Welcome to Model-Optimizer-YOLO Discussions!

[levipereira]

👋 Welcome to Model-Optimizer-YOLO Discussions

This is the place to discuss anything related to ONNX post-training quantization (PTQ) for YOLO-style models with NVIDIA Model Optimizer, COCO calibration, and TensorRT deployment using the model-opt-yolo CLI — questions, ideas, results, and workflows.

Most valuable posts here:

• 📊 Results — mAP, latency, throughput, and before/after PTQ on your hardware and dataset (share numbers and context).
• ⚙️ Parameter recipes — the best combinations you found per model (calibration size, image size, quant mode/method, autotune on/off, TensorRT flags, ORT/EP choices). Say which model + ONNX export they apply to so others can reproduce.

Some topics to get started:

• Pipeline and tuning — calibration set size, image size, quantization mode/method, optional Q/DQ autotune, trtexec / engine build flags
• Environments — Docker vs local install, CUDA / cuDNN / TensorRT / ONNX Runtime GPU alignment (especially CUDA 12 vs 13 and ORT builds)
• Datasets and calibration — COCO vs custom data, building calib.npy, preprocessing matching your ONNX
• Performance — latency, throughput, and mAP before/after PTQ on your hardware
• Exports and eval — ONNX export conventions (--dynamic, opset), eval-trt output formats (onnx_trt, ultralytics, deepstream_yolo, …)
• Unexpected behavior — when you are not sure it is a bug, ask here first
• Improving the project — suggestions for the CLI, docs, Docker image, examples, or the repo as a whole; feature ideas and constructive feedback are welcome

Support policy: Discussions are the main place for how-to, configuration, and sharing results. Issues are best reserved for confirmed bugs 🐛 with clear, reproducible steps (versions, commands, minimal inputs).

If you use this project in a pipeline, sharing benchmarks, lessons learned, and real-world setups helps everyone.

Welcome — share what you are building. 🚀