Ultralytics Settings

Ultralytics stores persistent, per-machine settings in a JSON file in the user configuration directory and exposes them through the settings object in Python and the yolo settings command. They control where datasets, weights, and runs are stored, the Ultralytics Platform API key, and which experiment trackers are enabled. Per-run arguments such as imgsz and epochs are documented on the Configuration page instead.

View, update, and reset settings

from ultralytics import settings

print(settings)  # view all settings
value = settings["runs_dir"]  # read one setting
settings.update({"runs_dir": "/path/to/runs", "tensorboard": False})  # update one or more settings
settings.reset()  # restore defaults

Settings Reference#

Name	Example Value	Data Type	Description
`settings_version`	`'0.0.8'`	`str`	Settings schema version, distinct from the `ultralytics` package version
`datasets_dir`	`'/path/to/datasets'`	`str`	Directory where datasets are stored
`weights_dir`	`'/path/to/weights'`	`str`	Download directory for URL and Platform (`ul://`) model weights, and fallback lookup for bare model names
`runs_dir`	`'/path/to/runs'`	`str`	Directory where experiment runs are stored
`uuid`	`'a1b2c3d4'`	`str`	Anonymized machine identifier (SHA-256 hash) used for analytics
`sync`	`True`	`bool`	Send anonymized analytics and crash reports to Ultralytics, see Privacy
`api_key`	`''`	`str`	Ultralytics Platform API key
`openai_api_key`	`''`	`str`	OpenAI API key for the Explorer Ask AI feature, available up to `ultralytics==8.3.11`
`clearml`	`True`	`bool`	Enable ClearML logging
`comet`	`True`	`bool`	Enable Comet ML experiment tracking
`dvc`	`True`	`bool`	Enable DVC experiment tracking
`mlflow`	`True`	`bool`	Enable MLflow experiment tracking
`raytune`	`True`	`bool`	Enable Ray Tune hyperparameter tuning reporting
`tensorboard`	`False`	`bool`	Enable TensorBoard logging
`wandb`	`False`	`bool`	Enable Weights & Biases logging
`vscode_msg`	`True`	`bool`	Show a prompt to install the Ultralytics-Snippets extension when a VS Code terminal is detected
`openvino_msg`	`True`	`bool`	On Intel CPUs, show a tip suggesting OpenVINO export for faster inference

A bare model name such as YOLO("yolo26n.pt") loads from the current working directory, then from weights_dir; if neither has it, the weights download to the current working directory. To download them into weights_dir instead, pass the full path, e.g. YOLO(Path(settings["weights_dir"]) / "yolo26n.pt").

Contributors

glenn-jocher¹

Echoslayer¹

raimbekovm¹

Created 2 months agoUpdated 1 week ago

Comments

Ask a question, share your results, or help someone else.