Skip to content

samplers

[ allennlp.nn.samplers.samplers ]

MultinomialSampler#

@Sampler.register("multinomial")
class MultinomialSampler(Sampler):
 | def __init__(
 |     self,
 |     temperature: float = 1.0,
 |     filter_val: float = -float("inf")
 | ) -> None

Represents a sampler to choose values from a multinomial distribution.

Registered as a Sampler with name "multinomial".

__call__#

class MultinomialSampler(Sampler):
 | ...
 | def __call__(
 |     self,
 |     logits: torch.Tensor,
 |     num_samples: int = 1,
 |     with_replacement: bool = True
 | ) -> torch.Tensor

TopKSampler#

@Sampler.register("top-k")
class TopKSampler(Sampler):
 | def __init__(
 |     self,
 |     k: int = 1,
 |     temperature: float = 1.0,
 |     filter_val: float = -float("inf")
 | )

Represents a Sampler which redistributes the probability mass function among the top k choices then selects from that subset logits is a tensor of log-probabilities to be selected from. k is the number of highest-probability options that the returned choice will be selected from temperature modules the probabilitis of the selected tokens. A temperature below 1.0 produces a sharper probability distribution and a temperature above 1.0 produces a flatter probability distribution.

Registered as a Sampler with name "top-k".

__call__#

class TopKSampler(Sampler):
 | ...
 | def __call__(
 |     self,
 |     logits: torch.Tensor,
 |     num_samples: int = 1,
 |     with_replacement: bool = True
 | ) -> torch.Tensor

TopPSampler#

@Sampler.register("top-p")
class TopPSampler(Sampler):
 | def __init__(
 |     self,
 |     p: float = 0.9,
 |     temperature: float = 1.0,
 |     filter_val: float = -float("inf")
 | )

Represents a Sampler which redistributes the probability mass function among the top choices with a cumulative probability of at least p then selects from that subset p if minimum cumulative probability of highest-probability options that the returned choice will be selected from temperature modules the probabilitis of the selected tokens. A temperature below 1.0 produces a sharper probability distribution and a temperature above 1.0 produces a flatter probability distribution.

Registered as a Sampler with name "top-p".

__call__#

class TopPSampler(Sampler):
 | ...
 | def __call__(
 |     self,
 |     logits: torch.Tensor,
 |     num_samples: int = 1,
 |     with_replacement: bool = True
 | ) -> torch.Tensor

Performs top-p sampling on the given logits. logits is a tensor of log-probabilities to be selected from. Returns the