YOLOv5 is an open-source object detection framework developed by Ultralytics that implements the YOLO (You Only Look Once) architecture in PyTorch with support for export to multiple formats including ONNX, CoreML, and TFLite. The repository represents Ultralytics' research into vision AI methods, incorporating lessons learned from thousands of hours of research and development. The project is written in Python and requires Python 3.7.0 or higher along with PyTorch 1.7 or newer.
The framework provides multiple model variants with different size and performance tradeoffs, including YOLOv5n, YOLOv5s, YOLOv5m, YOLOv5l, and YOLOv5x. Training times on a V100 GPU range from 1 day for the nano variant to 8 days for the extra-large variant. The repository includes pre-trained checkpoints that download automatically from the latest YOLOv5 release, and models are evaluated on the COCO dataset with metrics including [email protected]:0.95 measured on the COCO val2017 dataset. GPU speed measurements are taken on AWS p3.2xlarge instances with V100 GPUs at batch size 32.
The codebase supports inference through multiple methods, including PyTorch Hub integration and a detect.py script that runs inference on various sources and saves results to runs/detect. Training functionality includes support for custom datasets, multi-GPU training, and AutoBatch functionality that automatically determines optimal batch sizes. The repository provides tutorials covering training on custom data, test-time augmentation, model ensembling, model pruning and sparsity, hyperparameter evolution, and transfer learning with frozen layers.
YOLOv5 integrates with several external platforms and tools. Roboflow integration allows users to label and export custom datasets directly for training. ClearML provides automatic tracking, visualization, and remote training capabilities. Comet offers model saving, training resumption, and interactive prediction visualization. The Deci platform enables automatic compilation and quantization for improved inference performance. Additional integrations include support for NVIDIA Jetson Nano deployment and logging through ClearML and Comet.
The repository includes comprehensive documentation available at docs.ultralytics.com covering training, testing, and deployment. Export functionality supports TFLite, ONNX, CoreML, and TensorRT formats for deploying models across different platforms and hardware. The project emphasizes ease of use and learning, prioritizing real-world results in object detection tasks. Models are available through multiple cloud platforms including Google Colab, Kaggle, and Paperspace Gradient for accessible experimentation. The repository maintains continuous integration testing through GitHub Actions workflows and provides Docker images for containerized deployment.
