Tsai

New Features

- compatibility with Pytorch 1.13 ([619](https://github.com/timeseriesAI/tsai/issues/619))

- added sel_vars to get_robustscale_params ([610](https://github.com/timeseriesAI/tsai/issues/610))

- added sel_steps to TSRandom2Value ([607](https://github.com/timeseriesAI/tsai/issues/607))

- new walk forward cross-validation in tsai ([582](https://github.com/timeseriesAI/tsai/issues/582))

Bugs Squashed

- fixed issue when printing an empty dataset wo transforms NoTfmLists ([622](https://github.com/timeseriesAI/tsai/issues/622))

- fixed minor issue in get_robustscaler params with sel_vars ([615](https://github.com/timeseriesAI/tsai/issues/615))

- fixed issue when using tsai in dev with VSCode ([614](https://github.com/timeseriesAI/tsai/issues/614))

- issue when using lists as sel_vars and sel_steps in TSRandom2Value ([612](https://github.com/timeseriesAI/tsai/issues/612))

- fixed issue with feature_importance and step_importance when using metrics ([609](https://github.com/timeseriesAI/tsai/issues/609))

- renamed data processing tfms feature_idxs as sel_vars for consistency ([608](https://github.com/timeseriesAI/tsai/issues/608))

- fixed issue when importing 'GatedTabTransformer' ([536](https://github.com/timeseriesAI/tsai/issues/536))

Breaking Changes

- replaced TSOneHot preprocessor by TSOneHotEncode using a different API ([502](https://github.com/timeseriesAI/tsai/issues/502))

- replaced MultiEmbedding n_embeds, embed_dims and padding_idxs by n_cat_embeds, cat_embed_dims and cat_padding_idxs ([497](https://github.com/timeseriesAI/tsai/issues/497))

New Features

- added GaussianNoise transform ([514](https://github.com/timeseriesAI/tsai/issues/514))

- added TSSequencer model based on Sequencer: Deep LSTM for Image Classification paper ([508](https://github.com/timeseriesAI/tsai/issues/508))

- added TSPosition to be able to pass any steps list that will be concatenated to the input ([504](https://github.com/timeseriesAI/tsai/issues/504))

- added TSPosition preprocessor to allow the concatenation of a custom position sequence ([503](https://github.com/timeseriesAI/tsai/issues/503))

- added TSOneHot class to encode a variable on the fly ([501](https://github.com/timeseriesAI/tsai/issues/501))

- added token_size and tokenizer arguments to tsai ([496](https://github.com/timeseriesAI/tsai/issues/496))

- SmeLU activation function not found ([495](https://github.com/timeseriesAI/tsai/issues/495))

- added example on how to perform inference, partial fit and fine tuning ([491](https://github.com/timeseriesAI/tsai/issues/491))

- added get_time_per_batch and get_dl_percent_per_epoch ([489](https://github.com/timeseriesAI/tsai/issues/489))

- added TSDropVars used to removed batch variables no longer needed ([488](https://github.com/timeseriesAI/tsai/issues/488))

- added SmeLU activation function ([458](https://github.com/timeseriesAI/tsai/issues/458))

- Feature request: gMLP and GatedTabTransformer. ([354](https://github.com/timeseriesAI/tsai/issues/354))

- Pay Attention to MLPs - gMLP ([paper](https://arxiv.org/abs/2105.08050), [implementation](https://github.com/lucidrains/g-mlp-pytorch))

- The GatedTabTransformer ([paper](https://arxiv.org/abs/2201.00199), [implementation](https://github.com/radi-cho/GatedTabTransformer));

Bugs Squashed

- after_batch tfms set to empty Pipeline when using dl.new() ([516](https://github.com/timeseriesAI/tsai/issues/516))

- 00b_How_to_use_numpy_arrays_in_fastai: AttributeError: attribute 'device' of 'torch._C._TensorBase' objects is not writable ([500](https://github.com/timeseriesAI/tsai/issues/500))

- getting regression data returns _check_X() argument error ([430](https://github.com/timeseriesAI/tsai/issues/430))

- I wonder why only 'Nor' is displayed in dls.show_batch(sharvey=True). ([416](https://github.com/timeseriesAI/tsai/issues/416))

New Features

- added StratifiedSampler to handle imbalanced datasets ([479](https://github.com/timeseriesAI/tsai/issues/479))

- added seq_embed_size and seq_embed arguments to TSiT ([476](https://github.com/timeseriesAI/tsai/issues/476))

- added get_idxs_to_keep that can be used to filter indices based on different conditions ([469](https://github.com/timeseriesAI/tsai/issues/469))

- added SmeLU activation function ([458](https://github.com/timeseriesAI/tsai/issues/458))

- added split_in_chunks ([454](https://github.com/timeseriesAI/tsai/issues/454))

