After training two tower model, is it possible to use user-tower only to generate user embedding vector?

after training two tower model is done, it looks like we can take out the user-tower model only (i.e., without item tower) to generate a user embedding vector. Is it correct?

Also is there any easy way to develop a user embedding model to get user embedding vector instead of using two-tower model? I cannot think of how to set up a supervised learning (i.e., which label and which cost function).