Releases

1. ▸▾0.3.020 Oct 2026

   StateDetector(backend, roi, states) is now StateDetector(roi, states, *, backend=None) . Without backend= , the detector uses SqueezeNet on ncnn, load

   What's changed

   ⚠️ Breaking changes

   • StateDetector(backend, roi, states) is now StateDetector(roi, states, *, backend=None). Without backend=, the detector uses SqueezeNet on ncnn, loaded once per process and shared.
   • Creating the first detector without backend= loads the model. It blocks on file I/O and can raise BackendError, so create detectors in an executor, not in an event loop.
   • ncnn is a regular dependency (ncnn>=1.0.20261020) and import tarsius always imports it. The ncnn extra is gone, and tarsius[ncnn] only gives a pip warning.
   • NcnnBackend.id is now squeezenet11/2 because fp16 is off. Stored originals are re-extracted on load, so nothing needs to be taught again.

   ✨ New features

   • #8 feat!: ncnn as a regular dependency, StateDetector picks and shares its backend @RaHehl

   🧰 Maintenance

   • #5 chore: drop CodSpeed, strict mypy, manual publish run @RaHehl

   📚 Documentation

   • #7 docs: evaluate history with StateDetector, report unknown and timing @RaHehl

   Full Changelog: 0.2.0...0.3.0

   Open source →

   Additional notes

   0.3.0

   Compare

   Choose a tag to compare

   Open source →
2. ▸▾0.2.019 Oct 2026

   Reference is removed, with Features.reference() , Reference.to_bytes() / from_bytes() and FeaturesError . Classifier.add() takes Features only. Refere

   What's changed

   ⚠️ Breaking changes

   • Reference is removed, with Features.reference(), Reference.to_bytes()/from_bytes() and FeaturesError. Classifier.add() takes Features only. References are re-extracted from stored frames instead.
   • load_image(draft_side=) is now load_image(max_side=), and the output is at most max_side for every format. Features extracted after load_image() change for the same backend id.
   • majority() returns a Result instead of str | None.
   • Blank references and invalid regions raise BlankRegionError and RoiError, both ValueError subclasses. A frame that shows nothing raises BlankFrameError, a BlankRegionError subclass.

   ✨ New features

   • #2 feat: add StateDetector, shared frame normalization and region outline @RaHehl

   🚀 Enhancements

   • #4 ci: test on Python 3.15 and list it as supported @RaHehl

   📚 Documentation

   • #1 docs: add tarsier logo @RaHehl
   • #3 docs: explain the name @RaHehl

   Full Changelog: 0.1.0...0.2.0

3. ▸▾0.1.019 Oct 2026

   Classifier with k-nearest confirmed references per state and explicit unsure and not_ready results

   First release.

   • Classifier with k-nearest confirmed references per state and explicit unsure and not_ready results
   • NcnnBackend: SqueezeNet 1.1 features on ncnn (extra tarsius[ncnn]), model converted from the original release in CI
   • EdgeBackend: numpy-only baseline
   • Regions as fractions or grid cells, reduced JPEG decoding, reference serialization

   ncnn stays an optional extra until an ncnn release no longer requires opencv-python, see the README.

   Open source →

   Additional notes

   0.1.0

   Compare

   Choose a tag to compare

   Open source →

Your coding agent can read these notes before it upgrades. Set up the MCP server →