rf-mediapipe - fork of mediapipe-numpy2

A patching tool to remove the numpy<2 constraint from official mediapipe wheels.

Why does this repository exist?

The official mediapipe Python wheels published on PyPI include a strict dependency:

Requires-Dist: numpy<2

This prevents installation or use of mediapipe with numpy version 2 or higher. This constraint is problematic for users and downstream projects who want to use the latest numpy (see mediapipe issue #5612). As of now, a lot of mediapipe seems to work with numpy 2.x, but the PyPI wheels block installation.

What does this repository do?

This repository provides a script (patch_wheels.sh) that:

• Downloads all official mediapipe wheels for a given version from PyPI.
• Removes the numpy<2 requirement from the wheel metadata.
• Renames the package to mediapipe-numpy2 to avoid conflicts with the official package.
• Updates the wheel metadata to reference this repository.
• Rebuilds and outputs patched wheels that can be installed alongside numpy>=2.

Usage

To use mediapipe with numpy>=2, install the pre-patched wheels from the releases as a drop-in replacement:

pip install mediapipe-numpy2

Then import mediapipe as you normally would:

import mediapipe as mp

Patch Mediapipe

1. Clone this repository:

   git clone https://github.com/roboflow/mediapipe-numpy2.git
   cd mediapipe-numpy2

2. Run the patching script:

   ./patch_wheels.sh

   By default, this will:

   • Download the latest mediapipe wheels into ./downloaded_wheels
   • Output patched wheels into ./patched_wheels

   You can override the mediapipe version and directories using environment variables:

   VERSION=0.10.21 DOWNLOAD_DIR=./my_dl OUTPUT_DIR=./my_patched ./patch_wheels.sh

3. Install a patched wheel:

   pip install ./patched_wheels/rf_mediapipe-*.whl

Notes

• The patched wheels are named rf_mediapipe to avoid clashing with the official package.
• This repository is not affiliated with the official mediapipe project.
• Use at your own risk; always test compatibility with your workflow.