anderleich
November 5, 2021, 12:21pm
1
Hi,
I need the word-normalized conditional cross-entropy of the probability distribution for each translated sentence. I found this line in code which has the predicted sentece score.
start_time = time.time()
for batch in data_iter:
batch_data = self.translate_batch(
batch, data.src_vocabs, attn_debug
)
translations = xlation_builder.from_batch(batch_data)
for trans in translations:
all_scores += [trans.pred_scores[: self.n_best]]
pred_score_total += trans.pred_scores[0]
pred_words_total += len(trans.pred_sents[0])
if tgt is not None:
gold_score_total += trans.gold_score
gold_words_total += len(trans.gold_sent) + 1
n_best_preds = [
" ".join(pred) for pred in trans.pred_sents[: self.n_best]
]
if self.report_align:
align_pharaohs = [
Is this what I’m looking for?
Hi,
Yes, what you are looking for is PRED AVG SCORE in the verbose logs.