Support for unpooling operation?

Questions
1

I’m interested in using TVM to compile models that use unpooling operations (where an input feature map is upsampled by inserting rows and columns of zeros between input values). I’ve been looking at the NNVM Core Tensor operators listed here (https://docs.tvm.ai/nnvm_top.html) but am not sure if any of those operators (or a combination of them) could be used to implement unpooling.

I’m also aware of the TVM operator inventory (TOPI). If I am interested in supporting unpooling, would I need to write a custom TOPI function?

Thanks!

2

For average unpool probably could use upsampling with (nearest neighbour).
But maxpool is irreversible unless we know the location of maxima with in each window.
You may refer RFC https://arxiv.org/pdf/1311.2901v3.pdf for soem details on unpooling.