[Torch] Support aten::tensor, aten::empty and aten::numel

[zhiqwang]

Hi TVM teams,

TVM currently support pytorch 1.4 and torchvision 0.5. I want to know is there any plan to support pytorch 1.6 version to trace the object detection models in pytorch 1.6?

For faster rcnn, which is not partly scripted and thus can be traced completely, I get following missing ops with PyTorch 1.6

NotImplementedError: The following operators are not implemented: ['aten::tensor', 'aten::empty', 'aten::numel']

Originally posted by @masahi in #6449 (comment)

[masahi]

Yes, we want to support a newer version but our quantization support is completely broken in PyTorch 1.6 due to the breaking change they introduced pytorch/pytorch#42497. I don't know a simple workaround for this.

Either we wait for Torch people to fix this issue or we spend significant effort to add workaround. Until then we are stuck with 1.4 or 1.5 (CI is at 1.4)

The ops aten::tensor, aten::empty, and aten::numel are very trivial ops actually, so if you want to just convert 1.6 detection models, adding support for them shouldn't be hard. You don't need to wait for 1.6 support.

[masahi]

We'll track 1.6 support at #6594

[zhiqwang]

Hi, That's great news! Relink it to #6594 ?

[masahi]

ah sorry, the link is fixed