We have trained a good language model, using awd_lstm architecture. Then we have saved it using
Now we would like to build our own classifier. We can get the
fastai.text.models, but then we need to separately store at least the vocab_sz, emb_sz, n_hid and n_layers - the required arguments.
We thought about creating a new learner, loading the saved artifact and then extracting the encoder from there, but this seems like overkill.
Is there a cleaner way to load the encoder from the saved