwargs)

Contains logp_old, often needed for importance weights, in particular in PPO.

Builds on batches that contain advantages and values.

logp_old: Tensor

class ModelOutputBatchProtocol(*args, **kwargs)

In addition to state and action, contains model output: (logits).

logits: Tensor

class ObsBatchProtocol(*args, **kwargs)

Observations of an environment that a policy can turn into actions.

Typically used inside a policy’s forward

info: Tensor | ndarray

obs: Tensor | ndarray | BatchProtocol

class PrioBatchProtocol(*args, **kwargs)

Contains weights that can be used for prioritized replay.

weight: ndarray | Tensor

class QuantileRegressionBatchProtocol(*args, **kwargs)

Contains taus for algorithms using quantile regression.

See e.g. https://arxiv.org/abs/1806.06923

taus: Tensor

class RecurrentStateBatch(*args, **kwargs)

Used by RNNs in policies, contains hidden and cell fields.

cell: Tensor

hidden: Tensor

class RolloutBatchProtocol(*args, **kwargs)

Typically, the outcome of sampling from a replay buffer.

act: Tensor | ndarray

obs_next: Tensor | ndarray | BatchProtocol

rew: ndarray

terminated: Tensor | ndarray

truncated: Tensor | ndarray