🚀 Optimized YOLOv5 for NPU Deployment

📌 Overview

This repository is a fork of Ultralytics' YOLOv5 with added tools for pruning and quantizing the model to run efficiently on an NPU (Neural Processing Unit), specifically the Phytec i.MX8+. Our optimizations allow for significant performance gains while maintaining accuracy.

🔥 Key Features

• Unstructured Post-Training Pruning: Reduce model size while preserving accuracy.
• Static Quantization: Convert model to int8 for NPU compatibility.
• PyTorch & TensorFlow Pipeline: Pruning is done in PyTorch, while TensorFlow Lite handles quantization.
• Massive FPS Boost: Achieves 2-3x more FPS compared to CPU execution.
• Compatible with YOLOv5s and YOLOv5m: Ensuring efficiency on embedded devices.

🏗 Project Workflow

1. Model Selection

   • Chose YOLOv5s for initial experimentation due to its lightweight nature.
   • The final deployment was done on YOLOv5m for a balance between speed and accuracy.

2. Dataset Selection

   • Utilized COCO 2017 dataset (123,287 images, 80 classes, 886,284 labels).
   • Evaluated baseline model performance before applying pruning & quantization.

3. Performance Baseline Evaluation

   • Measured precision, recall, and mean Average Precision (mAP50/mAP50-95) on 10000 validation images.
   • Conducted inference time analysis for single images and video streams.

🚀 Streamlining your model

• In the yolov5 folder, use the PFE_YOLOv5_Dashboard to find all the tools required to streamline your model as you wish.

🚀 Deploying on NPU

• Deployed on Phytec i.MX8+ with TFLite.
• Used VX Delegate for optimized inference.
• Performance: Real-time inference with 2-3x FPS improvement over CPU.

🚀 Results & Future Improvements

✅ Achievements

• Successfully deployed YOLOv5m on the NPU.
• Achieved real-time inference with minimal accuracy loss.
• Pruning & quantization provided optimal speed/accuracy tradeoff.

🔧 Challenges & Future Work

• Detection inconsistencies: Some noise in NPU predictions.
• Bounding box accuracy variations: Potential improvements via fine-tuning.
• Exploring structured pruning for better layer-level compression.
• Investigating YOLOv5L/X models for potential future optimization.

📜 License

This project is licensed under the MIT License.

---

🔥 Supercharge your object detection with optimized YOLOv5 for NPUs! 🚀