This release contains several new features to highlight:
- A major new feature is to support multiple time series in one dataset with a new task named "ts_forecast_panel" and a neural network estimator from pytorch-forecast.
- Allow disabling shuffle for custom splitter.
- Allow explicit specification of whether the choices of a hp have an inherent order.
- Allow skipping data transformation to avoid overhead.
- Support AzureML pipeline tuning.
- Allow log file name to be specified in
tune.runand perform logging when ray is used.
There are other improvements for the transformer estimator and bug fixes for config constraints.
What's Changed
- Fixing the issue that FLAML trial number is significantly smaller than Transformers.hyperparameter_search by @liususan091219 in #657
- make test result more stable by @sonichi in #646
- Add pipeline tuner component and dependencies. by @ruizhuanguw in #671
- Skip transform by @jmrichardson in #665
- pull request template by @sonichi in #668
- Update Research.md by @liususan091219 in #672
- Documentation on search space and parallel/sequential tuning by @qingyun-wu in #675
- time series forecasting with panel datasets by @int-chaos in #541
- categorical choice can be ordered or unordered by @sonichi in #677
- Disable shuffle for custom CV by @jmrichardson in #659
- update time series forecast notebook by @int-chaos in #682
- check config constraints for the initial config by @sonichi in #685
- log_file_name in tune.run() by @sonichi in #681
- updating nlp notebook by @liususan091219 in #683
- VW version requirement and documentation on config_constraints vs metric_constraints by @qingyun-wu in #686
New Contributors
- @jmrichardson made their first contribution in #665
Full Changelog: v1.0.9...v1.0.10