def test_bleu_score(weights, n_gram, smooth_func, smooth):
nltk_output = sentence_bleu(
[REFERENCE1, REFERENCE2, REFERENCE3], HYPOTHESIS1, weights=weights, smoothing_function=smooth_func
)
pl_output = bleu_score([HYPOTHESIS1], [[REFERENCE1, REFERENCE2, REFERENCE3]], n_gram=n_gram, smooth=smooth)
assert torch.allclose(pl_output, torch.tensor(nltk_output))
nltk_output = corpus_bleu(LIST_OF_REFERENCES, HYPOTHESES, weights=weights, smoothing_function=smooth_func)
pl_output = bleu_score(HYPOTHESES, LIST_OF_REFERENCES, n_gram=n_gram, smooth=smooth)
assert torch.allclose(pl_output, torch.tensor(nltk_output))
def forward(self, translate_corpus: list, reference_corpus: list) -> torch.Tensor:
"""
Actual metric computation
Args:
translate_corpus: An iterable of machine translated corpus
reference_corpus: An iterable of iterables of reference corpus
Return:
torch.Tensor: BLEU Score
"""
return bleu_score(
translate_corpus=translate_corpus,
reference_corpus=reference_corpus,
n_gram=self.n_gram,
smooth=self.smooth,
).to(self.device, self.dtype)
def test_no_4_gram():
hyps = [["My", "full", "pytorch-lightning"]]
refs = [[["My", "full", "pytorch-lightning", "test"], ["Completely", "Different"]]]
assert bleu_score(hyps, refs) == torch.tensor(0.0)
def test_bleu_empty():
hyp = [[]]
ref = [[[]]]
assert bleu_score(hyp, ref) == torch.tensor(0.0)