- upgraded min Python version to 3.7 ([450](https://github.com/timeseriesAI/tsai/issues/450))

- added sampler argument to NumpyDataLoader and TSDataLoader ([436](https://github.com/timeseriesAI/tsai/issues/436))

- added TSMask2Value transform which supports multiple masks ([431](https://github.com/timeseriesAI/tsai/issues/431))

- added TSRandomStandardize for improved ood generalization ([428](https://github.com/timeseriesAI/tsai/issues/428))

- added get_dir_size function ([421](https://github.com/timeseriesAI/tsai/issues/421))

- Is there any ways of recording the wrong predictions into a txt files? ([397](https://github.com/timeseriesAI/tsai/issues/397))
- I am wondering if the net could record the wrong predictions of the dataset? So maybe I can find a pattern from the wrong files and adjust my method of preprocessing data.

Bugs Squashed

- slow import of MiniRocketMultivariate from sktime ([482](https://github.com/timeseriesAI/tsai/issues/482))

- Installing from source fails on Windows (UnicodeDecodeError) ([470](https://github.com/timeseriesAI/tsai/issues/470))
- Hi everyone,

trying to install the package from source does not work for on Windows 10; it fails with `UnicodeDecodeError: 'charmap' codec can't decode byte 0x8f in position [...]`. The problem also affects other packages (e.g. <https://github.com/iitzco/faced/issues/21>), but is easily solved by specifying an `encoding` in the `setup.py`, see PR.

- TSDataset error oindex is not an attribute ([462](https://github.com/timeseriesAI/tsai/issues/462))

- split_in_chunks incorrectly calculated ([455](https://github.com/timeseriesAI/tsai/issues/455))

- _check_X() got an unexpected keyword argument 'coerce_to_numpy' ([415](https://github.com/timeseriesAI/tsai/issues/415))

New Features

- Added function that pads sequences to same length ([410](https://github.com/timeseriesAI/tsai/issues/410))

- Added TSRandomStandardize preprocessing technique ([396](https://github.com/timeseriesAI/tsai/issues/396))

- New visualization techniques: model's feature importance and step importance ([393](https://github.com/timeseriesAI/tsai/issues/393))

- Allow from tsai.basics import * to speed up loading ([320](https://github.com/timeseriesAI/tsai/issues/320))

Bugs Squashed

- Separate core from non-core dependencies in tsai - pip install tsai[extras]([389](https://github.com/timeseriesAI/tsai/issues/318)). This is an important change that:
- reduces the time to pip install tsai
- avoid errors during installation
- reduces the time to load tsai using from tsai.all import *

Breaking Changes

- updated forward_gaps removing nan_to_num ([331](https://github.com/timeseriesAI/tsai/issues/331))

- TSRobustScaler only applied by_var ([329](https://github.com/timeseriesAI/tsai/issues/329))

- remove add_na arg from TSCategorize ([327](https://github.com/timeseriesAI/tsai/issues/327))

New Features

- added IntraClassCutMix1d ([384](https://github.com/timeseriesAI/tsai/issues/384))

- added learn.calibrate_model method ([379](https://github.com/timeseriesAI/tsai/issues/379))

- added analyze_array function ([378](https://github.com/timeseriesAI/tsai/issues/378))

- Added TSAddNan transform ([376](https://github.com/timeseriesAI/tsai/issues/376))

- added dummify function to create dummy data from original data ([366](https://github.com/timeseriesAI/tsai/issues/366))

- added Locality Self Attention to TSiT ([363](https://github.com/timeseriesAI/tsai/issues/363))

- added sel_vars argument to MVP callback ([349](https://github.com/timeseriesAI/tsai/issues/349))

- added sel_vars argument to TSNan2Value ([348](https://github.com/timeseriesAI/tsai/issues/348))

- added multiclass, weighted FocalLoss ([346](https://github.com/timeseriesAI/tsai/issues/346))

- added TSRollingMean batch transform ([343](https://github.com/timeseriesAI/tsai/issues/343))

- added recall_at_specificity metric ([342](https://github.com/timeseriesAI/tsai/issues/342))

- added train_metrics argument to ts_learner ([341](https://github.com/timeseriesAI/tsai/issues/341))

- added hist to PredictionDynamics for binary classification ([339](https://github.com/timeseriesAI/tsai/issues/339))

- add padding_idxs to MultiEmbedding ([330](https://github.com/timeseriesAI/tsai/issues/330))

Bugs Squashed

- sort_by data may be duplicated in SlidingWindowPanel ([389](https://github.com/timeseriesAI/tsai/issues/389))

- create_script splits the nb name if multiple underscores are used ([385](https://github.com/timeseriesAI/tsai/issues/385))

- added torch functional dependency to plot_calibration_curve ([383](https://github.com/timeseriesAI/tsai/issues/383))

- issue when setting horizon to 0 in SlidingWindow ([382](https://github.com/timeseriesAI/tsai/issues/382))

