timeseriesAI/tsai
Tsai
tsai is a deep learning library for time series classification, regression, and forecasting. Built on PyTorch and fastai, it provides a framework for assigning labels to sequential data, predicting future values in univariate or multivariate sequences, and training representations on unlabeled data through self-supervised learning.
The library distinguishes itself with specialized temporal engineering and scaling capabilities. It includes tools for cyclical temporal encoding to capture seasonal patterns and online window slicing to process datasets larger than available memory. It also supports multimodal input pipelines that combine static categorical features with dynamic continuous sequences.
The toolkit covers a broad range of preprocessing and evaluation needs, including sliding window segmentation, missing data imputation, and the conversion of tabular dataframes into structured tensors. Model performance is assessed through walk-forward validation and feature importance analysis to ensure temporal consistency.
Features
- Time Series Forecasting - Provides a comprehensive framework for predicting future values in univariate and multivariate time series.
- Forecasting Input Preparation - Provides tools to filter time-range data from tables to generate necessary features and targets for forecasting tasks.
- Self-Supervised Pre-training - Trains model backbones on unlabeled sequential data to initialize weights for downstream predictive tasks.
- Self-Supervised Time Series Learning - Provides frameworks for training models on unlabeled sequences to learn general temporal patterns before fine-tuning for specific tasks.
- Time Series Classification - Implements recurrent, convolutional, and transformer architectures for assigning categorical labels to sequential data.
- Time Series Deep Learning Libraries - Provides a complete deep learning library for time series classification, regression, and forecasting built on PyTorch and fastai.
- Time Series Regression - Estimates continuous numerical values based on univariate or multivariate time series inputs.
- Online Window Slicing - Enables processing of datasets larger than available memory by slicing time series windows directly from disk during the training loop.
- Self-Supervised Learning - Leverages unlabeled sequential data to learn general representations before fine-tuning for specific tasks.
- Time Series Representations - Implements self-supervised learning to train representations on unlabeled sequential data.
- Walk-Forward Validation - Implements walk-forward validation to evaluate predictive performance by sliding training and testing sets forward through time.
- Cyclical Time Feature Encoders - Provides specialized encoders that transform date and time attributes into sine and cosine arrays to capture periodic seasonal patterns.
- Time Series Transformations - Transforms raw sequences into model-ready formats using sliding windows, padding, and panel data organization.
- Sequence Padding Utilities - Implements utilities to standardize input sequence lengths and identify padding to separate structural filler from actual time series data.
- Large Scale Dataset Processing - Processes datasets larger than memory through online window slicing and meta-datasets.
- Multi-Dimensional Sequence Processing - Organizes input data structured by samples, variables, and sequence length using arrays or tabular formats.
- Multimodal Analytical Pipelines - Offers a unified pipeline to process a combination of static categorical features and dynamic continuous sequences.
- Time Series Model Training - Builds and trains deep learning models for classification and regression using sequence-based data.
- Sliding-Window Data Segmentation - Generates overlapping or non-overlapping samples by moving a window across sequential data to create structured training tensors.
- Synthetic Time Series Generation - Generates synthetic training samples using mixing and uncertainty-based techniques to improve time series model generalization.
- Tabular-to-Tensor Pipelines - Provides pipelines to convert tabular dataframes and numpy arrays into structured tensor formats for deep learning.
- Tabular-to-Tensor Sequence Mapping - Transforms tabular dataframes into multi-dimensional arrays organized by sample, variable, and sequence length for model input.
- Temporal Validation - Implements walk-forward validation and importance analysis to ensure temporal consistency in model evaluation.
- Multimodal Data Preprocessing - Processes datasets containing a combination of static, categorical, and continuous features within a single preprocessing pipeline.
- Missing Data Imputation - Fills gaps in sequential datasets using estimation techniques to ensure continuity for downstream modeling.
- Temporal Gap Handling - Inserts missing timestamps into a sequence and adds indicator columns to mark temporal gaps in the data.
- Temporal Data Processing - Includes specialized utilities for sliding window segmentation and missing data imputation for temporal data.
- Time Series Data Normalization - Provides tools for outlier clipping, log transforms, and robust scaling to normalize time series sequences.
- Sliding Window Algorithms - Generates overlapping or non-overlapping samples by moving a fixed-size window across sequential time series data.
- Time Series Analysis - Deep learning library for time series tasks using PyTorch.
- Analysis Toolkits - Deep learning library optimized for time series analysis.
Star history
How this analysis was created: This summary and feature list are AI-generated from collected project material and can contain mistakes. Stars, license and language are imported from GitHub. Inclusion does not mean that we have tested or audited this project. Check the source documentation for any feature you depend on. Learn more on our About page.
