Sonny Hu

+1. Is there a plan to release a version of `tensorflow-workspace` that work with kmodel format?

@lharries Thanks for the reply! I am hoping to edit words/sentences in the speaker cards then download the dubbed video with those edited words/sentences through API. I can do this...

Yup. Create a project in Dubbing Studio -> edit in speaker cards -> download the edited version, all through API.

@lharries hey, any update on if there is a plan to support this? Thanks

@bottler Thanks for the reply! The error msg `"RuntimeError: Cannot access data pointer of Tensor that doesn't have storage"` from #1533 happened in the **forward** pass `"idx, dists = _C.knn_points_idx(p1,...

@bottler I wrote a toy example that follows [this pytorch3d tutorial](https://pytorch3d.org/tutorials/deform_source_mesh_to_target_mesh) with following modifications: 1. Only use `_PointFaceDistance` in the objective. Because I only care if we can compute jacobian,...

@TimoRST Thanks for the inputs! I never thought of it and will look into it. Just one follow up, `jac_fn` is either the `cost_function` which is `th.AutoDiffCostFunction` or the `error_fn`...

@TimoRST Thanks. I will try it out. Can I ask what your use case was to use `knn` and `theseus`? Was it also used in the context of an optimization...

@TimoRST Thanks for the details. Really appreciate the help! @bottler could you take a look at the code I posted [above](https://github.com/facebookresearch/pytorch3d/issues/1796#issuecomment-2125804370)? I am going to follow @TimoRST suggestion to take...

Here are some findings. Conclusion: `_C.point_face_dist_backward` cannot take in `grad_dists` that is a 2D `BatchedVector` which is created by `vmap/jacrev`. - When calling `torch.autograd.functional.jacobian`, it computes jacobian row-by-row. `grad_dists` is...