Scaling design for NN (Nearest Neighbour) and BILINEAR


#21

I suggest two different API at TVM and single interface at NNVM.

bilinear approach require additional param weights to be passed, hence a different spec at TVM.
weights calculation will be handled part of NNVM compilation, hence frontend could have same API.

To unify the front_end scaling factor would be replace with tuple, hence support asymetric across height and width could be supported.


#22

Ref. https://github.com/dmlc/tvm/pull/1181

Bilinear scaling implementation.