-
Notifications
You must be signed in to change notification settings - Fork 107
PaddleTS 1.0.0 Release Note EN
kehuo edited this page Nov 3, 2022
·
13 revisions
- Support time series model for anomaly detection. PaddleTS provides a collection of abilities such as anomaly detection metrics, visualization analysis, third-party library integration, etc. Meanwhile, PaddleTS involves a set of state-of-the-art models for anomaly detection, see below New Models list.
- Support joint training by multiple time series datasets. Developers are now able to build time series models by using a list of time series datasets associated with related identifiers such as device ids.
- Support time series ensemble model.
- Representation models is now able to solve time series forecasting tasks.
- RNN now supports categorical type features and static covariates.
- TSDataset.load_from_dataframe() now requires significant less time than before.
- More time series analysis.
Time series anomaly detection is now availalbe in PaddleTS, below are newly released models:
- registry.baidubce.com/paddlets/paddlets:1.0.0
- registry.baidubce.com/paddlets/paddlets:1.0.0-gpu-paddle2.3.0-cuda11.2-cudnn8
- registry.baidubce.com/paddlets/paddlets:1.0.0-gpu-paddle2.3.1-cuda11.2-cudnn8
- registry.baidubce.com/paddlets/paddlets:1.0.0-gpu-paddle2.3.2-cuda11.2-cudnn8
Legend for changelogs:
Feature: new features that PaddleTS does not support previously.
Optimization: an existing feature now either requires less computation or memory.
API Change: some changes that require you to change your code.
Fix: some features that previously did not work as stated in documents, or for reasonable expectations that should now work.
- Feature Add Seasonality, Auto-correlation Function(ACF), Correlation time series analysis.
- API Change autots.AutoTS::fit() now allows the train_tsdataset and valid_tsdataset parameters to be either one TSDataset or a list of TSDataset instance(s) to support multiple time series joint training.
- Optimization TSDataset.load_from_dataframe() is 25% / 95% faster than previous releases, in the case if the time index of input data is / isn't standard format, respectively.
- API Change TSDataset.load_from_dataframe() and TSDataset.load_from_csv() now accept new group_id parameter to support multiple time series joint training.
- API Change TSDataset.load_from_dataframe() and TSDataset.load_from_csv() delete the old label_cols parameter and add the new label_col parameter to limit that the TSDataset can only have one label column.
- Feature PaddleTS introduces AutoEncoder, VAE, and Anomaly Transformer deep time series models for anomaly detection.
- Feature Third-party model integration module make_ml_model now supports building anomaly detection models based on third-party library pyod.
- Feature RNN now supports categorical type features. It also supports using static covariates as features.
- API Change make_ml_model module changed from models.forecasting.ml.ml_model_wrapper::make_ml_model to models.ml_model_wrapper::make_ml_model.
- API Change paddle_base_impl.PaddleBaseModelImpl::fit() now allows the train_tsdataset and valid_tsdataset parameters to be either one TSDataset or a list of TSDataset instance(s) to support multiple time series joint training.
- Feature PaddleTS now supports time series model ensemble with two types of ensemble forecaster, StackingEnsembleForecaster and WeightingEnsembleForecaster proposed.
- API Change pipeline.Pipeline::fit() now allows the train_tsdataset and.valid_tsdataset parameters to be either one TSDataset or a list of TSDataset instance(s) to support multiple time series joint training.
- API Change The difference, fill, ksigma, lag, statistical, time_feature transformations now delete fit and transform functions, and add fit_one and transform_one functions to perform time series data transformation based on one TSDataset instance.
- Feature utils::plot_anoms() supports plotting time series anomaly features and labels.