init from yolov5_quant_sample

[wjx052333]

🤖 Generated by Copilot at 7a1a558

Summary

🎛️🐳🔢

This pull request adds quantization-aware training (QAT) and ONNX export capabilities to the yolov5 model, and provides tools and scripts for converting the ONNX model to a TensorRT engine with int8 quantization and dynamic shape support. It also includes docker files, hyperparameters, and a demo for QAT. It modifies the models package to use pytorch_quantization modules, and disables layer fusion when loading a checkpoint. It adds new files in the trt and trt/trt_dynamic folders for processing, calibrating, evaluating, and visualizing the TensorRT engine.

To make yolov5 more efficient
We added QAT and made it persistent
We used pytorch_quantization
And ONNX export for conversion
To TensorRT engines with int8 inference

Walkthrough

• Add QAT support for yolov5 models and export them to ONNX and TorchScript formats (link, link, link, link, link, link, link, link, link)
• Build and launch docker images for QAT experiments with docker/build.sh and docker/launch.sh (link, link)
• Convert ONNX models with QAT to TensorRT engines with int8 quantization using trt/onnx_to_trt.py and trt/onnx_to_trt_partialquant.py, optionally with a calibrator (link, link, link)
• Process and visualize images using TensorRT engines with trt/Processor.py, trt/Visualizer.py, and trt/demo.py (link, link, link)
• Support dynamic shape for TensorRT engines with trt/trt_dynamic/Processor_dynamic.py and trt/trt_dynamic/eval_yolo_trt_dynamic.py (link, link)
• Add a comment to data/hyps/hyp.qat.yaml to describe its content and usage (link)

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Provide a comprehensive Quantization-Aware Training (QAT) workflow for YOLOv5 in a command-line script.

📊 Key Changes

• Introduce yolo_quant_flow.py with functions for QAT, calibration, and fine-tuning.
• Export trained YOLOv5 models to ONNX format with dynamic axis support.
• Implement data loaders and model preparation for QAT and fine-tuning.
• Evaluate model accuracy throughout the quantization process.

🎯 Purpose & Impact

• Enhances the YOLOv5 workflow to support quantization-aware training (QAT), potentially leading to smaller model sizes and faster inference without significant loss in accuracy.
• The ONNX export capability with dynamic shape support allows for more flexibility and compatibility with different deployment backends.
• The script automates the evaluation of model accuracy pre- and post-quantization, ensuring that the model maintains its performance throughout the process.

[github-actions]

👋 Hello @wjx052333, thank you for submitting a YOLOv5 🚀 PR! To allow your work to be integrated as seamlessly as possible, we advise you to:

• ✅ Verify your PR is up-to-date with ultralytics/yolov5 master branch. If your PR is behind you can update your code by clicking the 'Update branch' button or by running git pull and git merge master locally.
• ✅ Verify all YOLOv5 Continuous Integration (CI) checks are passing.
• ✅ Reduce changes to the absolute minimum required for your bug fix or feature addition. "It is not daily increase but daily decrease, hack away the unessential. The closer to the source, the less wastage there is." — Bruce Lee