Train YOLO

The node writes this tree into the task’s own working directory before training:

📁 dataset/
├── 📄 data.yaml
├── 📁 images/
│   ├── 📁 train/
│   └── 📁 val/
└── 📁 labels/
    ├── 📁 train/
    └── 📁 val/

Instance segmentation runs write the same tree, with polygon label lines instead of boxes. Both datasets are exported in a single pass so the two splits share one class index space. The tree is rebuilt from the datasets on every run and nothing downstream can read it.

Every run adds a new version to the model, or to Destination model when one is set. Earlier versions stay in the registry and the source datasets aren’t changed. When Model names the model itself rather than a pinned version, it follows the newest ready version, so a rerun that trains into the same model fine-tunes the previous run’s result. Pick a specific version to start every run from the same weights.

The version is created before training starts and receives the latest best and last checkpoints about every two minutes. A run that fails keeps what it uploaded, and its version is marked failed, so it isn’t offered in the model pickers. If uploading fails five times in a row, the node stops training rather than carry on with no way to save the result.

The version holds best.pt and last.pt (intermediate epoch checkpoints are not kept), the Ultralytics results and plots, your hyperparameters as typed, and the full argument set Ultralytics resolved. It also records the class list, the train and validation datasets, the version training started from, the training report, and the last epoch’s mAP50, mAP50-95, precision, and recall. Segmentation runs add mask mAP50 and mask mAP50-95.

When the train dataset is pinned to an ontology, the model learns that ontology’s object classes in its order, and labels of any other class are skipped. Otherwise the classes come from the annotations in the datasets. Tags never become classes.

Detection trains on boxes: polygons, masks, and other shapes are converted to their bounding boxes. Instance segmentation trains on polygons, with masks converted to polygons. Rectangles are skipped, and a run whose datasets hold only rectangles stops with an error.

For each epoch, the node adds up the box, classification, and distribution focal training losses, plus the segmentation loss on segmentation runs. The badge shows the lowest sum as Loss: value with four decimal places. It stays empty if Ultralytics logged no readable loss rows.

The datasets hold no annotated object classes, so there is nothing to train on. The check uses the datasets’ computed statistics before anything is downloaded, and the export repeats it. For instance segmentation the datasets must also hold polygon or mask annotations. Check the dataset contents, then rerun.

Hyperparameters may not be a YAML mapping, or it may contain an invalid Ultralytics argument. Fix the YAML and check its arguments against the pinned Ultralytics configuration reference.

The model’s task type must be object detection or instance segmentation. The version must also have weights to start from: a checkpoint in its asset, or a checkpoint name Ultralytics can download. A catalog checkpoint that contradicts the task type, such as a -seg checkpoint on a detection model, is rejected.

Checkpoint saving or validation may be disabled, or Ultralytics may not have produced the required files. Keep save and val enabled and check the effective training configuration.

Training continues when the report asset cannot be created. Check the run logs and workspace asset permissions.

A version produced by training or registered by upload starts from the checkpoint stored with it, the best one when there are several. A version copied from the catalog hosts no weights of its own, so Ultralytics downloads its checkpoint by name. Either way the new version records the picked version as its starting point.

OvalBee starts with these values in Hyperparameters:

Delete a value to use the pinned Ultralytics default instead.

Hyperparameters must be a YAML mapping. The effective configuration follows this order:

1. Values present in Hyperparameters override Ultralytics 8.4.33 defaults.
2. The selected Model overrides and removes model.
3. The node overrides data, project, name, and exist_ok so the exported dataset and the task’s artifacts directory are used.
4. Other valid Ultralytics training arguments pass through unchanged.

• Configured image: cr.internal.supervisely.com/ovalbee-internal/nodes/yolo:0.0.12.
• Upstream base: ultralytics/ultralytics:8.4.33.
• Added components: OvalBee SDK, TensorBoard, and the Ultralytics TensorBoard integration.
• Exact CUDA, PyTorch, NVIDIA driver, and GPU compatibility depends on the image digest deployed by the worker.

• Evaluate object detection
• Predict YOLO
• Export YOLO
• Ultralytics repository at v8.4.33
• Training mode documentation at v8.4.33
• Configuration and hyperparameter documentation at v8.4.33
• Default configuration at v8.4.33
• YOLOv8 benchmark table at v8.4.33
• YOLOv9 benchmark table at v8.4.33
• YOLO11 benchmark table at v8.4.33
• Ultralytics license at v8.4.33