Released Notes
Version 0.1.0
HyperTS is an AutoML and AutoDL tool which focuses on processing time series datasets. It supports the following features:
Support the following data and tasks
Forecasting, classification, and regression
Univariate, multivariates, covariates
Data cleaning:
Repeated columns cleaning
Columns types correction
ID column cleaning
Constant covariate columns cleaning
Deleting covariate columns with missing values
Deleting samples without targets
Data preprocessing:
TimeStamp impution
Missing value simple impution
Missing value average moving impution
Outliers processing
OrdinalEncoder
LabelEncoder
StandardScaler
MinMaxScaler
MaxAbsScaler
Log(x+1)
Data split:
Split training dataset and test dataset in the order of time sequence
Dataset creation
Create batches of inputs and targets by sliding window
Model & mode:
Statistical models: Prophet、ARIMA、VAR、TSForest、KNeighbors
Deep learning models: DeepAR、RNN、GRU、LSTM、LSTNet
Evaluation method:
Train-Validation-Holdout
Visualization:
Human interactive plot
Plot options: historical data, forecasting data and actural data
Confidence intervals
Version 0.1.2.1
Details of the HypertTS update are as follows:
Supports cross validation.
Supports greedy ensemble.
Supports time series forecasting data without timestamp column.
Supports time series forecasting truncation training set to train.
Supports time series forecasting of discrete data (no fixed time frequency).
Supports Fourier inference period.
Supports non-invasive parameters tuning.
Optimizes search space and architecture.
Fixes some bugs.
Version 0.1.3
Details of the HypertTS update are as follows:
Tuning search space hyperparameters;
Added report_best_trial_params;
Fixed ARIMA to be compatible with version 0.12.1 and above;
Fixed the pt issue of LSTNet;
Simplified custom search space, task, timestamp, covariables and metircs can not be passed;
Added OutliersTransformer, supported dynamic handling of outliers;
Adjusted final train processing - lr, batch_size, epcochs and so on;
Added time series meta-feature extractor;
Added Time2Vec, RevIN, etc. layers;
Added N-Beats time series forecasting model;
Added InceptionTime time series classification model;
Supported dynamic downsampling for time series forecasting;
Refactored positive label inference method;
Added neural architecture search mode;
Fixed some known issues.
Version 0.1.4
See Version 0.1.3.
Version 0.2.0
Details of the HypertTS update are as follows:
Supported time series anomaly detection task, and adapt to the full pipeline automation process;
Added IForest anomaly detection model (stats mode);
Added TSOneClassSVM anomaly detection model(stats mode);
Added ConvVAE anomaly detection model(dl mode);
Added realKnownCause anomaly detection dataset;
Supported the visualization of anomaly detection results, and can analyze the anomaly location and severity;
Compatible with Prophet version 1.1.1, now pip install hyperts for simultaneous successful prophet installation;
Compatible with all versions of scipy;
Added API documentation module;
Supported for model persistence (saving and reloading trained models);
In
`model.predict()
, fixed missing value handling;For the time series forecast task, the
`forecast`
function of DL model is calibrated;`DLClassRegressSearchSpace`
was refactored for better adaptation to regression task;Extend
`InceptionTime`
to solve the regression task;Fixed some known issues;
Thanks to @Peter Cotton for his contributions to hyperts.