Quick start¶
- Choose modern architectures, classic architectures or anomaly detection for the task. Use classification to identify classes, detection for ordinary boxes, instance segmentation for masks, and oriented detection for boxes with orientation.
- Prepare the appropriate dataset format. Classic detection and classification need one dataset folder; modern tasks each require their own label format.
- Select model, device and output options. Classic and anomaly tasks use their own independent environments. One of multiple NVIDIA GPUs can be selected; the current Windows implementation does not support joint multi-GPU training of one classic model.
- Use a small batch and few epochs to check paths, labels, GPU memory and outputs before full training. This workflow check does not assess final accuracy.
- After starting, use Console logs for charts and raw output. Charts show trends; raw logs show initialization, validation, export and error details.
- After successful training, inspect the automatically exported model. Classic training and export are separate stages; an export failure retains saved training weights.
- Open an experiment from Training history, or select a model for inference validation. Anomaly inference is on its own page.
A task takes a parameter snapshot when started. Editing controls afterward does not change the running task; make changes before starting the next one.
Closing the application asks for confirmation and defaults to cancel. Confirming exit while a task runs terminates it. Wait until training-file cleanup completes before exiting.
Three runtime families¶
| Architecture | Training method | Device selection |
|---|---|---|
| Four modern architectures | Modern training engine | auto / gpu / cpu; gpu uses GPU 0 |
| Two classic architectures | Independent classic environment | CPU or any one selected NVIDIA GPU |
| Anomaly detection | Independent anomaly detection environment | CPU or any one selected NVIDIA GPU |
After explicitly selecting GPU, a failure must not be interpreted as an automatic switch to CPU. Classic and anomaly tasks report errors when GPU is unavailable; inspect the log and environment.