Model Training Tab¶
Use Model Training to train YOLO pose models from the GUI.
Important scope note¶
The built-in training workflow is pose-oriented.
If you are working with a detection-only model, the usual path is to bring that checkpoint into Inference rather than train it here.
At a glance¶
| Best for | Typical output | Usually next |
|---|---|---|
Training YOLO pose checkpoints from dataset.yaml |
Model files, training measurements, and optional exports | Inference |
Main sections¶
1. Training paths¶
| Field | What it is |
|---|---|
| Dataset YAML Path | The dataset.yaml prepared in Setup |
| Model Save Directory | Where runs, weights, and logs are written |
2. Model and run configuration¶
| Field | What it is |
|---|---|
| Model Variant | Starting pose checkpoint |
| Run Name | Folder name for this training run |
The initial fields are yolo26n-pose.pt and
keypoint_behavior_run1. They are editable starting values, not requirements;
you may type another compatible Ultralytics pose checkpoint or a local model
path.
3. Training essentials¶
Common controls include:
- epochs
- learning rate
- batch size
- image size
Advanced training settings¶
This section is collapsed by default. Open it when you need finer control over:
- optimizer choice
- weight decay
- label smoothing
- early-stop patience
- device override
The device defaults to -1: automatic idle-GPU selection on CUDA/ROCm systems, mps on supported Apple systems, and CPU otherwise. Enter cpu, 0, cuda:1, or a multi-GPU list only when an explicit override is needed.
4. Augmentation settings¶
This section is also collapsed by default. Use it when you want to tune:
- HSV shifts
- rotation, translation, and scale
- flips
- mixup, mosaic, and copy-paste
5. Export and quantization¶
After training, expand the export section to select trained .pt weights, an
export directory, precision options, and a deployment format such as:
- TensorRT engine
- ONNX
- OpenVINO
- TorchScript
- CoreML
INT8 is available only for compatible TensorRT or OpenVINO setups. Export support ultimately depends on the active Ultralytics runtime and the target platform.
Model Registry integration¶
The Training tab works closely with Model Registry.
Typical behavior:
- completed runs can register
best.pt - recently trained models can be reused in
InferenceandWebcam Inference - export targets can be selected from the registry
Practical tips¶
- Start with default pose settings unless you have a reason to tune aggressively.
- Use Dataset QA before long training runs.
- Watch the Log tab during training for the first useful error message if a run fails.