Skip to content

High performance self-hosted photo and video management solution.

1.1 KB
Raw Blame

machine-learning/AGENTS.md

Python 3.12 FastAPI service running ONNX models for CLIP, facial recognition and OCR, managed with uv.

Rules

  • Install with uv sync --extra cpu; run tools through uv run or the mise tasks.
  • mise //machine-learning:check is mypy --strict, and ruff enforces line length 120. Code must pass both.
  • immich_ml/main.py is the app (/predict, /ping); pipeline.py composes the Clip, FacialRecognition and Ocr pipelines.

Skills

  • Add or change a model: models live in immich_ml/models/{clip,facial_recognition,ocr}/ and extend InferenceModel in models/base.py, which downloads weights with Hugging Face snapshot_download into the cache folder. Loaded models are kept in models/cache.py ModelCache, keyed by class, name and graph, with TTL eviction. Execution backends (ONNX Runtime, RKNN, Arm NN) are in immich_ml/sessions/.
  • Run tests: everything is in test_main.py. One class with uv run pytest test_main.py -k TestCache, one test with test_main.py::TestCLIP::test_name. mise //machine-learning:test-full downloads real models and runs the prediction endpoint tests.