Skip to content

Latest commit

 

History

History
337 lines (228 loc) · 14.6 KB

CHANGELOG.md

File metadata and controls

337 lines (228 loc) · 14.6 KB

Release notes

0.2.24

Breaking Changes

  • removed InceptionTSiT, InceptionTSiTPlus, ConvTSiT & ConvTSiTPlus (#276)

New Features

  • add stateful custom sklearn API type tfms: TSShrinkDataFrame, TSOneHotEncoder, TSCategoricalEncoder (#313)

  • Pytorch 1.10 compatibility (#311)

  • ability to pad at the start/ end of sequences and filter results in SlidingWindow (#307)

  • added bias_init to TSiT (#288)

  • plot permutation feature importance after a model's been trained (#286)

  • added separable as an option to MultiConv1d (#285)

  • Modified TSiTPlus to accept a feature extractor and/or categorical variables (#278)

Bugs Squashed

  • learn modules takes too long to load (#312)

  • error in roll2d and roll3d when passing index 2 (#304)

  • TypeError: unhashable type: 'numpy.ndarray' (#302)

  • ValueError: only one element tensors can be converted to Python scalars (#300)

  • unhashable type: 'numpy.ndarray' when using multiclass multistep labels (#298)

  • incorrect data types in NumpyDatasets subset (#297)

  • create_future_mask creates a mask in the past (#293)

  • NameError: name 'X' is not defined in learner.feature_importance (#291)

  • TSiT test fails on cuda (#287)

  • MultiConv1d breaks when ni == nf (#284)

  • WeightedPerSampleLoss reported an error when used with LDS_weights (#281)

  • pos_encoding transfer weight in TSiT fails (#280)

  • MultiEmbedding cat_pos and cont_pos are not in state_dict() (#277)

  • fixed issue with MixedDataLoader (#229), thanks to @Wabinab

0.2.23

Breaking Changes

  • removed torch-optimizer dependency (#228)

New Features

  • added option to train MVP on random sequence lengths (#252)

  • added ability to pass an arch name (str) to learner instead of class (#217)

  • created convenience fns create_directory and delete_directory in utils (#213)

  • added option to create random array of given shapes and dtypes (#212)

  • my_setup() print your main system and package versions (#202)

  • added a new tutorial on how to train large datasets using tsai (#199)

  • added a new function to load any file as a module (#196)

Bugs Squashed

  • Loading code just for inference takes too long (#273)

  • Fixed out-of-memory issue with large datasets on disk (#126)

  • AttributeError: module 'torch' has no attribute 'nan_to_num' (#262)

  • Fixed TypeError: unhashable type: 'numpy.ndarray' (#250)

  • Wrong link in paper references (#249)

  • remove default PATH which overwrites custom PATH (#238)

  • Predictions where not properly decoded when using with_decoded. (#237)

  • SettingWithCopyWarning: A value is trying to be set on a copy of a slice from a DataFrame (#221)

  • InceptionTimePlus wasn't imported by TSLearners (#218)

  • get_subset_dl fn is not properly creating a subset dataloader (#211)

  • Bug in WeightedPersSampleLoss (#203)

0.2.19

New Features

  • implemented src_key_padding_mask in TST & TSTPlus (#79)

Bugs Squashed

  • Problem with get_minirocket_features while using CUDA in training (#153)

0.2.19

New Features

  • Models:
    • implement src_key_padding_mask in TST & TSTPlus (#79)

Bugs Squashed

  • Models:
    • Problem with get_minirocket_features while using CUDA in training (#153)

0.2.18

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 natural mask that reuses missing data in the input
    • 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

0.2.17

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.

0.2.16

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

0.2.15

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.

0.2.14

New Features

  • Data:
    • External: added a new function get_Monash_data to get extrinsic regression data.
  • Models:
    • Added show_batch functionality to TSBERT.

0.2.13

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.

0.2.12

New Features

  • Data:
    • core: get_subset_dl and get_subset_dlsconvenience 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.

0.2.8

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

0.2.4

New Features

  • Added support to Pytorch 1.7.

0.2.0

tsai 0.2.0 is a major update to the tsai library. These are the major changes made to the library:

  • 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).