* read tensor data as numpy array instead of python list

this is faster by 700 times

* apply code-format changes

* read char tensor as numpy array instead of python list