PaddleTS 1.1.0 Release Note EN

Release Highlights for PaddleTS 1.1.0

New Models

PaddleTS releases below new models for time series anomaly detection, classification and forecasting:

• UnSupervised Anomaly Detection (USAD)
• Convolutional Neural Network (CNN) for time series classification
• Sample Convolution and Interaction Network (SCINet) for time series forecasting
• Temporal Fusion Transformer (TFT) for time series forecasting

CPU & GPU Capable Docker Images for PaddleTS

• registry.baidubce.com/paddlets/paddlets:1.1.0
• registry.baidubce.com/paddlets/paddlets:1.1.0-gpu-cuda11.2-cudnn8

Pulling latest tag will get 1.1.0 release now.

• registry.baidubce.com/paddlets/paddlets:latest
• registry.baidubce.com/paddlets/paddlets:latest-gpu-cuda11.2-cudnn8

Changelog for PaddleTS 1.1.0

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.

Changelog

paddlets.automl

• Feature search_space_configer.SearchSpaceConfiger::paddlets_default_search_space now supports hyperparameter optimization for Informer and DeepAR models.
• Feature autots.AutoTS::fit now supports customized maximum concurrency by adding max_concurrent_trials parameter, defaults to 1.

paddlets.ensemble

• Feature Ensemble now supports Representation based time series cluster and classification.
• Feature weighting_ensemble.WeightingEnsembleAnomaly now supports anomaly detection.
• API Change Ensemble removes old ensemble.ensemble_forecaster_base.py, ensemble.stacking_ensemble_forecaster.py and ensemble.weighting_ensemble_forecaster.py modules, replaces with new ensemble.base.py, ensemble.stacking_ensemble.py and ensemble.weighting_ensemble.py modules, respectively.

paddlets.metrics

• Optimization We made internal optimization regarding metrics.base.Metric, now it works more stable.
• API Change metric.metric.MetricContainer construct method removes old metric_names parameter, replaces with new metrics parameter.

paddlets.models

• Feature Paddle Inference is now available for PaddleTS time series forecasting and anomaly detection. See documentation to get more details.
• Feature PaddleTS now supports time series classification.
• Feature PaddleTS introduces 4 deep time series models. USAD for anomaly detection, CNN for classification, SCINet and TFT for forecasting.
• Feature All Anomaly Detection DL Models now support adjusting predict results according to ground truth label by adding pred_adjust and pred_adjust_fn parameters, defaults to False and None, respectively.
• Feature AutoEncoder Anomaly Detection Model now supports categorical features by adding embedding_size and pooling parameters in construct method.
• Optimization We made internal optimization regarding sample building to unify sample building logic.
• Optimization We made internal optimization regarding models.utils::check_tsdataset and time series models RNN, AutoEncoder, now they work more stable.
• API Change models.anomaly.dl.utils::smooth_l1_loss_vae removes y_true parameter.

paddlets.pipeline

• Optimization We made internal optimization regarding pipeline.Pipeline::recursive_predict and pipeline.Pipeline::recursive_predict_proba, now they work more stable.

paddlets.xai

• Feature PaddleTS now supports model interpretability. More specifically, model agnostic interpretability provides the ability of explaining how a prediction is generated. Meanwhile, model-specific interpretability provides visualization of explainable results of time series models, in this release, we implemented ante_hoc.tft_exp based on newly released TFT model.