Beam search score

kingomalek · December 3, 2017, 7:05pm

Hello,
As far as I know, beam search’s final result is the path with the highest probability. Where can I can the value of this probability inside OpenNMT-lua ? is there a way to add some paths manually to the search tree ?

guillaumekln · December 4, 2017, 8:09am

Hello,

The final scores are gathered here:

github.com

OpenNMT/OpenNMT/blob/master/onmt/translate/Translator.lua#L422




return feats
end


function Translator:translateBatch(batch)
local encStates, context = self.model.models.encoder:forward(batch)
if self.args.dump_input_encoding then
  return encStates[#encStates]
end


-- if we have a language model - initialize lm state with BOS
local lmStates, lmContext
if self.lm then
  local bos_inputs = torch.IntTensor(batch.size):fill(onmt.Constants.BOS)
  if #self.lm.dicts.src.features > 0 then
    local inputs = { bos_inputs }
    if #self.lm.dicts.src.features > 1 then
      table.insert(inputs, {})
      for _ = 1, #self.lm.dicts.src.features do
        table.insert(inputs[2], bos_inputs)
      end

You can hack you way in the search tree in the code that is looking for the next candidates:

github.com

OpenNMT/OpenNMT/blob/master/onmt/translate/BeamSearcher.lua#L101


    finished[finishedBatches[b]] = finishedHypotheses[b]
    histories[finishedBatches[b]] = finishedHistory[b]
  end
  t = t + 1
  remaining = beams[t]:getRemaining()
end
return finished, histories
end


-- Generate kBest tokens based on top scores.
function BeamSearcher:_findKBest(beams, vocabSize, kBest, expandedScores, expandedNormScores)


local function topk(tensor, ...)
  if torch.typename(tensor) == 'torch.CudaHalfTensor' then
    tensor = tensor:cuda()
  end
  return tensor:topk(...)
end


local t = #beams

I hope these entry points are useful to you.

kingomalek · December 4, 2017, 9:14am

thank you this is very helpful