Version 0.1.0

(Release Date: April 17, 2025)

Initial Public Release

This is the first public release of fusionlab, establishing the core framework for building and experimenting with advanced time series forecasting models based on Temporal Fusion Transformer architectures.

Key Features & Modules Included

• Feature Core Forecasting Models (fusionlab.nn.transformers):

  • TemporalFusionTransformer: A flexible implementation of the standard TFT model.

  • XTFT: The Extreme Temporal Fusion Transformer, featuring enhanced components for complex time series, including multi-scale processing, advanced attention, and integrated anomaly detection capabilities.

  • NTemporalFusionTransformer: A variant requiring static/dynamic inputs (point forecast only).

  • SuperXTFT: An experimental variant of XTFT.

• Feature Modular Components (fusionlab.nn.components):

  • Core blocks: GatedResidualNetwork, VariableSelectionNetwork, PositionalEncoding.

  • Sequence processing: MultiScaleLSTM, DynamicTimeWindow, aggregate_multiscale(), aggregate_time_window_output().

  • Attention mechanisms: TemporalAttentionLayer, CrossAttention, HierarchicalAttention, MemoryAugmentedAttention, MultiResolutionAttentionFusion, ExplainableAttention.

  • Input/Output layers: MultiModalEmbedding, LearnedNormalization, MultiDecoder, QuantileDistributionModeling.

• Feature Loss Functions (fusionlab.nn.losses, fusionlab.nn.components):

  • Support for point (MSE) and quantile forecasting (combined_quantile_loss()).

  • Components/factories for combined anomaly objectives: AnomalyLoss, MultiObjectiveLoss, prediction_based_loss(), combined_total_loss().

• Feature Anomaly Detection (fusionlab.nn.anomaly_detection):

  • Initial components: LSTMAutoencoderAnomaly, SequenceAnomalyScoreLayer.

• Feature Hyperparameter Tuning (fusionlab.nn.forecast_tuner):

  • Utilities (xtft_tuner(), tft_tuner()) using keras-tuner.

• Feature Utilities (fusionlab.utils, fusionlab.nn.utils):

  • Time series helpers (ts_utils) for feature engineering, analysis, etc.

  • Neural network helpers (nn.utils) for sequence preparation, forecasting execution, visualization.

• Feature Tools (fusionlab.tools):

  • Initial command-line applications for running workflows.

• Docs Documentation:

  • Initial Sphinx setup: User Guide, Examples, API Reference, Glossary.

Breaking Changes

• Breaking Initial release. No breaking changes from previous versions.

Known Issues / Limitations

• API Change SuperXTFT is experimental and its API may change or be removed.

• Enhancement Backend support is currently focused on TensorFlow/Keras.

• Enhancement Some utility functions might require optional dependencies (e.g., statsmodels, scikit-learn).

Contributors

• earthai-tech (Lead Developer: Laurent Kouadio)