New training wizard
The 11 steps that configure a training run, from choosing the task to review and start.
Required permission: DevelopDEVELOPER · OPERATOR · ADMIN
This screen configures a training run in 11 steps and starts it. You enter it with New training (새 학습) on the Training list, and it needs the develop permission (DEVELOPER, OPERATOR, ADMIN). Entering with Train with this dataset (이 데이터셋으로 학습) on a dataset detail preselects ④ Dataset. Entering with Train again with these settings (이 설정으로 다시 학습) on a training detail fills in every value and opens the ⑪ Review step. The experiment name is cleared; a runtime variant that can no longer be chosen is replaced with the default variant, and uploaded starting weights that were deleted or are not ready are replaced with the catalog weights; a notice tells you so.
| # | Area | Description |
|---|---|---|
| ① | Step list | 11 steps. Click back only to steps you have passed. A step the server returned an error for gets a red mark. |
| ② | Current step | Step title and description, input area. |
| ③ | Next (다음) | Enabled once this step's required values are filled in. On the last step it becomes Start training (학습 시작). |
| ④ | Cancel (취소) | Discards your input and returns to the list (asks for confirmation once). |
Previous (이전) goes back one step. Your input is not saved, so leaving for another menu while you are filling in the wizard asks for confirmation ("Leave the training settings? (학습 설정을 두고 나갈까요?)").
Changing an earlier step clears the later steps that depend on it. Changing the task resets the framework, model, classes and hyperparameters; changing the framework resets the model, hyperparameters, starting weights and runtime variant; changing the model resets the hyperparameters and starting weights; changing the dataset resets the classes.
Step by step
① Task
Choose the computer vision task to train. Object Detection is the default. The number at the top right of each card is how many frameworks can run in this tenant. A task with no runnable framework, because no training runtime variant is published, does not appear in the list. Only a task already chosen by cloning stays, greyed out. If no task is left, the notice "There is no task you can train (학습할 수 있는 태스크가 없습니다)" appears.
| Task | Description |
|---|---|
| Object Detection | Finds objects with bounding boxes and attaches class labels. |
| Instance Segmentation | Also predicts a pixel mask per object. For point clouds, it groups the points of each object. |
| Semantic Segmentation | Assigns every pixel (every point, for point clouds) to a category. |
| Classification | Classifies one input into one of a fixed set of classes. |
② Framework
Cards for the frameworks that support the task. Each card shows badges (real-time, pointcloud and so on), the number of models, license, year and additional export formats (for example onnx, tensorrt). The default for Object Detection is YOLO. As with tasks, a framework with no training runtime variant does not appear. If no card is left, the notice "There is no framework you can train (학습할 수 있는 프레임워크가 없습니다)" appears. In that case, publish a variant on the training runtime screen or ask the system administrator for a shared image.
③ Model
Choose a model variant of the framework (versions with different size and accuracy). Use the search box above to find a variant. In the table, Params(M) is the parameter count, mAP is the accuracy on a public benchmark, and pretrained is the catalog checkpoint (file name and size). After you choose a variant, pick the catalog weights (default) or weights you uploaded under Starting weights (시작 가중치) below. Only uploaded weights for the chosen variant that have finished verification are listed. If there are none, a link opens the pretrained weights screen in a new tab.
④ Dataset
Only datasets with files usable for training (images, PLY, tabular) are shown. Each row shows the type, asset count and last update time. A point cloud dataset without the labels the model needs is greyed out, with the reason in orange inside the row. Semantic segmentation, where classes are required, needs class labels; instance segmentation (PointGroup) needs an instance property. Set the labels in the dataset overview's point label reading. Use dataset cache (데이터셋 캐시 사용) below (on by default) skips the download when you train again with the same dataset.
⑤ Classes
Shows the classes collected from the dataset's annotations, with object counts and shape types. For tabular models, the values of the label column are shown as classes. Depending on the model variant, classes are Required (필수) (choose at least one; all compatible classes are preselected), Optional (선택) (choose none to train all) or Not applicable (해당 없음) (there is nothing to choose). Classes with shapes the model cannot handle are greyed out with the reason. Above the list are the selection count (N / M개 선택됨) and the Select all compatible classes (호환 클래스 전체 선택) and Clear selection (선택 해제) buttons. For point cloud instance segmentation, if you choose no class, every object is treated as the single class object and the model learns only to separate objects one by one. In a dataset that assigns a class per instance, if assignment is unfinished, the warning "N unassigned instances (x% of points) are ignored in training (미할당 인스턴스 N개(점 x%)는 학습에서 무시됩니다)" appears. It does not block training.
