YunKan-Trainer martin888/yunkan-trainer:latest https://hub.docker.com/r/martin888/yunkan-trainer host sh false https://github.com/mrtian2016/yunkan-unraid-templates/issues https://yun-kan.com YunKan-Trainer (云瞰本地自训练) — an **optional NVIDIA-GPU sidecar** for YunKan that fine-tunes the on-device object-detection model on YOUR cameras' own footage, entirely on your LAN. It is a separate, single-purpose container: install it only when you want a detector that is more accurate for the specific people / vehicles / packages at your door, in your lighting. How it works: in the YunKan web admin you confirm or correct the detector's pre-labels (active-learning). When you have enough samples you start a training job — the main YunKan container ships the dataset to this trainer over HTTP, the trainer fine-tunes with ultralytics on your NVIDIA GPU, exports an ONNX model, and hands it back. Nothing leaves your network. **This container is not a full app** — it exposes a small HTTP API on port 28900 (a `/health` endpoint reports GPU / device / busy). It does the training; the main YunKan container drives it. You do not open a normal web UI here. **Requirements:** 1. An **NVIDIA GPU** + the Unraid **"NVIDIA Driver"** plugin (reboot, confirm `nvidia-smi` works on the host). This template already sets `--runtime=nvidia` and the NVIDIA_* env vars — no extra GPU config needed. 2. The **main YunKan** container running on your LAN (any variant). In its **Settings → Training** page, enter this machine's trainer URL `http://<this-host-ip>:28900` and the same **TRAINER_TOKEN** you set below. 3. The image is several GB (it bundles torch + ultralytics + the base weights), so the first pull takes a while. Training is **serial** (one job at a time; the GPU is used exclusively while a job runs). Each job runs in a short-lived subprocess that **fully releases GPU memory and RAM the moment it finishes** — when idle the container just polls and sleeps, using almost nothing, so it can safely share a GPU with other workloads between jobs. No NVIDIA GPU? You do not need this container — you can still fine-tune on **on-demand cloud compute** from within the YunKan admin. **Keywords:** YunKan, 云瞰, self-training, fine-tuning, active learning, custom model, on-device AI, edge AI, object detection, NVIDIA, CUDA, GPU, ultralytics, YOLO, trainer, local training, privacy, self-hosted, NVR AI. ### 0.9.25 (2026-07-09) - **Initial release** — YunKan-Trainer is an optional NVIDIA-GPU sidecar for the main YunKan container. It fine-tunes YunKan's on-device object-detection model on your own cameras' footage, entirely on your LAN: in the YunKan admin you confirm or correct the detector's pre-labels (active learning), then start a training job that this container runs on your GPU and hands back an improved model. Training is serial and fully releases GPU memory and RAM when idle, so it can share a GPU with other workloads between jobs. - **Requirements** — an NVIDIA GPU plus the Unraid "NVIDIA Driver" plugin, and the main YunKan container (any variant) on your LAN. In the main container's Settings / Training page, enter this trainer's URL and the same token you set here. No shared volume needed; datasets and models transfer over HTTP. No NVIDIA GPU? You can fine-tune on on-demand cloud compute from the YunKan admin instead. HomeAutomation: Productivity: http://[IP]:[PORT:28900]/health https://raw.githubusercontent.com/mrtian2016/yunkan-unraid-templates/main/templates/yunkan-trainer.xml https://raw.githubusercontent.com/mrtian2016/yunkan-unraid-templates/main/icon.png --runtime=nvidia --shm-size 1gb Install the Unraid "NVIDIA Driver" plugin so `nvidia-smi` works on the host, otherwise the trainer cannot use the GPU. Also install the main YunKan container (any variant) on your LAN and point its Settings → Training page at this trainer's URL + token. all compute,utility 28900 4096 Etc/UTC 28900 /mnt/user/appdata/yunkan-trainer/work