Why are x86_64 Linux wheels so large?

[johnthagen]

If you look at a release that shows all supported platforms: https://pypi.org/project/open3d/0.14.1/#files

You'll notice the Linux x84_64 wheel is 400MB while most other wheels are between 30-70MB.

We are deploying Open3D within a Linux x86_64 Docker image, and so we are trying to minimize the final image size. Why are these x86_64 Linux wheels so large, especially when they still dynamically link to libgl1 and libgomp1?

Is there any way these wheels could be compiled in a more optimal way such that their size would better match ARM Linux, macOS, or Windows wheels?

[johnthagen]

Looking into the wheel itself I see cpu/pybind.cpython-39-x86_64-linux-gnu.so at 177MB and cuda/pybind.cpython-39-x86_64-linux-gnu.so at 752MB. These seem to be adding support for Tensorflow and Pytorch.

I wonder if there would be a way to publish an open3d-lite or similar where users could opt for a lighter weight Open3D wheel if they are not using these heavier features.

Alternatively, open3d could be the stripped down version and open3d-ml could be packaged that includes Tensorflow and Pytorch support.

OpenCV does this with opencv-python-headless, which is a stripped down version of OpenCV.

[johnthagen]

Feature request: #5124