add support for extremely high dimensional time series forecasting

shane-huang commented 4 years ago

For high demensional time series forecasting. Implementation is based on the paper: https://arxiv.org/abs/1905.03806

Total Task List: ------------- basic ---------

[x] add interfaces in automl and zouwu and UT
[x] update licenses
[x] full training
[x] fixed bugs in cpu-only training
[x] expose arguments such as 'alt_iters'
[x] predict the future steps
[x] model load and restore
[x] rolling prediction without retrain
[x] code clean up (unecessary flags including gpu, unused branches)
[x] memory leak and bug fixes
[x] remove fixed random seeds and put in test ------------- advanced -----------
[x] parallel version of TCMF (multi-model parallel) based orca (without evaluate & incremental training)
[x] parallel version of TCMF (single-model paralle) based on ray
[x] tune training performance ( scalability, convergence, early stop, partitioning) -------------- optional ------------
[x] an example notebook

shane-huang commented 4 years ago

merged pr intel-analytics/analytics-zoo#2406 as the first version. Further changes will be based on the current base code for easier review.

shanyu-sys commented 4 years ago

Predicting the future steps is in PR intel-analytics/analytics-zoo#2460 and has been merged.

intel-analytics / analytics-zoo