search

harvardnlp / seq2seq-attn

Sequence-to-sequence model with LSTM encoder/decoders and attention
http://nlp.seas.harvard.edu/code
MIT License
1.26k stars 278 forks source link

model ensemble #21

Closed zcyang closed 7 years ago

zcyang commented 8 years ago

Is model ensemble part available? How can we use the code to do model ensemble?

yoonkim commented 8 years ago

its not part of the current code. you want to modify beam.lua so it takes in multiple model file paths and use them to generate multiple (log) probabilities over the next word, then average them during beam search.