Onnx auto-tune problem

I have a problem when I do auto-tune by using onnx model.Any ideas to solve this?

I find that the reason is that upsample.I make model by using feature size.It will make inputs[1] is not var class.How to solve this?@tqchen Could ypu please support this way to make model?Upsample details in onnx as following:

I modify upsample and new problem as following:

Now upsample like this: