FAQ
What common messages on the training platform mean and what to do about them, plus a few things to think through before you click.
FAQ
| What you see | Meaning | What to do |
|---|---|---|
| "SSO 登录失败" (SSO sign-in failed), "票据已失效" (The ticket has expired) | The link was already used | Go back to the admin console and click 打开训练平台 (Open training platform) again |
| The operation's project is not in the project list | Not submitted for training yet, or it belongs to another company | In the admin console, click 打开训练平台 (Open training platform) → 送训并打开 (Submit for training and open) |
| Submitted for training, but there is no teach video in the footage | The video is large and still uploading in the background | Wait a few minutes and refresh. If it never appears, upload the teaching video manually |
| "同名素材已存在…" (Footage with the same name already exists…) | A file with the same name was already uploaded | No need to upload again. If you really want to replace it, delete the old one first |
| "零件视频(part)必须填写零件名" (A part video (part) must have a part name) | You chose part video but did not choose a class | Choose a class, or choose "混合类别" (Mixed classes) |
| "混合类别素材缺参考帧" (Mixed-class footage is missing reference frames) | Some parts have no separate footage | Add a separate video of that part, or click 仍然标注 (Label anyway) |
| "该帧标注已被其他人改动…" (The labels on this frame were changed by someone else…) | Two people edited the same frame at the same time | Refresh, then make your changes again |
| "GPU 正忙(训练进行中)，点击出框请稍后再试" (GPU is busy (training in progress). Try click-to-box again later) | Training is using the graphics card | Wait for training to finish before using E for click-to-box, or draw boxes by hand |
| "点击处未分割出目标，请换个位置再点" (No object was found where you clicked. Click somewhere else) | The AI did not recognize anything where you clicked | Click in the center of the part, or draw the box by hand |
| Training failed after about 1 second | The dataset is empty (0 / 0) | Extract frames, label and review first |
| "缺少校准图，无法量化" (Missing calibration images, cannot quantize) | Fewer than 10 reviewed images | Review more before quantizing, or just use FP16 |
| "无 rknn 产物，不能发布" (No rknn output, cannot release) | This version was not exported successfully | Export again, or use the pipeline's 只导出+发布 (Export + release only) |
| You cannot see the 发正式版 (Release stable) button | You do not have release permission | Ask your company's admin |
| "原视频不存在(超过 30 天已自动删除…)" (The original video does not exist (deleted automatically after 30 days…)) | The hard sample's original video has expired | Just look at the captured images |
| "积分不足" (Not enough credits) | Your company's billing account balance is too low | Contact your account manager to top up |
| mAP is very high, but the workstation still does not recognize parts accurately | The validation set is too similar to the training set, or the lighting and placement on site differ from the footage | Record more footage on site and handle more hard samples |
Notes
- Review matters more than quantity: 1,000 reviewed images are more useful than 5,000 AI boxes nobody has checked.
- Change classes in the admin console: on the training platform, you cannot change classes for projects sent from the platform. If you change classes (especially label names), submit for training again and train again.
- Release a new model as a test version first, verify it on the workstation, then release the stable release. Once a stable release goes out, all workstations switch.
- Roll back to an old version: in the model list, click 发正式版 (Release stable) on the old version, or in "训练版本" (Training versions) in the admin console, click 使用此版本 (Use this version) on the old version.
These actions lose data
全量重抽 (Re-extract all), 全量重标 (Re-label all), 删帧 (Delete frames), deleting footage, and deleting a project (彻底删除 (Delete permanently)) all lose data. Think before you click. Re-labeling everything does not lose boxes labeled by people, but deleting frames or footage deletes the boxes along with them.