I understand that Jeremy replaced the one hot encodings with Word2Vec embeddings, but it seems he used
two types of Word2Vec embeddings?
I am a little confused about the entire process.
First he downloaded the Word2Vec embeddings from Google"s site
He then searched the label names from Imagenet and got the Word2embeddings for those he could find.
He replaced the label one hot encodings with these word vectors
I didnt quite understand what he did with those he didnt find the label for.
Did he manually search for a similar item, and get the word embeddings that way?
He trained the model, and it seemed to be working?
FROM here on is where I am confused
- He then decided to expand the word embeddings to use 90,000 or nouns.
However if Imagenet is confined to 1000 categories, wouldnt that mean the word embeddings would be limited to hose 1000 categories? I watch the video two times and went through the notebooks, but I"m still a little confused.