Nms implemented in torch purly, supporting export onnx

Off-Topic Community Showcase
,
1

(Here)[https://github.com/VectorElectron/yolov11\] is a sleek, professional, and exciting draft for your post, tailored for platforms like GitHub, LinkedIn, or developer forums (like WeChat Moments or Twitter/X).

:rocket: Fully End-to-End YOLO: Merging Resize, Core Model, and NMS into a Single ONNX File!

Tired of dealing with messy preprocessing (resize/normalization) and heavy Python/C++ post-processing (NMS) outside of your model? I’ve wrapped the entire pipeline—from raw uint8 images to final bounding boxes—into one single ONNX file.

This approach delivers true plug-and-play, end-to-end deployment for web frontends, edge devices, or localized runtimes.

:glowing_star: Key Features

  • Zero-Dependency Pipeline: No more manual bounding box parsing or custom NMS C++ modules. The model accepts a raw [H, W, 3] uint8 image and outputs final, scaled coordinates directly.

  • Mathematical NMS: Uses a batch-and-class isolated IoU graph lookup table (LUT) implemented directly in PyTorch operators—fully compatible with ONNX export.

  • Fully Dynamic Input: Supports dynamic image resolutions, automatically scaling coordinates back to the original input size.

  • Seamless Merging: Combines three independent stages (Preprocess, Core YOLO, and Post-NMS) into a unified computational graph using onnx.compose.