Add a predictor method to return more than one possible sequence

Would be possible to add to this library a predictor_n (or to modify the current predictor) method to return more than one sequence as result? I think it would be a great tool to have when using beam search (with TopKDecoder).

I coded a first attempt to do that (it seems to work, https://github.com/juan-cb/pytorch-seq2seq/commit/442431001b122fa15c4b6476a9d7411570f53f20), but I’m not sure if it is the best way to implement that or is completely correct. The desired behavior is to return the n most probable sequences given an src_seq.

Thanks in advance

[kylegao91]

Hi @juan-cb, sorry for the late reply. Please find my changes to your predict_n below. It should work, and let’s make a pr.

    def predict_n(self, src_seq, n=None):
        src_id_seq = Variable(torch.LongTensor([self.src_vocab.stoi[tok] for tok in src_seq]),volatile=True).view(1, -1)
        if torch.cuda.is_available():
            src_id_seq = src_id_seq.cuda()

        softmax_list, _, other = self.model(src_id_seq, [len(src_seq)])

        result = []
        for x in range(0, int(n)):
            length = other['topk_length'][0][x]
            tgt_id_seq = [other['topk_sequence'][di][0,x,0].data[0] for di in range(length)]
            tgt_seq = [self.tgt_vocab.itos[tok] for tok in tgt_id_seq]
            result.append(tgt_seq)

        return result

[kylegao91]

I think it’s nice feature to have and thanks for the attempted implementation.

I noticed that you use other["sequence"], which contains sequences for the input batch rather than the beams of each input. Instead, as noted here, the other['topk_sequence'] should be used to get the beam sequences.

Please feel free to make a pr from this, we can discuss from there.