Skip to content

v0.4.0

Latest
Latest

Choose a tag to compare

View all tags
@paulnovello paulnovello released this 12 Jan 15:30
v0.4.0
70ec56a

Release Notes: v0.4.0

OODEEL v0.4.0 - Enhanced Multi-Layer OOD Detection & HuggingFace Integration

Release Date: January 2026

This release introduces major enhancements to OODEEL, including HuggingFace model support, multi-layer scoring capabilities, and significant improvements to feature-based OOD detection methods.

🎯 Major Features

HuggingFace Integration

  • New HFTorchFeatureExtractor: Seamlessly apply OOD detection on HuggingFace models
  • HuggingFace dataset loading: Added load_from_huggingface method for both TensorFlow and PyTorch data handlers
  • Unified API: DictDataset and HuggingFace Dataset API unification with column_names and get_columns_shapes methods
  • Streamlined transforms: Improved transform handling for HuggingFace and torchvision datasets

Multi-Layer Scoring & Aggregators

  • Aggregator system: New Fisher and variance score aggregators for combining multi-layer features
  • Enhanced methods: Multi-layer scoring support for DKNN, Mahalanobis, RMDS, and Gram methods
  • Per-layer improvements: Refactored VIM, SHE, RMDS, Gram, and Mahalanobis for better per-layer fit and score operations

Method Enhancements

  • Input perturbation: Added epsilon/temperature arguments for all methods via FeatureBasedDetector class
  • head_layer_id support: ReAct, ASH, and SCALE now properly handle head_layer_id argument
  • ReAct improvements: Better penultimate layer computation, especially for models with F.avg_pool
  • return_penultimate: New argument for feature extractors to support ReAct quantile computation

🔧 Improvements & Refactoring

Code Organization

  • Refactored _fit_to_dataset and _score_layer into OODBaseDetector for feature-based methods
  • Reorganized methods with clear public API → per-layer → internal helpers structure
  • Removed fit_to_dataset from logit-based methods for cleaner separation
  • Moved scoring code outside of _fit_layer for better modularity

Data Handling

  • Default features post-processing functions for PyTorch and TensorFlow extractors
  • Improved numpy concatenation of features in extractors
  • Better batch processing with concatenation moved outside loops
  • Removed deprecated merge method from data handlers

Performance

  • Numpy concatenation optimizations in ReAct and t-SNE plotting
  • More efficient feature handling in torch and TensorFlow extractors

🐛 Bug Fixes

  • Dependencies: Replaced faiss_cpu with faiss_gpu and added numpy version constraints
  • Operator improvements: Check if already numpy array in operator.convert_to_numpy
  • sklearn compatibility: Check sklearn version using packaging.version.parse
  • SHE fix: Proper reset between two fits in tests
  • Softmax perturbation: Fixed input perturbation for RMDS and SHE
  • Feature references: Fixed features reference in torch feature extractor
  • Plotting: Fixed number of bins in plot_ood_scores
  • Deprecated warnings: Updated t-SNE n_iter → max_iter
  • Type handling: Improved typing and label conversion when numpy_concat is True
  • Segfault: Fixed segfault in GitHub Actions
  • Postproc functions: Resolved postproc_fns conflicts and migration from init to fit

📚 Documentation

  • New tutorial: Comprehensive data_handler tutorial
  • Improved docstrings: Better documentation for ASH, ReAct, and SCALE methods
  • README updates: Updated with test versions and removed deprecated OODDataset references
  • Method documentation: Added sources in Fisher aggregator and reordered methods
  • Notebook updates: Updated both TensorFlow and PyTorch feature-based methods notebooks

🧪 Testing

  • Extended coverage: GitHub Actions now test on more recent TensorFlow and PyTorch versions
  • HuggingFace tests: Comprehensive tests for HFTorchFeatureExtractor
  • Stability improvements: Fixed tests that could fail due to random seed
  • VIM tests: Updated TensorFlow and PyTorch VIM tests for aggregators

🏗️ Build & Maintenance

  • Version bump: 0.3.0 → 0.4.0
  • Pre-commit hooks: Updated for license insertion
  • Dependencies: Updated pylint in pre-commit config
Assets 2
Loading