3. New training (11-step wizard)
Set everything from the task to the runtime in the 11-step wizard and start training
Required permission: DevelopDEVELOPER · OPERATOR · ADMIN
New training collects settings in 11 steps, following the step list on the left, and submits them once, at the end. The Next (다음) button stays disabled until you make a choice in the step. To change a step you have passed, click it in the list on the left.
| # | Step | What you choose |
|---|---|---|
| 1 | Task (태스크) | Object Detection, Instance Segmentation, Classification … |
| 2 | Framework (프레임워크) | A framework that supports the task (for example YOLO, RF-DETR) |
| 3 | Model (모델) | Model variant and Starting weights (시작 가중치) |
| 4 | Datasets (데이터셋) | A dataset with files usable for training |
| 5 | Classes (클래스) | Classes to include in training |
| 6 | Train/Val split (Train/Val 분할) | Split method, train ratio, seed (default Random 80%, seed 0) |
| 7 | Hyperparameters (하이퍼파라미터) | Edit the framework's default YAML (an input form for some frameworks) |
| 8 | Evaluation (평가) | Evaluate on the validation set after training (on by default), speed test |
| 9 | Export (Export) | Extra ONNX and TensorRT conversion |
| 10 | Runtime (실행 환경) | Training runtime variant, number of GPU slots (default 0 = CPU) |
| 11 | Review and start (검토·시작) | Experiment name, model registry registration, submit |
- In the sidebar, open Training (학습) and press New training (새 학습) ①. (You can also come in through Train with this dataset (이 데이터셋으로 학습) on the dataset detail.)
Training list — ① New training - 1. Task (태스크) — you want to find helmets with bounding boxes, so choose Object Detection ① and press Next (다음). Tasks with no ready training runtime do not appear in the list. The same goes for frameworks in step 2.
Step 1, Task — ① Object Detection - In 2. Framework (프레임워크) choose YOLO, and in 3. Model (모델) choose a model variant ①. The Params(M) and mAP in the table are public benchmark values for gauging size and performance. Starting weights (시작 가중치) ② defaults to Catalog pretrained weights (카탈로그 사전학습 가중치); choose your own uploaded checkpoint here.
Step 3, Model — ① model variant ② starting weights - In 4. Dataset (데이터셋), choose
안전모 현장 사진(the helmet field photos dataset) ① that you created on the previous page. Datasets without files usable for training do not appear in the list. If you keep Use dataset cache (데이터셋 캐시 사용) ② on (the default), training again with the same dataset skips the download.Step 4, Dataset — ① dataset to train on ② dataset cache - 5. Classes (클래스) — the classes read from the annotations and their object counts are shown. Use Select all compatible classes (호환 클래스 전체 선택) ① to include both
headandhelmet. Depending on the model variant, classes are required, optional or not applicable.Step 5, Classes — classes and object counts read from the annotations, ① select all compatible classes - 6. Train/Val split (Train/Val 분할) — the default is Random (ratio-based random) at 80%. Stratified keeps the label ratio of the validation set in line with the training set. The same seed reproduces the same split.
Step 6, Train/Val split — method, train ratio and seed - 7. Hyperparameters (하이퍼파라미터) — the framework defaults are filled in as YAML ①. Change only the values you need, such as
epochs,batchandimgsz. The server checks the allowed keys and ranges on submit; if something is wrong, you are sent back to this step with the error. Reset to defaults (기본값으로 재설정) ② restores the defaults.Step 7, Hyperparameters — ① edit as YAML ② reset to defaults - Run model evaluation (모델 평가 실행) in 8. Evaluation (평가) computes mAP and other metrics on the validation set after training and records them in MLflow. Turn on Speed test (속도 테스트) to also measure inference latency and throughput. In 9. Export, the PyTorch weights are always saved; turn on ONNX export (ONNX 변환) and TensorRT export (TensorRT 변환) when you need them (the conversion results are also uploaded as MLflow run artifacts).
- 10. Runtime (실행 환경) — training runs as a Kubernetes Job in the tenant namespace. If you leave Training runtime variant (학습 런타임 변형) ① empty (Auto (default) (자동 (기본))), the tenant's default image is used. Choose the number of GPU slots in GPU resources (GPU 자원) ②. The default is 0, and with 0 it runs on CPU. You can choose up to the slots still free among those wired to the tenant.
Step 10, Runtime — ① training runtime variant ② GPU resources (the warning for a tenant without GPU wiring) - 11. Review and start (검토·시작) — enter Experiment name (optional) (실험 이름 (선택)) ①. Training runs submitted with the same name are grouped as runs under one experiment in MLflow (if you leave it empty, a name is generated and each run becomes a new experiment). Turn on Register in model registry (모델 레지스트리 등록) ② (off by default) and enter Registered model name (등록 모델 이름) (required once it is on), and the model is registered automatically as a new version when training finishes. Look over the summary and press Start training (학습 시작) ③.
Step 11, Review and start — ① experiment name ② register in model registry ③ start training
When you submit, the server checks the settings (hyperparameter keys and ranges, dataset, training runtime image, GPU slots). If they pass, a "Training started (학습 시작됨)" notification appears, you are taken to the training run detail screen, and the training run waits for its turn with status Queued (대기 중) in the Waiting in queue (실행 대기) step. If they are rejected, a notification gives the reason, and a per-step setting error sends you back to that step.