Change TopK operator to allow dynamic 'k'

[hariharans29]

This PR is modifying TopK operator from having 'k' as attribute to having it as an input

[CLAassistant]

All committers have signed the CLA.

[hariharans29]

CC: @linkerzhang

[linkerzhang]

LGTM. Thank you!

[linkerzhang]

@houseroad any comments please? thanks a lot!

[hariharans29]

@houseroad @linkerzhang - Is this change good to go ? Thanks!

[spandantiwari]

@hariharans29 - there was a failing backend test which is why the CI has failed. That test (test_densenet) is disabled for now. Could you try running the CI job again to see if this passes?

[houseroad]

Sorry to block:

Yes, we do use topk in pytorch-onnx-caffe2
https://github.com/pytorch/pytorch/blob/master/torch/onnx/symbolic.py#L1305

[spandantiwari]

@houseroad - just FYI - the failing test was test_densenet, which we, coincidentally, disabled today (pytorch/pytorch#17696) for the Upsample PR (#1773).

[hariharans29]

Sorry to block:

Yes, we do use topk in pytorch-onnx-caffe2
https://github.com/pytorch/pytorch/blob/master/torch/onnx/symbolic.py#L1305

@houseroad - Sorry I don't quite understand. Could you please elaborate on the changes you are requesting ?

[houseroad]

@hariharans29 I don't mean you need to change anything, but we need to prepare for the change. Otherwise, it will break our internal pipeline. I am thinking of setting the pytorch onnx exporter to some stable opset by default. So such changes won't cause problem anymore.

[hariharans29]

@houseroad - Got it. I will wait for your go-ahead. Thanks.

[houseroad]

Good to ship now.

[hariharans29]

@houseroad - Thanks a lot!