espnet2.speechlm.core_lm.abs_core_lm.SpeechLMInferenceOptions

source

class espnet2.speechlm.core_lm.abs_core_lm.SpeechLMInferenceOptions(device: str = 'cpu', search_algo: str = 'sampling', nbest: int = 1, sampling_temperature: float = 1.0, top_k: int = 20, maxlenratio: float = 0.0, minlenratio: float = 0.0, eos: int = 5, start: int = 1, masks: Tensor | None = None, nq: int | None = None)

Bases: object

Options for inference in Speech Language Models.

This class holds various parameters that control the behavior of the inference process for Speech Language Models. Users can adjust these parameters to fine-tune the model’s output according to their needs.

device

The device to run the model on. Default is “cpu”.

• Type: str

search_algo

The algorithm used for searching the next token. Default is “sampling”.

• Type: str

nbest

The number of best candidates to consider. Default is 1.

• Type: int

sampling_temperature

The temperature parameter for sampling. Higher values lead to more randomness. Default is 1.0.

• Type: float

top_k

The number of top candidates to sample from. Default is 20.

• Type: int

maxlenratio

The maximum length ratio for the generated output compared to the input. Default is 0.0.

• Type: float

minlenratio

The minimum length ratio for the generated output compared to the input. Default is 0.0.

• Type: float

eos

The end-of-sequence token ID. Default is 5.

• Type: int

start

The start token ID. Default is 1.

• Type: int

masks

Optional masks for the input sequences. Default is None.

• Type: torch.Tensor

nq

Number of queries for the model. Default is None.

• Type: int

Examples

>>> options = SpeechLMInferenceOptions(
...     device="cuda",
...     search_algo="beam_search",
...     nbest=5,
...     sampling_temperature=0.8
... )
>>> print(options.device)
"cuda"

device

*: str* *= 'cpu'*

eos

*: int* *= 5*

masks

*: Tensor* *= None*

maxlenratio

*: float* *= 0.0*

minlenratio

*: float* *= 0.0*

nbest

*: int* *= 1*

nq

*: int* *= None*

sampling_temperature

*: float* *= 1.0*

search_algo

*: str* *= 'sampling'*

start

*: int* *= 1*

top_k

*: int* *= 20*