Track and debug your YOLOv5 models

Introducing automatic bounding box debugging, system metrics, model performance metrics in the cloud, and shareable and reproducible model training for YOLOv5, using Weights & Biases.
Ayush Chaurasia

YOLOv5 now comes with a native Weights & Biases integration that tracks your model pipelines – including model performance, hyperparameters, GPU usage, predictions, and datasets.

Getting Started

Your model metrics are automatically tracked by YOLOv5 if you have wandb installed.

python train.py  --batch 16 --epochs 500 --data coco128.yaml --cfg yolov5s.yaml

Utilities you get out-of-the-box with W&B

1. The Dashboard

Each W&B project has a dashboard that contains information about all the experiments in that project. Here's an example dashboard of a YOLOv5 project. 6fBE0hz.png You can sort, filter, group runs based on any of the logged metrics or configuration parameters.

2. Bounding Box Debugging

Judging object detection models manually is painful. Are you tired of running inferences on model checkpoints at various stages of training? With YOLOv5, you get an interactive bounding box debugging plot where you can play around with confidence parameters to choose the optimal model and thresholds. Here's an example: 4d482f8f.gif

Try the bounding box debugging panel on Pascal VOC dataset below:

Section 2

3. Interactive model metrics saved in the cloud

Section 4

4. System Usage Statistics

Section 6

5. Track configurations useful for reproducibility

Screenshot (151).png

6. Share your model insights with the world

You can write a W&B Report(like this one) about your projects and share it with the world.