def computeBLEU(outputs, targets, corpus=False, tokenizer=None, segmenter=None):
outputs = [tokenizer(o) for o in outputs]
targets = [tokenizer(t) for t in targets]
if segmenter is not None:
outputs = segmenter(outputs)
targets = segmenter(targets)
if not corpus:
return torch.Tensor([sentence_gleu(
[t], o) for o, t in zip(outputs, targets)])
return corpus_bleu([[t] for t in targets], [o for o in outputs], emulate_multibleu=True)
def computeGroupBLEU(outputs, targets, tokenizer=None, bra=10, maxmaxlen=80):
outputs = [tokenizer(o) for o in outputs]
targets = [tokenizer(t) for t in targets]
maxlens = max([len(t) for t in targets])
print(maxlens)
maxlens = min([maxlens, maxmaxlen])
nums = int(np.ceil(maxlens / bra))
outputs_buckets = [[] for _ in range(nums)]
targets_buckets = [[] for _ in range(nums)]
for o, t in zip(outputs, targets):
idx = len(o) // bra
if idx >= len(outputs_buckets):
idx = -1
outputs_buckets[idx] += [o]
targets_buckets[idx] += [t]
for k in range(nums):
print(corpus_bleu([[t] for t in targets_buckets[k]], [o for o in outputs_buckets[k]], emulate_multibleu=True))