v1.3.0

Improvements

• ONNX model backends: add Perch v2 via ONNX and replace BirdNET’s TFLite / birdnetlib path with
  ONNX Runtime — faster model loading and embedding extraction, no TensorFlow dependency for these models
  (#185)
• Data layer updates: add versioned eBird taxonomy data, restore BirdSet train split support,
  improve backend-compatible sampling and train/val splitting, and add BEANS label-filling support for
  benchmark configs (#186, #180)
• Rename esp-data / esp_data usage to alp-data / alp_data after the package moved public,
  including docs, configs, jobs, imports, and generated requirements (#199)

Bug fixes

• Fix safetensors device handling so torch.device works while integer GPU indices and map_location
  behavior remain valid (#196)
• Fix RLSubsampleTransform with backend-provided datasets by using backend sampling when available
  (#180)

Dependencies

• Move TensorFlow and TensorFlow Hub to the optional avex[tensorflow] extra; Perch v1 and SurfPerch
  now warn users toward perch2 or the TensorFlow extra (#185)
• Move librosa to the optional avex[examples] extra and declare scikit-learn explicitly as a core
  dependency (#185)
• Drop Python 3.10 support; supported Python versions are now >=3.11,<3.14 (#185)
• Replace private esp-data dependency usage with public alp-data (#199)

Full Changelog: v1.2.0...v1.3.0

Acknowledgements

Thanks to everyone who opened issues, reviewed pull requests, and contributed commits that made this
release possible.

Issue reports: @benjaminsshoffman, @aharmax

Code & commits: @nkundiushuti (+Claude,Codex)

Reviews & discussion: @GaganNarula

We are grateful for your time and feedback