add None option for chamfer distance point reduction

[haritha-j]

The chamfer_distance function currently allows "sum" or "mean" reduction, but does not support returning unreduced (per-point) loss terms. Unreduced losses could be useful if the user wishes to inspect individual losses, or perform additional modifications to loss terms before reduction. One example would be implementing a robust kernel over the loss.

This PR adds a None option to the point_reduction parameter, similar to batch_reduction. In case of bi-directional chamfer loss, both the forward and backward distances are returned (a tuple of Tensors of shape [D, N] is returned). If normals are provided, similar logic applies to normals as well.

I'm not sure if this is the best return format, or if it should be returned as a padded Tensor (shape [2, D, N]) instead. It would be great if this could be clarified. I didn't add any test cases since they would depend on the return format.

This PR addresses issue #622.

[bottler]

Thank you!

I know we had a None option which was not useful and was removed in 790eb8c . I think this is not the same as that, right?

Can you specify (in the docstring(s)) the shape of the output in the new case and add a test for the new case?

[haritha-j]

Yes, this is quite different. Before 790eb8c setting point_reduction to None would really just result in point reduction anyway because batch_reduction would cause at least one reduction, and there was a check that prevented both point_reduction and batch_reduction from being None. The commit was to prevent this.

This PR allows both of these to be None, allowing the return of the individual loss terms for each point.

The return shape is specified in the doc string and tests have been added.

[facebook-github-bot]

@bottler has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[facebook-github-bot]

@bottler merged this pull request in d84f274.

[bottler]

Thank you!