[Relay][Frontend][Onnx] MaxUnpool Operator

[jwfromm]

I recently encountered a model with that uses the MaxUnpool op, a pretty interesting operator that is designed to reverse a maxpool. I've added support for it in our frontend through some sneaky use of scatter and found that it matches onnx's execution for each special case of the operator.

[jwfromm]

@masahi @mbrookhart can you guys take a look at this PR?

[jwfromm]

Whoops, I tested this with a newer version of onnxruntime that can do lenient shape merges. I'll fix the test to specify the expected output shape.

[masahi]

maybe it is time to upgrade ONNX and ORT on CI? We are at ONNX v1.6, while the latest is 1.8. Our ORT is hopelessly old: We are at 1.00 while the latest is 1.52.

I can take a look at CI upgrade for ONNX this month @jwfromm. I did that for PyTorch last month.

[jwfromm]

@masahi that would be nice, although in this case I was just being lazy and not calculating the output shape. I've fixed the test now if you want to take another look.

[masahi]

Interesting that we can support max unpool without modifying max pooling to return indices.

[masahi]

thanks @jwfromm

[jwfromm]

@masahi you're right that most models will use MaxUnpool with MaxPool returning indices, and we probably want to add that option to MaxPool operators soon.