⑥ Train/Val split
Set the split method (분할 방식), the Train ratio (Train 비율) (1–99%, default 80%) and the Seed (default 0; the same value reproduces the same split). Split methods include Random (ratio-based random) and Stratified (keeps label ratios); only the ones the framework supports are enabled. Datasets and Collections are not available yet and cannot be chosen.
⑦ Hyperparameters
The model variant's default hyperparameters are filled in as YAML. Your edits are saved as they are, and the server validates the allowed keys and ranges when you submit. Leave it empty to train with the framework defaults. Revert with Reset to defaults (기본값으로 재설정). Some frameworks show input fields instead of YAML, with the range and default written under each field; there, View YAML (YAML 보기) lets you edit the YAML directly.
⑧ Evaluation · ⑨ Export
| Step | Item | Description |
|---|---|---|
| ⑧ Evaluation | Run model evaluation | After training, evaluates the checkpoint that will be registered on the validation set (mAP, mIoU and so on). On by default. |
| ⑧ Evaluation | Speed test | Measures inference latency (p50, p95) and throughput at batch size 1 on the training device. Serving may be faster or slower. Off by default. |
| ⑨ Export | PyTorch (.pt) | Always saved. |
| ⑨ Export | ONNX (.onnx) | After training, converts to ONNX and uploads it as a run artifact. Off by default. Cannot be chosen for frameworks that do not support it. |
| ⑨ Export | TensorRT (.engine) | Depends on the GPU and driver environment. Off by default. Cannot be chosen for frameworks that do not support it. |
Evaluation and the speed test run after the model is logged to MLflow. If either fails, the run still completes and the model is kept. Results appear on the Post-processing results (후처리 결과) (post-processing results) card of the training run detail. A stopped run skips both.
⑩ Execution environment
Training runs as a container (Kubernetes Job). Choose the training image under Training runtime variant (학습 런타임 변형). Only variants for the chosen framework and variants without a framework restriction are listed, and a variant with no image (unpublished) cannot be chosen. Leave it empty (automatic) to use the tenant's default variant. Under GPU resources (GPU 자원), decide whether to run on CPU or how many wired GPU slots (shared pool or dedicated reservation) to use. You can choose up to the slots that are free right now. If no GPU is wired to the tenant, a warning appears and training runs on CPU only.
GPU slots are checked once at submission and again just before a queued run starts. If another workload takes the slots while the run waits, the run fails with "Could not secure the GPU (GPU 를 확보하지 못했습니다)". The queue itself comes from the server-wide limit on concurrent training runs (default 1, shared by all tenants), not from GPUs.
⑪ Review and start
| # | Area | Description |
|---|---|---|
| ① | Experiment name (optional) | Runs submitted under the same name are grouped as runs under one MLflow experiment. If left empty, it is generated as {preset or model}-{date-time}, so each run gets a new experiment. |
| ② | Register in model registry (모델 레지스트리 등록) | When on (off by default), registers best.pt as a new model version after training, and you must enter the Registered model name (등록 모델 이름). Each run with the same name adds a version, and the new version enters Staging. |
| ③ | Start training (학습 시작) | Submits. On success, you move to the training run detail. |
The summary below lets you check the values of every step at once.
When the submission is refused
If server validation finds a problem, a "Check the settings (설정을 확인하세요)" notice appears, the wizard returns to the first step with a problem, and that step lists the errors in red. Errors that belong to no particular step (hyperparameter keys, YAML format and so on) are attached to ⑦. Fixing a value clears that step's red mark. Typical cases that send you back:
- A GPU-only framework (Pointcept and others) with no GPU slot assigned (⑩; at least one needed).
- Point cloud semantic segmentation with only one class chosen (⑤; at least two needed).
- A point cloud instance segmentation dataset with no instance property set (④).