Closed michelole closed 5 years ago
Our dl4j-based code seems to "freeze" them already, because the input is featurized by ourselves and is not subject to training.
For fasttext, follow https://github.com/facebookresearch/fastText/issues/811.
In #101 we saw that pre-trained embeddings show a difference with fewer epochs, so there's no need to freeze them now.
Both on fasttext and dl4j.