Skip to content

Latest commit

 

History

History
54 lines (45 loc) · 2.98 KB

CHANGELOG.md

File metadata and controls

54 lines (45 loc) · 2.98 KB
Raw

New in version ...

0.2.2

  • Deleted keras_contrib dependecy as it was a frequent source of installation problems, and was only used peripherally.
  • Deteled scikit-optimize and tqdm as dependencies, as they were only used in the demos/hp_search.py script.
  • Added demos/atlas_specific_usecases/use_trained_model/inference.py which provides a single function for doing inference with a trained model.
  • Added both the master's thesis that this package was made in conjunction with, along with a short overview of the knowledge that came out of that thesis.
  • Added strides as an available hyperparameter.
  • Saved images of models as .png, as .pdf was causing some trouble on some platforms.
  • Made minor convenient changes to utils functions.

0.2.1

  • Updated demos:
    • Included a more realistic use case of training a model.
    • Included a demo showing how to load and use a trained model.
    • Made hp_search.py more memory efficient in that different processes don't have their own copy of the data.
  • Added functionality for multiplying the model output with a scalar variable (see the multiply_output_name argument of deepcalo/utils/load_atlas_data).
  • Added bias correction classes, which use a (1D or 2D) spline to fit the median error of a model.

0.2.0

  • Added network_in_network model
  • Changed name of time_net to gate_net, as multiple types of cell data can be processed using this - this breaks backward compatibility!
  • Added possibility to scan learning rates logarithmically in LRFinder (#1)
  • Made it possible to divide target by a scalar variable (e.g. the total accordion energy, when doing ER) in load_atlas_data
  • Changed naming convention of ECAL layers in load_atlas_data to fit the new data
  • Bugfix when trying to load gate_net weights into TimeDistributed
  • Bugfix when trying to plot the FiLM generator
  • Bugfix when giving tracks to self.cnn_with_upsampling in model_container.py
  • Bugfix in merge_dicts

0.1.5

  • Added 1Cycle learning rate schedule and improved docs for learning rate schedules in general
  • Bugfix in SGDR_lr_schedule.py (missing imports)
  • Bugfix in load_atlas_data (targets were divided by 1000)
  • Changed the way GPUs were counted in the demos, as the old code was wrong if more than 9 GPUs were used

0.1.4

  • Updated load_data to load_atlas_data, which now works with the newly uploaded data
  • Added custom model checkpoint callback that allows models to be saved as jsons
  • Bugfix in get_track_net
  • Bugfix in datagenerator

0.1.3

  • Bugfix in utils.set_auto_lr
  • Save models instead of just weights

0.1.2

  • Deleted unneeded Python path insertion

0.1.1

  • Deleted deepcalo.utils.apply_preprocessing, such that there is no dependency on scikit-learn

0.1.0

  • Initial release