- replace learn by self in calibrate_model patch ([381](https://github.com/timeseriesAI/tsai/issues/381))

- Argument `d_head` is not used in TSiTPlus ([380](https://github.com/timeseriesAI/tsai/issues/380))
- <https://github.com/timeseriesAI/tsai/blob/6baf0ba2455895b57b54bf06744633b81cdcb2b3/tsai/models/TSiTPlus.py#L63>

- replace default relu activation by gelu in TSiT ([361](https://github.com/timeseriesAI/tsai/issues/361))

- sel_vars and sel_steps in TSDatasets and TSDalaloaders don't work when used simultaneously ([347](https://github.com/timeseriesAI/tsai/issues/347))

- ShowGraph fails when recoder.train_metrics=True ([340](https://github.com/timeseriesAI/tsai/issues/340))

- fixed 'se' always equal to 16 in MLSTM_FCN ([337](https://github.com/timeseriesAI/tsai/issues/337))

- ShowGraph doesn't work well when train_metrics=True ([336](https://github.com/timeseriesAI/tsai/issues/336))

- TSPositionGaps doesn't work on cuda ([333](https://github.com/timeseriesAI/tsai/issues/333))

- XResNet object has no attribute 'backbone' ([332](https://github.com/timeseriesAI/tsai/issues/332))

- import InceptionTimePlus in tsai.learner ([328](https://github.com/timeseriesAI/tsai/issues/328))

- df2Xy: Format correctly without the need to specify sort_by ([324](https://github.com/timeseriesAI/tsai/issues/324))

- bug in MVP code learn.model --> self.learn.model ([323](https://github.com/timeseriesAI/tsai/issues/323))

- Colab install issues: importing the lib takes forever ([315](https://github.com/timeseriesAI/tsai/issues/315))

- Calling learner.feature_importance on larger than memory dataset causes OOM ([310](https://github.com/timeseriesAI/tsai/issues/310))

Breaking Changes

- removed InceptionTSiT, InceptionTSiTPlus, ConvTSiT & ConvTSiTPlus ([276](https://github.com/timeseriesAI/tsai/issues/276))

New Features

- add stateful custom sklearn API type tfms: TSShrinkDataFrame, TSOneHotEncoder, TSCategoricalEncoder ([313](https://github.com/timeseriesAI/tsai/issues/313))

- Pytorch 1.10 compatibility ([311](https://github.com/timeseriesAI/tsai/issues/311))

- ability to pad at the start/ end of sequences and filter results in SlidingWindow ([307](https://github.com/timeseriesAI/tsai/issues/307))

- added bias_init to TSiT ([288](https://github.com/timeseriesAI/tsai/issues/288))

- plot permutation feature importance after a model's been trained ([286](https://github.com/timeseriesAI/tsai/issues/286))

- added separable as an option to MultiConv1d ([285](https://github.com/timeseriesAI/tsai/issues/285))

- Modified TSiTPlus to accept a feature extractor and/or categorical variables ([278](https://github.com/timeseriesAI/tsai/issues/278))

Bugs Squashed

- learn modules takes too long to load ([312](https://github.com/timeseriesAI/tsai/issues/312))

- error in roll2d and roll3d when passing index 2 ([304](https://github.com/timeseriesAI/tsai/issues/304))

- TypeError: unhashable type: 'numpy.ndarray' ([302](https://github.com/timeseriesAI/tsai/issues/302))

- ValueError: only one element tensors can be converted to Python scalars ([300](https://github.com/timeseriesAI/tsai/issues/300))

- unhashable type: 'numpy.ndarray' when using multiclass multistep labels ([298](https://github.com/timeseriesAI/tsai/issues/298))

- incorrect data types in NumpyDatasets subset ([297](https://github.com/timeseriesAI/tsai/issues/297))

- create_future_mask creates a mask in the past ([293](https://github.com/timeseriesAI/tsai/issues/293))

- NameError: name 'X' is not defined in learner.feature_importance ([291](https://github.com/timeseriesAI/tsai/issues/291))

- TSiT test fails on cuda ([287](https://github.com/timeseriesAI/tsai/issues/287))

- MultiConv1d breaks when ni == nf ([284](https://github.com/timeseriesAI/tsai/issues/284))

- WeightedPerSampleLoss reported an error when used with LDS_weights ([281](https://github.com/timeseriesAI/tsai/issues/281))

- pos_encoding transfer weight in TSiT fails ([280](https://github.com/timeseriesAI/tsai/issues/280))

- MultiEmbedding cat_pos and cont_pos are not in state_dict() ([277](https://github.com/timeseriesAI/tsai/issues/277))

- fixed issue with MixedDataLoader ([229](https://github.com/timeseriesAI/tsai/pull/229)), thanks to [Wabinab](https://github.com/Wabinab)