EmbeddingDotBias, changing Embeddings after train

hi, I setup my learner i. e.

learn = collab_learner(dls, n_factors=5, y_range=[0,5.5])

I could change i.e. the user embeddings before I start to train

learn.model.u_weight = Embedding(9,5)

and it works, I can fit and the embedding is trained.

But once I trained with fit_one_cycle, if I still change learn.model.u_weight the embedding sounds freezed and it doesn’t change anymore.

Anyone knows why? Thank you.