Tsai

New Features

- General: Added min requirements for all package dependencies.

- Data:
- Validation: added split visualization (show_plot=True by default).
- Data preprocessing: add option to TSStandardize or TSNormalize by_step.
- Featurize time series: added tsfresh library to allow the creation of features from time series.
- Models:
- Updated ROCKET to speed up feature creation and allow usage of large datasets.
- Added change_model_head utility function to ease the process of changing an instantiated models head.
- conv_lin_3d_head function to allow generation of 3d output tensors. This may be useful for multivariate, multi-horizon direct (non-recursive) time series forecasting, multi-output regression tasks, etc.
- Updated TST (Time series transformer) to allow the use of residual attention (based on He, R., Ravula, A., Kanagal, B., & Ainslie, J. (2020). Realformer: Transformer Likes Informed Attention. arXiv preprint arXiv:2012.11747.)
- provided new functionality to transfer model's weights (useful when using pre-trained models).
- updated build_ts_model to be able to use pretrained model weights.
- Training:
- TSBERT: a new callback has been added to be able to train a model in a self-supervised manner (similar to BERT).
- Tutorial notebooks:
- I've added a new tutorial notebook to demonstrate how to apply TSBERT (self-supervised method for time series).

Bug Fixes

- Data:
- ROCKET: fixed a bug in `create_rocket_features`.

New Features

- Data:
- core: `get_subset_dl` and `get_subset_dls`convenience function have been added.
- data preparation: `SlidingWindow` and `SlidingWindowPanel` functions are now vectorized, and are at least an order of magnitude faster.
- Models:
- `XCM`: An Explainable Convolutional Neural Network for Multivariate Time Series Classification have been added. Official code not released yet. This is a stete-of-the-art time series model that combines Conv1d and Conv2d and has good explainability.
- Training:
- learner: `ts_learner` and `tsimage_learner` convenience functions have been added, as well as a `get_X_preds` methods to facilitate the generation of predictions.

New Features

- Data:
- data preparation: a new `SlidingWindowPanel` function has been added to help prepare the input from panel data. `SlidingWindow` has also been enhanced.
- new preprocessors: TSRobustScaler, TSClipOutliers, TSDiff, TSLog, TSLogReturn
- Models:
- `MLP` and `TCN` (Temporal Convolutional Network) have been added.
- Training:
- Callback: Uncertainty-based data augmentation
- Label-mixing transforms (data augmentation): MixUp1D, CutMix1D callbacks
- Utility functions: build_ts_model, build_tabular_model, get_ts_dls, get_tabular_dls, ts_learner

New Features

- Added support to Pytorch 1.7.

- New tutorial nbs have been added to demonstrate the use of new functionality like:
- Time series __data preparation__
- Intro to __time series regression__
- TS archs comparison
- __TS to image__ classification
- TS classification with __transformers__

New Features

- More ts data transforms have been added, including ts to images.

- New callbacks, like the state of the art noisy_student that will allow you to use unlabeled data.
- New time series, state-of-the-art models are now available like XceptionTime, RNN_FCN (like LSTM_FCN, GRU_FCN), TransformerModel, TST (Transformer), OmniScaleCNN, mWDN (multi-wavelet decomposition network), XResNet1d.
- Some of the models (those finishing with an plus) have additional, experimental functionality (like coordconv, zero_norm, squeeze and excitation, etc).