128ea5274db96163c58a99405e7205abcab36192,onmt/Translator.py,Translator,translateBatch,#Translator#Any#Any#,84

Before Change

        def unbottle(m):
            return m.view(beamSize, batchSize, -1)

        while i < self.opt.max_sent_length \
              or any([len(b.finished) == 0 for b in beam]):
            // Construct batch x beam_size nxt words.
            // Get all the pending current beam words and arrange for forward.
            inp = var(torch.stack([b.getCurrentState() for b in beam])
                      .t().contiguous().view(1, -1))
            
            // Turn any copied words to UNKs
            // 0 is unk
            if self.copy_attn:
                inp = inp.masked_fill(inp.gt(len(self.fields["tgt"].vocab) - 1), 0)
                            
            // Run one step.
            decOut, decStates, attn = \
                self.model.decoder(inp, src, context, decStates)
            decOut = decOut.squeeze(0)
            // decOut: beam x rnn_size
            
            // (b) Compute a vector of batch*beam word scores.
            if not self.copy_attn:
                out = self.model.generator.forward(decOut).data
                out = unbottle(out)
                // beam x tgt_vocab
            else:                
                out = self.model.generator.forward(decOut, attn["copy"].squeeze(0),
                                                   srcMap)
                // beam x (tgt_vocab + extra_vocab)
                out = dataset.collapseCopyScores(
                    unbottle(out.data),
                    batch, self.fields["tgt"].vocab)
                // beam x tgt_vocab
                out = out.log()

            // (c) Advance each beam.
            for j, b in enumerate(beam):
                is_done = b.advance(out[:, j], unbottle(attn["copy"]).data[:, j])
                decStates.beamUpdate_(j, b.getCurrentOrigin(), beamSize)
                if is_done:
                    break
            i += 1
            
        if "tgt" in batch.__dict__:
            allGold = self._runTarget(batch, dataset)
        else:
            allGold = [0] * batchSize

After Change

            return m.view(beamSize, batchSize, -1)

        for i in range(self.opt.max_sent_length):
            if all((b.done() for b in beam)):
                break

            // Construct batch x beam_size nxt words.

Instances