aepo.mbr.policy package

Submodules

aepo.mbr.policy.diverse_mbr module

aepo.mbr.policy.diverse_mbr.compute_dmbr(hyp: list | None = None, score_function: callable | None = None, matrix: array | None = None, weights: list | None = None, src: str | None = None, k: int = 1, div_pen: float = 0.0) array
Parameters:
  • hyp (list) – the list of hypotheses.

  • score_function (callable) – the scoring function.

  • matrix (np.array) – the similarity matrix.

  • weights (list) – the weights for the scoring function.

  • src (str) – the source.

  • k (int) – the number of elements to select.

  • div_pen (float) – the diversity penalty.

Compute the diverse MBR.

aepo.mbr.policy.diverse_mbr.gbfs(func: callable, n: int, k: int) array
Parameters:
  • func (callable) – the objective function.

  • n (int) – the number of elements.

  • k (int) – the number of elements to select.

Returns:

the selected elements.

Return type:

np.array

Greedy Best First Search for the diverse MBR. Greedy search is guaranteed to find a solution with an approximation factor of (1−1/e), provided that λ is small enough to ensure the function is non-decreasing; otherwise, the approximation factor is slightly worse than (1−1/e). Nemhauser, G. L., Wolsey, L. A., & Fisher, M. L. (1978). An analysis of approximations for maximizing submodular set functions—I. Mathematical programming, 14, 265-294.

aepo.mbr.policy.diverse_mbr.generate_objective(k: int, div_pen: float, matrix: array) callable
Parameters:
  • k (int) – the number of elements to select.

  • div_pen (float) – the diversity penalty.

  • matrix (np.array) – the similarity matrix.

Returns:

the objective function of the diverse MBR.

Return type:

callable

Generate an objective function for the diverse MBR. See https://github.com/CyberAgentAILab/diverse-mbr/tree/master for more details of diverse MBR.

Parameters:
  • func (callable) – the objective function.

  • init (np.array) – the initial solution.

  • iterations (int) – the number of iterations.

  • neighbor (int) – the number of neighbors to remove.

Returns:

the selected elements.

Return type:

np.array

Local search for the diverse MBR. This is an alternative to the greedy search.

aepo.mbr.policy.mbr module

aepo.mbr.policy.mbr.compute_mbr(hyp: list | None = None, compute_similarity: callable | None = None, matrix: array | None = None, weights: list | None = None, src: str | None = None, incremental: bool = False)
Parameters:
  • hyp (list) – the list of hypotheses.

  • compute_similarity (callable) – the similarity function.

  • matrix (np.array) – the similarity matrix.

  • weights (list) – the weights for the scoring function.

  • src (str) – the source.

  • incremental (bool) – the flag for incremental MBR.

Returns:

the best hypothesis index.

Return type:

int

Compute the minimum bayes risk decoding.

aepo.mbr.policy.mbr.compute_score_matrix(samples: list, score_function: callable, src_input: str | None = None) array
Parameters:
  • samples (list) – the list of samples.

  • score_function (callable) – the score function.

  • src_input (str) – the source input.

Returns:

the score matrix.

Return type:

np.array

Compute the score matrix for a list of samples.

Module contents