v0.8.0

In this release, we add two nlp tasks: sequence classification and sequence regression to flaml.AutoML, using transformer-based neural networks. Previously the nlp module was detached from flaml.AutoML with a separate API. We redesigned the API such that the nlp tasks can be accessed from the same API as other tasks, and adding more nlp tasks in future would be easy. Thanks for the hard work @liususan091219 !

We've also continued to make more performance & feature improvements. Examples:

• We added a variation of XGBoost search space which uses limited max_depth. It includes the default configuration from XGBoost library. The new search space leads to significantly better performance for some regression datasets.
• We allow arguments for flaml.AutoML to be passed to the constructor. This enables multioutput regression by combining sklearn's MultioutputRegressor and flaml's AutoML.
• We made more memory optimization, while allowing users to keep the best model per estimator in memory through the "model_history" option.

What's Changed

• Unify regression and classification for XGBoost by @sonichi in #276
• when max_iter=1, skip search only if retrain_final by @sonichi in #280
• example update by @sonichi in #281
• Merge exp into flaml by @liususan091219 in #210
• add best_loss_per_estimator by @qingyun-wu in #286
• model_history -> save_best_model_per_estimator by @sonichi in #283
• datetime feature engineering by @sonichi in #285
• add warmstart test by @qingyun-wu in #298
• empty search space by @sonichi in #295
• multioutput regression by @sonichi in #292
• add max_depth to xgboost search space by @sonichi in #282
• custom metric function clarification by @sonichi in #300
• checkpoint naming in nonray mode, fix ray mode, delete checkpoints in nonray mode by @liususan091219 in #293

Full Changelog: v0.7.1...v0.8.0

v0.7.1

What's Changed

• make default verbose level > 0 when using ray by @sonichi in #272
• default to cfo for single estimator by @sonichi in #273
• update docstr by @sonichi and @qingyun-wu in #274
• fixed a bug in #278 by @sonichi in #274

Full Changelog: v0.7.0...v0.7.1

v0.7.0

New feature: multivariate time series forecasting.

What's Changed

• Fix exception in CFO's _create_condition if all candidate start points didn't return yet by @Yard1 in #263
• Integrate multivariate time series forecasting by @int-chaos in #254
• Update Dockerfile by @wuchihsu in #269
• limit time and memory consumption by @sonichi in #264

New Contributors

• @wuchihsu made their first contribution in #269

Full Changelog: v0.6.9...v0.7.0

v0.6.9

What's Changed

• no search when max_iter < 2 by @sonichi in #253
• fix bug in subspace identification by @sonichi in #257
• Add documentation for warm-start by @qingyun-wu in #255

Full Changelog: v0.6.8...v0.6.9

v0.6.8

What's Changed

• fix the bug in hierarchical search space (#248); make dependency on lgbm and xgboost optional (#252) by @sonichi in #250
• Add conda forge badge by @MichalChromcak in #251

New Contributors

• @MichalChromcak made their first contribution in #251

Full Changelog: v0.6.7...v0.6.8

v0.6.7

What's Changed

• remove big objects after fit by @sonichi in #176
• remove catboost training dir by @sonichi in #178
• Forecast v2 by @int-chaos in #182
• Fix decide_split_type bug. by @gianpdomiziani in #184
• Cleanml by @qingyun-wu in #185
• warmstart blendsearch by @sonichi in #186
• variable name by @sonichi in #187
• notebook example by @sonichi in #189
• make flaml work without catboost by @sonichi in #197
• package name in setup by @sonichi in #198
• clean up forecast notebook by @sonichi in #202
• consider num_samples in bs thread priority by @sonichi in #207
• accommodate nni usage pattern by @sonichi in #209
• random search by @sonichi in #213
• add consistency test by @qingyun-wu in #216
• set converge flag when no trial can be sampled by @sonichi in #217
• seed for hpo method by @sonichi in #224
• update config if n_estimators is modified by @sonichi in #225
• warning -> info for low cost partial config by @sonichi in #231
• Consistent California by @cdeil in #245
• Package by @sonichi in #244

New Contributors

• @cdeil made their first contribution in #245

Full Changelog: v0.6.0...v0.6.7

v0.6.6

What's Changed

• remove big objects after fit by @sonichi in #176
• remove catboost training dir by @sonichi in #178
• Forecast v2 by @int-chaos in #182
• Fix decide_split_type bug. by @gianpdomiziani in #184
• Cleanml by @qingyun-wu in #185
• warmstart blendsearch by @sonichi in #186
• variable name by @sonichi in #187
• notebook example by @sonichi in #189
• make flaml work without catboost by @sonichi in #197
• package name in setup by @sonichi in #198
• clean up forecast notebook by @sonichi in #202
• consider num_samples in bs thread priority by @sonichi in #207
• accommodate nni usage pattern by @sonichi in #209
• random search by @sonichi in #213
• add consistency test by @qingyun-wu in #216
• set converge flag when no trial can be sampled by @sonichi in #217
• seed for hpo method by @sonichi in #224
• update config if n_estimators is modified by @sonichi in #225
• warning -> info for low cost partial config by @sonichi in #231

Full Changelog: v0.6.0...v0.6.6

v0.6.0

In this release, we added support for time series forecasting task and NLP model fine tuning. Also, we have made a large number of feature & performance improvements.

• data split by 'time' for time-ordered data, and by 'group' for grouped data.
• support parallel trials and random search in AutoML.fit() API.
• support warm-start in AutoML.fit() by using previously found start points.
• support constraints on training/prediction time per model.
• new optimization metric: ROC_AUC for multi-class classification, MAPE for time series forecasting.
• utility functions for getting normalized confusion matrices and multi-class ROC or precision-recall curves.
• automatically retrain models after search by default; options to disable retraining or enforce time limit.
• CFO supports hierarchical search space and uses points_to_evaluate more effectively.
• variation of CFO optimized for unordered categorical hps.
• BlendSearch improved for better performance in parallel setting.
• memory overhead optimization.
• search space improvements for random forest and lightgbm.
• make stacking ensemble work for categorical features.
• python 3.9 support.
• experimental support for automated fine-tuning of transformer models from huggingface.
• experimental support for time series forecasting.
• warnings to suggest increasing time budget, and warning to inform users there is no performance improvement for a long time.

Minor updates

• make log file name optional.
• notebook for time series forecasting.
• notebook for using AutoML in sklearn pipeline.
• bug fix when training_function returns a value.
• support fixed random seeds to improve reproducibility.
• code coverage improvement.
• exclusive upper bounds for hyperparameter type randint and lograndint.
• experimental features in BlendSearch.
• documentation improvement.
• bug fixes for multiple logged metrics in cv.
• adjust epsilon when time per trial is very fast.

Contributors

• @sonichi
• @qingyun-wu
• @int-chaos
• @liususan091219
• @Yard1
• @bnriiitb
• @su2umaru
• @eduardobull
• @sek788432
• @ekzhu
• @anshumandutt
• @yue-msr
• @sadtaf
• @fzanartu
• @dsbyprateekg
• @hanhanwu
• @PardeepRassani
• @gianpdomiziani
• @stepthom
• @anhnht3
• @zzheng93
• @flippercy
• @luizhemelo
• @nabalamu
• @lostmygithubaccount
• @suryajayaraman

v0.5.0

Major update:

• Online automl. For example, we support tuning online machine learning library vowpal wabbit.

Minor updates:

• log best model in mlflow
• utility functions to produce normalized confusion matrix and roc or precision-recall curves for each class in multi-class tasks

v0.4.0

Support for general config constraints and metric constraints in hyperparameter tuning