Regression Only Repo

TO DOs

An example plotting notebook, based on functions defined in plotting.py that loads the model, does inference (saving to .npy), and plots.

Conversion of train_models.py to a class structure, so we can instantiate the class for standalone inference.

Introduction

The idea of this repo is to learn a regression model conditioned on detector properties. This can then be combined with a post-hoc optimizer. The followup to this will be the combination of a regression model and a generative model.

Deep Sets Training

The train_models.py is a training script for DeepSets and GNNs (coming soon). It's primary advantage is its ability to train on input data in a permutation invariant way. The GNNs can also directly encode geometric data.

To run train_models.py for the first time, go to the configs directory, and either edit default.yaml, or add your own. In the configuration file, make sure data_dir points to a directory with appropriated data, created using the generate_data repository, ideally in the form of several small ROOT files, each with a few thousand events.

Next, edit already_preprocessed in the config file to False only when running over a dataset for the first time. Afterwards, try running the training:

python train_models.py or python train_models.py --config [config file name]

One may need to limit num_procs and batch_size according to what their computer can handle.