Why do you normalize the vectors different compared to the original word2vec code? med = sqrt(med / vector_size); len = sqrt(len);