bioacoustic-ai/bacpipe: BacPipe v1.3.4 Release Notes

Résumé fourni par la source

🚀 bacpipe v1.3.4 Release Notes We are excited to announce bacpipe v1.3.4! This release focuses on polishing the API and developer experience, adding BirdNET v3 support, a comprehensive documentation overhaul, a refreshed set of example notebooks, and a large expansion of the test suite to lock in stability. 🔥 Key Highlights 🦜 BirdNET v3 Support A new birdnet_v3 feature extractor brings the BirdNET+ V3.0 model into bacpipe, powered by the BirdNET+_V3.0-preview3.1_Global_11K_Labels.csv label set. It operates at 32 kHz with a default 3 s segment length and produces 1280-dimensional embeddings. 📚 Documentation Overhaul Extensive docstrings and usage examples were added across the public API, the Sphinx documentation build was restructured and updated (including a revamped API reference and Read the Docs configuration), and the README.md was significantly expanded. 📓 Revamped & New Example Notebooks The basic examples were reorganized — the old simple use cases.ipynb was rewritten as simple_use_cases.ipynb, and a brand new using_different_bacpipe_functions.ipynb walks through the individual bacpipe functions. The advanced examples (clustering, extending_an_existing_model, probing_a_model, using_a_custom_model) were updated, and notebook workflows are now covered by automated tests. 🐛 Key Bug Fixes Fixed a perch_v2 bug by removing incorrect audio normalization and using softmax instead of sigmoid. Fixed a classifier prediction leak and a crash when indexing single-bin classes. Better handling of annotation files containing duplicate start/end times. 🍎 Improved Mac (MPS) Support Better support for Apple's Metal Performance Shaders (MPS — CUDA for Mac) so GPU acceleration works more smoothly on Apple silicon. 🧰 Friendlier API & Configuration Handling Improved and more consistent handling of kwargs and settings.yaml configuration, so options cascade correctly through functions and modules. Several API points were re-exposed for more convenient programmatic use. 📊 Better Dashboard & Prediction Visualization Predictions render more cleanly above spectrograms and now include hover information. The dashboard dynamically loads only what is visible, and overview plots were fixed. ⚙️ New / Updated Settings Parameters The following options in your settings.yaml (or via bacpipe.settings) were added or changed: # Minimum annotation duration (in seconds) required for inclusion in ground truth. min_annotation_length: 0 # Maximum number of labels kept per timestamp (reduces prediction file size). max_labels_per_timestamp: 30 # Choose 2D or 3D embedding plots. visualization_dimensions: 2 # Number of top classifier predictions shown above spectrograms. nr_predictions_to_display: 3 # Maximum discrete categories rendered with distinct colors before using a colorbar. max_nr_categories: 55 Other notable changes: model_base_path was renamed from bacpipe/model_checkpoints to bacpipe_model_checkpoints to avoid conflicts with the package name. rm_embedding_on_keyboard_interrupt now defaults to False, so embeddings are kept when a run is interrupted. Metadata labels were renamed from default_label_keys to metadata_label_keys, with week_of_year added as a new key. 📦 Dependency & Installation Changes Requirements were split into requirements_no_tf.txt and requirements_no_cuda.txt for lighter environments that do not need TensorFlow or GPU support. kaleido (>= 1) was added for static plot/picture export. onnxruntime was added to support ONNX-based models (e.g. BirdNET v3 and the ONNX version of Perch v2). ❤️ Acknowledgements & Contributor Thanks A massive thank you to everyone who contributed code, feedback, and models for this release! Justin Chu and the Perch/BirdNET communities for maintaining the ONNX conversions and model checkpoints that power perch_v2 and the new birdnet_v3 support. A warm thank you to all researchers, model creators, testers, and community members providing feedback to make bacpipe better for bioacoustic research!

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.