Lightweight framework for streamlined maching learning development for high energy physics, including
- data conversion from ROOT trees to PyTables
- data preprocessing, including input transformation/standardization, dynamic sample reweighting, etc.
- neural network training with Apache MXNet, checkpoint save, model exportation, etc.
This framework has been used for the development of the DeepAK8 tagger in CMS, more details of which can be found in CMS-DP-2017-049 and NIPS 2017 workshop paper.
More details about how to use this framework can be found in the README files of the preprocessing module and the training module.