YOLOv5

Use the built-in W&B integration in YOLOv5 for experiment tracking, model versioning, and prediction visualization.

Ultralytics’ YOLOv5 (“You Only Look Once”) model family enables real-time object detection with convolutional neural networks.

W&B is directly integrated into YOLOv5, providing experiment metric tracking, model and dataset versioning, rich model prediction visualization, and more. Run a single pip install before you run your YOLO experiments.

All W&B logging features are compatible with data-parallel multi-GPU training, such as with PyTorch DDP.

Track core experiments

To get started, install wandb alongside YOLOv5 and run training as usual. By installing wandb, you activate the built-in W&B logging features: system metrics, model metrics, and media logged to interactive Dashboards.

pip install wandb
git clone https://github.com/ultralytics/yolov5.git
python yolov5/train.py  # train a small network on a small dataset

Follow the links that wandb prints to standard out.

Customize the integration

Once experiment tracking is working, you can enable additional W&B features (such as model versioning, dataset versioning, and prediction visualization) by passing a few command-line arguments to YOLO.

• If you pass a number to --save_period, W&B saves a model version at the end of every save_period epochs. The model version includes the model weights and tags the best-performing model in the validation set.
• Turning on the --upload_dataset flag also uploads the dataset for data versioning.
• Passing a number to --bbox_interval turns on data visualization. At the end of every bbox_interval epochs, W&B uploads the outputs of the model on the validation set.

Model versioning only

python yolov5/train.py --epochs 20 --save_period 1

Model versioning and data visualization

python yolov5/train.py --epochs 20 --save_period 1 \
  --upload_dataset --bbox_interval 1

Every W&B account comes with 100 GB of free storage for datasets and models.

Here’s what that looks like.

With data and model versioning, you can resume paused or crashed experiments from any device, no setup necessary. See the Colab for details.