How to use fastai.text with word embedding but not transfer learning


#1

I am wondering if there is a way to use fastai.text with word embedding, such as fasttext, without using transfer learning?

in lesson 3, we are using wt103 pretrained model to train an imdb specific language model then a classification model. I suppose we can also train a specific language model using the word embedding without a pretrained language mode. But I am not sure how to do it with fastai.text. Could some one point me into the right direction so that I can experiment with it?

Thanks in advance!


(Miguel) #2

I tried to do that for this competition but without success. My approach was to directly replace the embedding matrices on the fastai model by the word embedding. I haven’t explored this much so I may have missed something or have some bug on my code.