Tsai

New Features

- Models:
- implement src_key_padding_mask in TST & TSTPlus ([79](https://github.com/timeseriesAI/tsai/issues/79))

Bugs Squashed

- Models:
- Problem with get_minirocket_features while using CUDA in training ([153](https://github.com/timeseriesAI/tsai/issues/153))

New features

- Data:
- Update TSStandardize to accept some variables and/or groups of variables when using by_var.
- added option to pad labeled and unlabed datasets with SlidingWindow with a padding value
- added split_idxs and idxs to mixed_dls
- added sklearn preprocessing tfms
- added functions to measure sequence gaps
- added decodes to TSStandardize

- Callbacks:
- change mask return values in MVP to True then mask
- updated MVP to accept nan values
- Models:
- updated mWDN to take either model or arch
- added padding_var to TST
- added MiniRocketFeatures in Pytorch
- Losses & metrics:
- added WeightedPerSampleLoss
- added mean_per_class_accuracy to metrics
- added mape metric
- added HuberLoss and LogCoshLoss
- Learner:
- added Learner.remove_all_cbs
- updated get_X_preds to work with multilabel datasets
- Miscellaneous:
- added rotate_axis utility functions

Bug Fixes

- Callbacks:
- fixed and issue with inconsistency in show_preds in MVP

- Models:
- Fixed an issue in InceptionTimePlus with stochastic depth regularization (stoch_depth parameter)
- Fixed issue with get_X_preds (different predictions when executed multiple times)
- fixed stoch_depth issue in InceptionTimePlus
- fixed kwargs issue in MultiInceptionTimePlus
- Data:
- fixed issue in delta gap normalize
- Learner:
- fixed bug in get_X_preds device
- updated get_X_preds to decode classification and regression outputs

Bug Fixes

- Models:
- Fixed an issue in TST and TSTPlus related to encoder layer creation.
- Fixed issue in TSStandardize when passing tensor with nan values

New features

- Models:
- Added TabTransformer, a state-of-the-art tabular transformer released in Dec 2020.
- TSTPlus now supports padding masks (passed as nan values) by default.

- Data:
- Added a Nan2Value batch transform that removes any nan value in the tensor by zero or median.
- Faster dataloader when suffle == True.
- Added TSUndindowedDataset and TSUnwindowedDatasets, which apply window slicing online to prepare time series data.
- Added TSMetaDataset and TSMetaDatasets, which allow you to use one or multiple X (and y) arrays as input. In this way, you won't need to merge all data into a single array. This will allow you to work with larger than memory datasets.
- Added a new tutorial notebook that demonstrates both multi-class and multi-label classification using tsai.
- Upgraded df2Xy to accept y_func that allows calculation of different types of targets
- Callbacks:
- MVP is now much faster as masks are now created directly as cuda tensors. This has increased speed by 2.5x in some tests.

Breaking changes

- Data:
- train_perc in get_splits has been changed to train_size to allow both floats or integers.
- df2Xy API has been modified

Updates

- Learner:
- Updated 3 new learner APIs: TSClassifier, TSRegressor, TSForecaster.

- ShowGraph callback:
- Callback optionally plots all metrics at the end of training.

Bug Fixes

- Data:
- Updated df2xy function to fix a bug.

Updates

- Tutorial notebooks:
- Updated 04 (regression) to use the recently released Monash, UEA & UCR Time Series Extrinsic Regression Repository (2020).

New features

- Models:
- Added new pooling layers and 3 new heads: attentional_pool_head, universal_pool_head, gwa_pool_head

New Features

- General:
- Added 3 new sklearn-type APIs: TSClassifier, TSRegressor and TSForecaster.

- Data:
- External: added a new function get_forecasting_data to access some forecasting datasets.
- Modified TimeSplitter to also allow passing testing_size.
- Utilities: add a simple function (standardize) to scale any data using splits.
- Preprocessing: added a new class (Preprocess) to be able to preprocess data before creating the datasets/ dataloaders. This is mainly to test different target preprocessing techniques.
- Utils added Nan2Value batch transform to remove any nan values in the dataset.
- Added a new utility function to easy the creation of a single TSDataLoader when no splits are used (for example with unlabeled datasets).
- Added a new function to quickly create empty arrays on disk or in memory (create_empty_array).

- Models:
- TST: Added option to visualize self-attention maps.
- Added 3 new SOTA models: MiniRocketClassifier and MiniRocketRegressor for datasets <10k samples, and MiniRocket (Pytorch) which supports any dataset size.
- Added a simple function to create a naive forecast.
- Added future_mask to TSBERT to be able to train forecasting models.
- Added option to pass any custom mask to TSBERT.

- Training:
- PredictionDynamics callback: allows you to visualize predictions during training.

- Tutorial notebooks:
- New notebook demonstrating the new PredictionDynamics callback.

Bug Fixes

- Models:
- Fixed bug that prevented models to freeze or unfreeze. Now all models that end with Plus can take predefined weights and learn.freeze()/ learn.unfreeze() will work as expected.

New Features

- Data:
- External: added a new function get_Monash_data to get extrinsic regression data.

- Models:
- Added show_batch functionality to TSBERT.