Support n_best option for CTranslate

guillaumekln · January 19, 2017, 10:24am

Like translate.lua, CTranslate should support the n_best option.

wsnooker · February 21, 2017, 1:15pm

Which version the n_best option will be supported for CTranslate ?

You could give some tips for us to finish it if you are busy for something else.

Be wish to take n_best as soon as possible.

guillaumekln · February 21, 2017, 2:03pm

No definite plans for it. Are you building an application around CTranslate? In the meantime, is using translate.lua an option in your use case?

wsnooker · February 22, 2017, 1:48pm

Yes, I’m planning to build an application around CTranslate.

And could you give some advice for supporting n_best output ?

guillaumekln · February 22, 2017, 2:12pm

If you want to take a look:

github.com

OpenNMT/CTranslate/blob/master/include/onmt/Translator.hxx#L425-L429


// Copy encoder states to decoder states.
std::vector<MatFwd> rnn_state_dec;
rnn_state_dec.reserve(rnn_state_enc.size());
for (size_t l = 0; l < rnn_state_enc.size(); ++l)
{

These data structures hold the beam search result. In particular, you want to extract N hypotheses from next_ys (with their corresponding next_features, scores and all_attention) following beam paths described in prev_ks.

Here is the piece of code that extracts the best hypotheses for a batch:

github.com

OpenNMT/CTranslate/blob/master/include/onmt/Translator.hxx#L739-L788


      input_feed = new_input_feed;
  }


  remaining_sents = new_remaining_sents;
}


// Build final translation by following the beam path for each batch.
std::vector<std::vector<std::string> > batch_tgt_tokens;
std::vector<std::vector<std::vector<std::string> > > batch_tgt_features;
std::vector<std::vector<std::vector<float> > > batch_attention;


for (size_t b = 0; b < batch_size; ++b)
{
  size_t start_k = best_k[b];
  size_t len = best_finished_at[b] + 1;
  if (end_score[b] > max_score[b]) // End score is the score of the top beam.
  {
    start_k = 0;
    len = end_finished_at[b] + 1;
  }

This file has been truncated. show original