YOLOv9: programmable gradients for object detection

What it does

This is the official PyTorch implementation of the YOLOv9 paper, which adds “Programmable Gradient Information” to the standard YOLO object-detection pipeline. It trains and runs inference on the MS COCO dataset, supports single- and multi-GPU setups, and ships with five model sizes from Tiny (2.0M parameters) to Extra (57.3M parameters).

The interesting bit

The core claim is PGI — a mechanism to control what gradient information flows through the network so the model learns task-relevant features rather than spurious correlations. The repo also includes GELAN, a lighter architecture family, and a re-parameterization notebook for compressing trained models.

Key highlights

• Five pretrained checkpoints: YOLOv9-T/S/M/C/E, plus GELAN variants
• MS COCO validation numbers published: YOLOv9-E hits 55.6% AP at 640px
• Docker setup with NVIDIA PyTorch base image; distributed training via torch.distributed.launch
• Export recipes scattered across issues: ONNX, TensorRT, TFLite, OpenVINO, ONNXSlim
• Community integrations tracked in issues: ROS, Julia, MLX, ByteTrack, DeepSORT, Comet, MLflow

Caveats

• README is mostly a performance table and shell snippets; the actual PGI mechanism is not explained in the repo — you need the paper
• “Useful Links” section is a dump of GitHub issue URLs, not organized documentation
• Some model variants and the ReLU ablation are marked “will be released after the paper be accepted”

Verdict

Worth a look if you’re running YOLO in production and want to experiment with gradient-level architectural tweaks. Skip if you need polished tooling or a stable API — this is a research release with community glue code.