Adding support for 3D deep learning models.

[VikasSoni1]

Feature request

Hi. I am planning to add a new pipeline and a model for 3d deep learning tasks which can work on point clouds for classification and detection as there is no support for 3d data right now. I just wanted to confirm if the process will be similar to the guides for adding a new pipeline and model to hugging face transformers or there are more complexities which I have not thought about? And is it going to be too much work to add GPU support and batching?

Motivation

I have been working with 3d deep learning and wanted to implement the whole process from scratch. So, why not contribute to hugging face so other people can use and build upon it?

Your contribution

Submitting a PR

[amyeroberts]

Hi @VikasSoni1, thanks for adding this feature request!

Pipelines are typically built around groups of models that tackle the same task and share inputs/outputs using the AutoXxx API. We don't currently have models which operate on point clouds or 3D data (additions always welcome!) and so there isn't a defined API yet and these should be defined before adding such a pipeline.

The great thing about open source, is anyone can fork this repo and built upon it. If you've developed your own pipelines - please feel free to share here for the community to find 🤗