Bigdl2.0 migration plan for ray/automl/orca.automl/chronos

[shanyu-sys]

Overall plan

The migration plan for ray/automl is

• ray will be put into orca.

• automl will be integrate with orca.automl (part with chronos).

And the pip install dependency for chronos will be orca only.

Therefore, the overall migration plan could be divided into two steps to reduce error and avoid confliction, the first step is to clean up the code structure for orca.automl and chronos, note that automl will be removed. And the second step is to migrate zoo.ray, zoo.orca.automl and zoo.chronos to BigDL2.0.

(Ordered) task list

Step 1: clean up code structure on zoo master

Automl

Step 1: Code structure in automl

Move to chronos

• [x] zoo/automl/pipeline/base -> zoo/chronos/pipeline intel-analytics/analytics-zoo#4633
• [x] zoo/automl/regression/base -> zoo/chronos/regression intel-analytics/analytics-zoo#4633
• [x] zoo/automl/recipe/base -> zoo/chronos/config intel-analytics/analytics-zoo#4633

Internal clean up

• [x] zoo/automl/common/metrics -> zoo/automl/metrics intel-analytics/analytics-zoo#4636
• [x] zoo/automl/search/ray_tune_search_engine -> zoo/automl/search/ray_tune intel-analytics/analytics-zoo#4648
• [x] zoo/automl/logger/tensorboardxlogger.py -> zoo/automl/search intel-analytics/analytics-zoo#4650
• [x] zoo/automl/common/utils -> zoo/automl/search, utils.py needs refactor intel-analytics/analytics-zoo#4641
• [x] zoo/automl/common/parameters -> zoo/automl/search intel-analytics/analytics-zoo#4642

Step 2: Move to orca.automl

• [x] zoo/automl/metrics -> zoo/orca/automl intel-analytics/analytics-zoo#4661
• [x] zoo/automl/model -> zoo/orca/automl intel-analytics/analytics-zoo#4667
• [x] zoo/automl/search -> zoo/orca/automl intel-analytics/analytics-zoo#4670
• [x] remove zoo/automl intel-analytics/analytics-zoo#4680

Chronos

• [x] chronos/model/forecast -> chronos/forecaster intel-analytics/analytics-zoo#4651

• [x] chronos/model/anomaly -> chronos/anomaly intel-analytics/analytics-zoo#4669

  Below tasks are part of AutoTSTrainer deprecation (issue intel-analytics/analytics-zoo-internal#109)

• [ ] add deprecated warning in code/docs for AutoTSTrainer and TSPipeline intel-analytics/analytics-zoo#4786

• [ ] chronos/autots/experimental -> chronos/autots intel-analytics/analytics-zoo#4786

• [ ] chronos/autots/forecast.py -> chronos/autots/deprecated intel-analytics/analytics-zoo#4786

• [ ] chronos/pipeline -> chronos/autots/deprecated intel-analytics/analytics-zoo#4786

• [ ] chronos/regression -> chronos/autots/deprecated intel-analytics/analytics-zoo#4786

• [ ] chronos/config -> chronos/autots/deprecated intel-analytics/analytics-zoo#4786

• [ ] chronos/feature -> chronos/autots/deprecated intel-analytics/analytics-zoo#4786

• [ ] chronos/preprocessing -> chronos/autots/deprecated intel-analytics/analytics-zoo#4786

Step 2: Migrate to BigDL 2.0

We will split into small features and migrated to BigDL2.0 separately.

• [x] zoo/ray -> bigdl/orca/ray
• [x] zoo/orca/automl -> bigdl/orca/automl
• [ ] zoo.chronos -> bigdl.chronos (needs to wait till orca has been migrated)

Note that we will not keep git history when migrating to BigDL2.0.

[shanyu-sys]

The code structure for Chronos will be

• autots
  • model
  • autotsestimator.py
  • tspipeline.py
  • deprecated
  • forecast.py
  • pipeline
  • regression
  • config
  • feature
  • preprocessing
• forecast
• anomaly
• data
• model (time series models in chronos.model)
• examples
• use_case

[shanyu-sys]

The code structure for orca.automl will be

• xgboost
• model
  • base_pytorch_model.py
  • pytorch_utils.py
  • base_keras_model.py
• search
  • ray_tune
  • ray_tune_search_engine.py
  • utils.py
  • tensorboardlogger.py
  • parameters.py
  • utils.py
  • base.py
• auto_estimator.py
• hp.py
• metrics.py

And chronos.autots will depend on orca.automl only.

[shane-huang]

For chronos:

• if we deprecate autots/forecast.py and related files, does it mean we remove the old AutoTSTrainer as a whole? Are we confident to let users switch to the new AutoTSEstimator in the next release? @TheaperDeng
• anomaly detectors does not have wrappers as Forecasters, so we don't have an folder for anomaly now. It is okay to move chronos/model/anomaly to chronos/anomaly. Later we may add change point detection and we can create a new folder.
• autotsestimator currently only works for forecasters. That's why we put it in package forecast.autotsestimator in the past. It is okay for me we just move it to the package autots.autotsestimator. Later we may enable autotsesitmator to tune other models besides forecasters.

[TheaperDeng]

For chronos:

• if we deprecate autots/forecast.py and related files, does it mean we remove the old AutoTSTrainer as a whole? Are we confident to let users switch to the new AutoTSEstimator in the next release? @TheaperDeng
• anomaly detectors does not have wrappers as Forecasters, so we don't have an folder for anomaly now. It is okay to move chronos/model/anomaly to chronos/anomaly. Later we may add change point detection and we can create a new folder.
• autotsestimator currently only works for forecasters. That's why we put it in package forecast.autotsestimator in the past. It is okay we just move it to the package autots.autotsestimator.

AutoTSEstimator and TSDataset can handle the features provided by AutoTSTrainer either in the same/a better way. The autotuning of MTNet is the only feature left behind and we will support it when it is requested.

AutoTSTrainer will also be deprecated in API doc, user guide and quick start this week.

[shanyu-sys]

For chronos:

• if we deprecate autots/forecast.py and related files, does it mean we remove the old AutoTSTrainer as a whole? Are we confident to let users switch to the new AutoTSEstimator in the next release? @TheaperDeng
• anomaly detectors does not have wrappers as Forecasters, so we don't have an folder for anomaly now. It is okay to move chronos/model/anomaly to chronos/anomaly. Later we may add change point detection and we can create a new folder.
• autotsestimator currently only works for forecasters. That's why we put it in package forecast.autotsestimator in the past. It is okay for me we just move it to the package autots.autotsestimator. Later we may enable autotsesitmator to tune other models besides forecasters.

chronos/model/anomaly has been moved to chronos/anomaly (PR intel-analytics/analytics-zoo#4669)

I will work with junwei to deprecate AutoTSTrainer and move AutoTSEstimator, TSPipepline out of experimental. (Issue intel-analytics/analytics-zoo-internal#109)

[shane-huang]

We can have a few modifications on the first-level folders. The major entrance of chronos will be 5

• chronos/forecaster - for forecasting
• chronos/detector - for anomaly, changepoint, seasonality dections
  • chronos/detector/anomaly
  • chronos/detector/changepoint
• chronos/simulator - for time series generation and simulations
• chronos/data - for data processing, i.e. TSDataset
• chronos/autots - automated training

[shanyu-sys]

Task list for the modified chronos structure.

• [x] chronos/forecast -> chronos/forecaster
• [x] chronos/anomlay -> chronos/detector/anomaly