theSearchLife/MantaWatch

Open-source ML pipeline for detecting manta rays in underwater images with automated training, GitHub Actions workflows, and easy-to-use model releases. # MantaWatch Automated image-classification training pipeline. Trigger one GitHub Action and a GPU worker (RunPod Serverless) trains a YOLO11 classifier on images from a Google Drive folder, evaluates it against the previous model, publishes a visual report, and — only if the new model is not worse — releases it. It's generic: the classes come from your dataset's folder names, so the same pipeline retrains on any animal without code changes (currently configured for **manta rays**). ## Companion tool — RunML MantaWatch is the **training half** of a two-part system. The models it releases are used with **RunML** — a desktop tool that runs a released model locally to sort a folder of images into per-class folders. You train a model here, then use its release (`model.onnx`) with RunML to classify new images. ## Documentation - **Setup** — one-time configuration: secrets, variables, the RunPod endpoint, GitHub Pages. - **Usage** — adding images, running training, reading the report, getting the model. - **Architecture** — how the pipeline works under the hood.