import math
from typing import Any, Callable, Sequence, Tuple, Union
import torch
from ignite.exceptions import NotComputableError
from ignite.metrics.metric import Metric, reinit__is_reduced, sync_all_reduce
from ignite.metrics.nlp.utils import modified_precision
__all__ = ["Bleu"]
def _closest_ref_length(references: Sequence[Sequence[Any]], hyp_len: int) -> int:
ref_lens = (len(reference) for reference in references)
closest_ref_len = min(ref_lens, key=lambda ref_len: (abs(ref_len - hyp_len), ref_len))
return closest_ref_len
class _Smoother:
"""
Smoothing helper
http://acl2014.org/acl2014/W14-33/pdf/W14-3346.pdf
"""
def __init__(self, method: str):
valid = ["no_smooth", "smooth1", "nltk_smooth2", "smooth2"]
if method not in valid:
raise ValueError(f"Smooth is not valid (expected: {valid}, got: {method})")
self.smooth = method
def __call__(self, numerators: torch.Tensor, denominators: torch.Tensor) -> Sequence[float]:
method = getattr(self, self.smooth)
return method(numerators, denominators)
@staticmethod
def smooth1(numerators: torch.Tensor, denominators: torch.Tensor) -> Sequence[float]:
epsilon = 0.1
denominators_ = [max(1, d.item()) for d in denominators]
return [n.item() / d if n != 0 else epsilon / d for n, d in zip(numerators, denominators_)]
@staticmethod
def nltk_smooth2(numerators: torch.Tensor, denominators: torch.Tensor) -> Sequence[float]:
denominators_ = torch.tensor([max(1, d.item()) for d in denominators])
return _Smoother._smooth2(numerators, denominators_)
@staticmethod
def smooth2(numerators: torch.Tensor, denominators: torch.Tensor) -> Sequence[float]:
return _Smoother._smooth2(numerators, denominators)
@staticmethod
def _smooth2(numerators: torch.Tensor, denominators: torch.Tensor) -> Sequence[float]:
return [
(n.item() + 1) / (d.item() + 1) if i != 0 else n.item() / d.item()
for i, (n, d) in enumerate(zip(numerators, denominators))
]
@staticmethod
def no_smooth(numerators: torch.Tensor, denominators: torch.Tensor) -> Sequence[float]:
denominators_ = [max(1, d) for d in denominators]
return [n.item() / d for n, d in zip(numerators, denominators_)]
[docs]class Bleu(Metric):
r"""Calculates the `BLEU score <https://en.wikipedia.org/wiki/BLEU>`_.
.. math::
\text{BLEU} = b_{p} \cdot \exp \left( \sum_{n=1}^{N} w_{n} \: \log p_{n} \right)
where :math:`N` is the order of n-grams, :math:`b_{p}` is a sentence brevety penalty, :math:`w_{n}` are
positive weights summing to one and :math:`p_{n}` are modified n-gram precisions.
More details can be found in `Papineni et al. 2002`__.
__ https://www.aclweb.org/anthology/P02-1040
In addition, a review of smoothing techniques can be found in `Chen et al. 2014`__
__ http://acl2014.org/acl2014/W14-33/pdf/W14-3346.pdf
- ``update`` must receive output of the form ``(y_pred, y)`` or ``{'y_pred': y_pred, 'y': y}``.
- `y_pred` (list(list(str))) - a list of hypotheses sentences.
- `y` (list(list(list(str))) - a corpus of lists of reference sentences w.r.t hypotheses.
Remark :
This implementation is inspired by nltk
Args:
ngram: order of n-grams.
smooth: enable smoothing. Valid are ``no_smooth``, ``smooth1``, ``nltk_smooth2`` or ``smooth2``.
Default: ``no_smooth``.
output_transform: a callable that is used to transform the
:class:`~ignite.engine.engine.Engine`'s ``process_function``'s output into the
form expected by the metric. This can be useful if, for example, you have a multi-output model and
you want to compute the metric with respect to one of the outputs.
By default, metrics require the output as ``(y_pred, y)`` or ``{'y_pred': y_pred, 'y': y}``.
device: specifies which device updates are accumulated on. Setting the
metric's device to be the same as your ``update`` arguments ensures the ``update`` method is
non-blocking. By default, CPU.
average: specifies which type of averaging to use (macro or micro)
for more details refer https://www.nltk.org/_modules/nltk/translate/bleu_score.html
Default: "macro"
Examples:
.. code-block:: python
from ignite.metrics.nlp import Bleu
m = Bleu(ngram=4, smooth="smooth1")
y_pred = "the the the the the the the"
y = ["the cat is on the mat", "there is a cat on the mat"]
m.update(([y_pred.split()], [[_y.split() for _y in y]]))
print(m.compute())
.. versionadded:: 0.4.5
.. versionchanged:: 0.4.7
- ``update`` method has changed and now works on batch of inputs.
- added ``average`` option to handle micro and macro averaging modes.
