Interpretation of 'decoderInputs' in dataset.lua

[nabihach]

I'm looking at dataset.lua, where decoderInputs are being set up.

    decoderInputs = torch.IntTensor(maxTargetOutputSeqLen-1,size):fill(0)

    for samplenb = 1, #targetSeqs do
      trimmedEosToken = targetSeqs[samplenb]:sub(1,-2)
      for word = 1, trimmedEosToken:size(1) do
        if size == 1 then
          decoderInputs[word] = trimmedEosToken[word]
        else
          decoderInputs[word][samplenb] = trimmedEosToken[word]
        end
      end
    end

This tensor is then used in model.decoder:forward(decoderInputs) in train.lua. My question is, how (and why) is this tensor different from encoderInputs?

I ask this because when we evaluate the model using eval.lua, we simply pass the decoder's own output at each time step into model.decoder:forward (see Seq2Seq:eval(input), where we have the line
local prediction = self.decoder:forward(torch.Tensor(output))[#output] ).

Help would be appreciated.