model ensemble

[zcyang]

Hi,

Is model ensemble part available? Or at least can you explain how to do model ensemble.

Here is my currently understanding:

Give the sequence x_1, x_2,...x_n, and two models m_1, m_2. First use the encoder part of m_1, m_2 to get the states s_1, s_2, then feed the states to the decoder parts of m_1, m_2, the first input to the decoder is EOS, then the decoder generates the probability for the next words p_1, p_2, average the probability to get (p_1 + p_2)/2, then pick the top k words according to the average probability as candidates, and then use them to feed into the next steps, the states s_1, s_2 are both updated after the first step.

Is this correct? I tried to implement using the above way, but don't find improvement using several models.