YOLOv5 Just Got Stronger in v6.1!

Important Updates

Continuing our relentless pursuit of excellence in Vision AI, these are the ground-breaking enhancements that you'll find in YOLOv5 v6.1:

• TensorRT support: Improved integration for TensorFlow, Keras, TFLite, and TF.js model exports using python export.py --include saved_model pb tflite tfjs (#5699 by @imyhxy). This is a significant milestone as NVIDIA's TensorRT is a high-performance deep learning inference optimizer and runtime that delivers low-latency, high-throughput for deep learning applications.
  ‍
• TensorFlow Edge TPU support ⭐ NEW: Introducing the new smaller YOLOv5n (1.9M params) model that stands below YOLOv5s (7.5M params) in complexity, yet shines in its ability to export to a mere 2.1 MB INT8 size. This is particularly ideal for ultralight mobile solutions, bringing powerful machine learning to the very edge of technology (#3630 by @zldrobit).
  ‍
• OpenVINO support: YOLOv5 ONNX models are now compWith OpenVINO, models can now harness the full power of Intel CPUs and integrated GPUs for a versatile array of applications (#6057 by @glenn-jocher).
  ‍
• Export Benchmarks: We have introduced a new benchmarking tool to assess mAP (Mean Average Precision) and speed across all YOLOv5 export formats with python utils/benchmarks.py --weights yolov5s.pt. Currently operating on CPUs, we plan to extend this to include GPU benchmarks in future updates (#6613 by @glenn-jocher).
  ‍
• Hyperparameters: There has been a minor yet crucial adjustment to our hyperparameters - in hyp-scratch-large.yaml the learning rate factor (lrf) has been reduced from 0.2 to 0.1 (#6525 by @glenn-jocher).
  ‍
• Training: The default Learning Rate (LR) scheduler has been updated to one-cycle linear, replacing the previous one-cycle with cosine, for improved training results (#6729 by @glenn-jocher).

Unveiling the full spectrum of our support across different formats, YOLOv5 now officially works with 11 formats, supporting not just export but also inference with detect.py and PyTorch Hub, and validation to profile mAP and speed:

• ✅ PyTorch
• ✅ TorchScript
• ✅ ONNX
• ✅ OpenVINO
• ✅ TensorRT
• ✅ CoreML
• ✅ TensorFlow SavedModel
• ✅ TensorFlow GraphDef
• ✅ TensorFlow Lite
• ✅ TensorFlow Edge TPU
• ✅ TensorFlow.js

Together for Everyone's AI

At Ultralytics, we are driven not merely by the desire to lead but by the passion to participate and contribute to the community. The YOLOv5 family has been instrumental in our journey, supporting us through triumphs and challenges alike. This update is a collective triumph, representing the hard work of 271 PRs from 48 new contributors. We stand committed to our mission of democratizing AI, making it accessible and operational for everyone.

From AI Enthusiasts to the Most Popular Object Detection of 2022

This year, our Ultralytics/YOLOv5 repository has achieved a significant milestone by surpassing Joseph Redmon's pjreddie/darknet YOLOv3 in the total number of GitHub stars, now boasting over 22.4k stars. This is a testament to the trust and enthusiasm of the community, and it motivates us to keep pushing the boundaries of Vision AI. We are deeply honored to carry forward the You Only Look Once legacy.

Visit our YOLOv5 GitHub repository for comprehensive details about the new release and join the vibrant community of YOLO object detection enthusiasts.