reagent.samplers package

Submodules

reagent.samplers.frechet module

class reagent.samplers.frechet.FrechetSort(shape: float = 1.0, topk: Optional[int] = None, equiv_len: Optional[int] = None, log_scores: bool = False)

Bases: reagent.gym.types.Sampler

EPS = 1e-12

log_prob(scores: torch.Tensor, action: torch.Tensor, equiv_len_override: Optional[torch.Tensor] = None) → torch.Tensor

What is the probability of a given set of scores producing the given list of permutations only considering the top equiv_len ranks?

We may want to override the default equiv_len here when we know the having larger action space doesn’t matter. i.e. in Reels

sample_action(scores: torch.Tensor) → reagent.core.types.ActorOutput

Sample a ranking according to Frechet sort. Note that possible_actions_mask is ignored as the list of rankings scales exponentially with slate size and number of items and it can be difficult to enumerate them.

Module contents

class reagent.samplers.FrechetSort(shape: float = 1.0, topk: Optional[int] = None, equiv_len: Optional[int] = None, log_scores: bool = False)

Bases: reagent.gym.types.Sampler

EPS = 1e-12

log_prob(scores: torch.Tensor, action: torch.Tensor, equiv_len_override: Optional[torch.Tensor] = None) → torch.Tensor

What is the probability of a given set of scores producing the given list of permutations only considering the top equiv_len ranks?

We may want to override the default equiv_len here when we know the having larger action space doesn’t matter. i.e. in Reels

sample_action(scores: torch.Tensor) → reagent.core.types.ActorOutput

Sample a ranking according to Frechet sort. Note that possible_actions_mask is ignored as the list of rankings scales exponentially with slate size and number of items and it can be difficult to enumerate them.