I want to try something quite adventurous: using Voyager Wingman to port a model to Axelera.
ZipDepth is a 6.1 million parameter monocular depth estimation model which has been released recently and has achieved SOTA status. The goal is a live, entirely local demo that turns an ordinary webcam feed into a real-time relative depth map. ZipDepth is an edge oriented model and it's made of CNNs, lightweight attention, pulling and resizing operations. This project will test whether Wingman can help to integrate and debug a newly released model that is not yet in Axelera's model zoo.
$ THE PROMPT $
Wingman, help me deploy LibreYOLO’s ZipDepth implementation on Metis.
Use:
- Repository: https://github.com/LibreYOLO/libreyolo
- Branch: dev
- Model: LibreZipDepthbnpu-depth.pt
Export a static batch-one ONNX model, compile and quantize it for Metis, and
report any unsupported operations or host fallbacks. Do not change the
architecture without asking me first.
Then build a live webcam pipeline showing RGB and colourised relative inverse
depth side by side, with measured latency and FPS. Validate the Metis output
against LibreYOLO’s PyTorch and ONNX outputs, and preserve all code,
configuration, prompts and instructions needed to reproduce the project.
