I’m novice in DL and embeddings theory.
I want to use embeddings(from trained fast.ai model) instead of one-hot-encoding in xgboost/lightgbm and etc
for example consider lesson6-rossmann.ipynb:
I’ve got weights of embedding of first categorical variable use code below:
emb_0 = to_np(learn.model.embeds.weight.data)
emb_0.shape = (1116, 81) but the number of objects in train_df is 844338
and I want to convert ebm_0 to features matrix with size (844338, ?) for gbm model
Is it possible to make such thing? And how?