Add and configure a Model
Keep the model file, task, and execution interpretation aligned.
A saved Model holds its file and configuration independently of inference-engine installation. A usable file still needs settings that match its export and Dataset task.
Import and inspect
Use Vault → Models and the import action. Confirm the task, input shape or preprocessing, output interpretation, and classes as applicable. Compare those settings with the model producer’s information.
PyTorch and YOLO checkpoints require ZIP-based serialization. Older weights need to be re-saved with current
torch.saveor exported to ONNX before importing for execution.Configure ONNX outputs
Supported interpretations include classification, tabular classification/regression, decoded rotated boxes, and raw YOLO detection, segmentation, and rotated boxes.
Choose Output interpretation to match the export. Raw Ultralytics exports use letterbox preprocessing; rotated-box outputs include angles in radians. For ambiguous outputs, choose Output tensor name and Mask prototype tensor name where relevant.
Outputs with embedded suppression or different conventions require a matching supported interpretation. Keep interpretation and preprocessing in the Model’s versioned execution settings.
Prepare execution
Complete engine setup and license consent, then select the Model in a Run.