glcLucky opened this issue 4 years ago · 1 comments
I guess this formula may be wrong. Should we change that to this: np.sqrt(epsilon / embedding_dim)？
Why do you say so ?