Added
- * Auto3DSeg `AutoRunner` added option `self.algos` to manually specify algorithms to run (instead of always downloading a zip)
- * Auto3DSeg `AutoRunner` added input options `train`, `algo_gen`, `analyze` will default to None, indicating to automatically decide whether this step is necessary based on cache (history).
- * Auto3DSeg `AutoRunner` added method `set_analyze_params` to set `DataAnalyser` parameters (e.g. run it on gpu)
- * Auto3DSeg `AutoRunner` added a dry-run mode for hyper-parameter optimization(HPO) module so that users can get the commands to run NNI without actually using them.

(pending)
- * Auto3DSeg `BundleGen` added a new feature that automatically optimizes the parameter of model training configuration and maximize GPU memory utilization according to the devices in the house.

(also pending to merge because of the update of ALGO_HASH)
- * Auto3DSeg `SegResNet` is refracted to be more efficient, modular, with added support on minimal custom component (BYOC). Currently testing performance. The minimal user input config for the hecktor22 dataset would include a BYOC transform.

Fixed
- * Auto3DSeg `AutoRunner` will no longer fail if the users first set their ensemble method to `AlgoEnsembleBestByFold` and then change the `num_fold` by `set_num_fold`
- * Auto3DSeg `DataAnalyzer` now accepts `None` as the label key to indicate there is no label in the command line mode
- * Auto3DSeg `DataAnalyzer` will no longer crash when label image is empty (no foreground voxels are in this case, previously it crashed)
- * Auto3DSeg ensemble module can find the algorithm folders generated by the hyper-parameter optimization module.
- * Auto3DSeg ensemble module currently can use the voting method under sigmoid mode (one-hot encoding).
- * Auto3DSeg ensemble method `AlgoEnsembleBestN` will no longer fail if the number of trained algorithms is less than the pre-defined N best algorithms.


Changed
- * Auto3DSeg `AutoRunner` changed `self.cache` dict to be a single level (since it only has 4 items), instead of a nested dictionary
- * Auto3DSeg `AutoRunner` removed default setting `sigmoid=True` in `set_prediction_params`, since it's somewhat arbitrary
- * Auto3DSeg `AutoRunner` will raise errors if previous steps are not completed (e.g. ensemble requires train step)
- * Auto3DSeg `AutoRunner` changed its defaults to run data analysis on GPU (previously it was CPU)
- * Auto3DSeg `AutoRunner` will save ensemble prediction labels in uint8 (unless specified; previously it was floats)
- * Auto3DSeg `AutoRunner` will print training progress in Console (previously no output was produced at all for a very long time)
- * Auto3DSeg `AutoRunner` will execute training in sorted order (fold0, fold1, fold2..), Previously it was in random order
- * Auto3DSeg `AutoRunner` now picks the best algorithm in each folder (`AlgoEnsembleBestByFold`) to ensemble algorithm by default, instead of finding the best N algorithms (`AlgoEnsembleBestN`)
- * Auto3DSeg `DataAnalyzer` uses GPU to improve speed by running data loading in parallel
- * Auto3DSeg `DataAnalyzer` can automatically fix shape mismatch (instead of crashing) when label is slightly different in size compared to image
- * Auto3DSeg `DataAnalyzer` saves summary stats on the top in `datastats.yaml` (previously it was at the very bottom)
- * Auto3DSeg `DataAnalyzer`'s connected component stastical analysis switched its default to false, since it’s currently not used, and takes a long time to comput)

(also pending to merge because of the update of ALGO_HASH)
- * All default algorithm templates in Auto3DSeg are unified to use epochs to mark the progress of training and validation.