HashingTF knows the output feature dimension, which should be in the output ML attributes.
[Github] Pull Request #6308 (mengxr)