"""
def __init__(
self,
ngram: int = 4,
smooth: str = "no_smooth",
output_transform: Callable = lambda x: x,
device: Union[str, torch.device] = torch.device("cpu"),
average: str = "macro",
):
if ngram <= 0:
raise ValueError(f"ngram order must be greater than zero (got: {ngram})")
self.ngrams_order = ngram
self.weights = [1 / self.ngrams_order] * self.ngrams_order
self.smoother = _Smoother(method=smooth)
if average not in ["macro", "micro"]:
raise ValueError(f'Average must be either "macro" or "micro" (got: {average})')
self.average = average
super(Bleu, self).__init__(output_transform=output_transform, device=device)
def _n_gram_counter(
self,
references: Sequence[Sequence[Sequence[Any]]],
candidates: Sequence[Sequence[Any]],
p_numerators: torch.Tensor,
p_denominators: torch.Tensor,
) -> Tuple[int, int]:
if len(references) != len(candidates):
raise ValueError(
f"nb of candidates should be equal to nb of reference lists ({len(candidates)} != "
f"{len(references)})"
)
hyp_lengths = 0
ref_lengths = 0
# Iterate through each hypothesis and their corresponding references.
for refs, hyp in zip(references, candidates):
# For each order of ngram, calculate the numerator and
# denominator for the corpus-level modified precision.
for i in range(1, self.ngrams_order + 1):
numerator, denominator = modified_precision(refs, hyp, i)
p_numerators[i] += numerator
p_denominators[i] += denominator
# Calculate the hypothesis lengths
hyp_lengths += len(hyp)
# Calculate the closest reference lengths.
ref_lengths += _closest_ref_length(refs, len(hyp))
return hyp_lengths, ref_lengths
def _brevity_penalty_smoothing(
self, p_numerators: torch.Tensor, p_denominators: torch.Tensor, hyp_length_sum: int, ref_length_sum: int,
) -> float:
# Returns 0 if there's no matching n-grams
# We only need to check for p_numerators[1] == 0, since if there's
# no unigrams, there won't be any higher order ngrams.
if p_numerators[1] == 0:
return 0
# If no smoother, returns 0 if there's at least one a not matching n-grams]
if self.smoother.smooth == "no_smooth" and min(p_numerators[1:]).item() == 0:
return 0
# Calculate corpus-level brevity penalty.
if hyp_length_sum < ref_length_sum:
bp = math.exp(1 - ref_length_sum / hyp_length_sum) if hyp_length_sum > 0 else 0.0
else:
bp = 1.0
# Smoothing
p_n = self.smoother(p_numerators[1:], p_denominators[1:])
# Compute the geometric mean
s = [w_i * math.log(p_i) for w_i, p_i in zip(self.weights, p_n)]
gm = bp * math.exp(math.fsum(s))
return gm
def _sentence_bleu(self, references: Sequence[Sequence[Any]], candidates: Sequence[Any],) -> float:
return self._corpus_bleu([references], [candidates])
def _corpus_bleu(
self, references: Sequence[Sequence[Sequence[Any]]], candidates: Sequence[Sequence[Any]],
) -> float:
p_numerators: torch.Tensor = torch.zeros(self.ngrams_order + 1)
p_denominators: torch.Tensor = torch.zeros(self.ngrams_order + 1)
hyp_length_sum, ref_length_sum = self._n_gram_counter(
references=references, candidates=candidates, p_numerators=p_numerators, p_denominators=p_denominators,
)
bleu_score = self._brevity_penalty_smoothing(
p_numerators=p_numerators,
p_denominators=p_denominators,
hyp_length_sum=hyp_length_sum,
ref_length_sum=ref_length_sum,
)
return bleu_score
[docs] @reinit__is_reduced
def reset(self) -> None:
if self.average == "macro":
self._sum_of_bleu = torch.tensor(0.0, dtype=torch.double, device=self._device)
self._num_sentences = 0
if self.average == "micro":
self.p_numerators = torch.zeros(self.ngrams_order + 1)
self.p_denominators = torch.zeros(self.ngrams_order + 1)
self.hyp_length_sum = 0
self.ref_length_sum = 0
[docs] @reinit__is_reduced
def update(self, output: Tuple[Sequence[Sequence[Any]], Sequence[Sequence[Sequence[Any]]]]) -> None:
y_pred, y = output
if self.average == "macro":
for refs, hyp in zip(y, y_pred):
self._sum_of_bleu += self._sentence_bleu(references=refs, candidates=hyp)
self._num_sentences += 1
elif self.average == "micro":
hyp_lengths, ref_lengths = self._n_gram_counter(
references=y, candidates=y_pred, p_numerators=self.p_numerators, p_denominators=self.p_denominators
)
self.hyp_length_sum += hyp_lengths
self.ref_length_sum += ref_lengths
@sync_all_reduce("_sum_of_bleu", "_num_sentences")
def _compute_macro(self) -> torch.Tensor:
if self._num_sentences == 0:
raise NotComputableError("Bleu must have at least one example before it can be computed.")
return self._sum_of_bleu / self._num_sentences
@sync_all_reduce("p_numerators", "p_denominators", "hyp_length_sum", "ref_length_sum")
def _compute_micro(self) -> float:
bleu_score = self._brevity_penalty_smoothing(
p_numerators=self.p_numerators,
p_denominators=self.p_denominators,
hyp_length_sum=self.hyp_length_sum,
ref_length_sum=self.ref_length_sum,
)
return bleu_score
[docs] def compute(self) -> None:
if self.average == "macro":
return self._compute_macro()
elif self.average == "micro":
return self._compute_micro()