How ensemble decoding


(Xiang Zhu0718) #1

I have trained 4 models with the different seed of 1000, 2000,3000 and 4000 . The source and target files are all the same for the 4 models.
Now, what can i do for the ensemble decoding? or does the currenr version of OpenNMT support ensemble decoding? and if ture, how to ?


(Eva) #2

Hi @XiangZhu0718 !

in this post you can find more information about the ensemble decoding available in OpenNMT:

You will find there the link to the OpenNMT-ensemble feature implementation by @guillaumekln . This will allow you to average the probabilities produced by your models just using the translate.lua script by indicating the list of models.

It’s been a while since the last time I used it, but I guess it will still be compatible with the last versions of OpenNMT.

Cheers,
Eva


(Xiang Zhu0718) #3

thanks a lot for your answers. i trained the model in pytorch, it seems that dose not support ensemble decoding feature.
any tools to convert the pytorch model to torch ?


(Guillaume Klein) #4

There is no model conversion tool at the moment.

For ensembling, see here: