v1.0.0

Important

• Package renamed to avex - The package is now available on PyPI as avex (previously representation-learning)
• Repository is now public at https://github.com/earthspecies/avex
• Install with pip install avex or uv add avex

Improvements

• Improve documentation structure and clarity - streamlined README, consolidated detailed docs in docs/index.md
• Add CODEOWNERS for PR approval requirements
• Update GitHub workflows for build and publish with trusted publishing
• Skip tests requiring esp_data to allow standalone package usage

Bug Fixes

• Fix BEATs model: load config instead of default model
• Fix transformers>=5 compatibility for EAT model

Documentation

• Streamlined README.md to focus on quick start and essential info
• Consolidated detailed documentation in docs/index.md
• Updated CONTRIBUTING.md with clearer guidelines
• Added list of official models to documentation

v0.4.3

Release 0.4.3

Important

• All official models have been renamed to use the esp-aves2- prefix naming convention
• All official models are now hosted on Hugging Face in safetensors format https://huggingface.co/collections/EarthSpeciesProject/esp-aves2
• After this release the package and the repo will be renamed to avex and will be made public

Improvements

• Add Hugging Face Hub integration for model hosting and downloading
• Add Safetensor support for model weights
• Add probing API to the library for training and evaluating probes
• Add GCS support for YAML config file loading
• Relax numpy, tensorflow, and torch version requirements for better compatibility
• Add extra_config workaround to allow for outside-of-pydantic configuration
• Update to esp-data 1.5.0

Bug Fixes

• Fix EfficientNet float32 dtype mismatch in process_audio (auto-converts float64 input)
• EfficientNet.forward now accepts optional padding_mask consistent with other models
• Fix BirdNet/birdnetlib tensorflow embedding extraction returning constant embeddings
• Fix classifier head loading issue
• Fix EfficientNet checkpoint loading
• Fix: don't create classification head when return_features_only=True (TensorFlow and PyTorch models)
• Fix beats_naturelm model name reference to official esp_aves2_naturelm_audio_v1_beats

Documentation & Examples

• Shift docs from README.md to docs
• Improve documentation across modules
• Update official models to use safetensors format hosted on Hugging Face
• Fix example_05 to load EAT with return_features_only=True for embedding demo
• Add fully fine-tuning results and plots
• Fix and prune integration tests

---

Acknowledgements

Thanks to all the contributors who made this release possible:

• @benjaminsshoffman
• @david-rx
• @GaganNarula
• @inesnolas
• @mil-ad

v0.2.0

Version 0.2.0

⚙️ Improvements

• return_features_only now returns unpooled features (preserving temporal/spatial information) in all model forward() methods
• Separated API and development dependencies, making the package installable without esp-data for basic API usage
• Created internal IO module to replace esp-data.io dependencies for API-only usage
• Optimized test suite by using class-scoped fixtures to load models once, significantly reducing test execution time
• Added deterministic embedding regression tests for BEATs, EAT-HF, EfficientNet, and AVES models
• Fixed deprecated warnings: replaced torchvision 'pretrained' with weights parameter, updated TypedStorage usage, replaced weight_norm with parametrizations
• Force torch and CUDA initialization to prevent CUDA becoming unavailable during model loading

🐛 Bug Fixes

• Fixed operator precedence issue in test_beats_model.py where assert not sample_audio.grad is not None was changed to assert sample_audio.grad is None
• Fixed model state isolation in test_base_model.py by resetting the model to eval mode after gradient propagation tests using try/finally block
• Added missing .eval() calls in BEATs and EAT-HF deterministic regression tests to ensure models are in evaluation mode
• Fixed BirdNet model shape conversion and tensor handling issues
• Fixed sklearn metrics warnings by adding warning filter for "No positive class found"
• Fixed torch.load weights_only warning by adding explicit weights_only=False parameter

📚 Documentation & Examples

• Completely rewrote installation section in README with clear separation between API Usage and Development Usage
• Updated examples to detect and display 3D/4D output shapes for unpooled features
• Added detailed shape breakdowns and explanatory text about temporal/spatial information preservation

🧪 Testing

• Optimized all model tests (BEATs, AVES, EAT-HF, EfficientNet, base model, probes) by using class-scoped fixtures, reducing test count by ~60% while maintaining coverage
• Consolidated overlapping test files: merged probe, embedding, and projector tests into unified test suites
• Pruned API integration tests from 12 to 3 tests (75% reduction) while maintaining coverage
• Added autouse cleanup fixtures and setup_and_cleanup_hooks pattern for proper test isolation
• Fixed PytestReturnNotNoneWarning by replacing return True/False with pytest.fail()

---

Version 0.1.1

⚙️ Improvements

• Human-readable output for list_models() and describe_model() showing classifier head status
• Simplified model loading API (removed num_classes=None requirement)
• Updated esp-data to 1.3.0 (thanks, Gagan!)
• Renamed load_class_mapping() to load_label_mapping()
• list_models() now always prints table and returns human-readable dictionaries (removed print_table and return_raw parameters)

📚 Documentation & Examples

• New Google Colab demo notebook (examples/colab_sl_beats_demo.ipynb)
• Installation instructions for internal PyPI (esp-pypi) using uv
• YAML configuration examples for custom ModelSpec definitions
• Updated README with current API and file paths
• Updated all example scripts to new API conventions