Cross Op

[mtngld]

Currently, ONNX does not include the Cross op.

PyTorch cross:
https://pytorch.org/docs/stable/torch.html#torch.cross

TF cross:
https://www.tensorflow.org/api_docs/python/tf/linalg/cross

[ebarsoum]

Which model need it?

[mtngld]

It is used often for geometric related models, for example in this pytorch_geometric library

https://github.com/rusty1s/pytorch_geometric

https://github.com/rusty1s/pytorch_geometric/blob/5692a8613d9340c76545e59b4d37cd4ced6d340e/torch_geometric/nn/conv/ppf_conv.py#L11

[ALLinLLM]

@mtngld I met the same problem and work around it by the following code

        # normal = tu.cross(tv, dim=3) 
        normal = torch.stack((
        tu[:,:,:,1]*tv[:,:,:,2] - tu[:,:,:,2]*tv[:,:,:,1],
        tu[:,:,:,2]*tv[:,:,:,0] - tu[:,:,:,0]*tv[:,:,:,2],
        tu[:,:,:,0]*tv[:,:,:,1] - tu[:,:,:,1]*tv[:,:,:,0]), dim=3)
        # print("cross all close", torch.allclose(normal, normal2))

the code above is following the cross formula on wikipedia.

BTW, @ebarsoum In the 3d or rendering tasks, the cross operator is frequently used to get normal vector of a panel. For example, the CVPR 2020 best paper unsup3d, and ECCV 2020 best paper are both about 3d and rendering. So I'm glad to see the support of cross operator. This will benefit the convert from the research to application.

[stale]

Is this still relevant? If so, what is blocking it? Is there anything you can do to help move it forward?

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.

[saeed1262]

Any chance this will be added to the supported ops?

[ivan-alles]

I need this too, to computing elements of a rotation matrix.

[justinchuby]

microsoft/onnxscript#512

[Francis235]

I need this op too, is this operator supported now?

[andife]

Being added to onnxscript means that the export should work for Pytorch?
But not a separate operator but a combination of other operators?

[justinchuby]

Being added to onnxscript means that the export should work for Pytorch? But not a separate operator but a combination of other operators?

That is correct. I am not aware of a strong motivation to standardize this operator in the ONNX spec