Performance Predictors for MLPF hyperparameter optimization

Predicting the performance of a Graph Neural Network for Particle-Flow reconstruction using (Quantum-)SVR

This project aims to asses the potential of using performance prediction techniques based on https://arxiv.org/abs/1705.10823 whit the goal of optimizing the hyperparameters of Machine Learning Particle Flow (MLPF) .

Practical Info

Use code/QSVR_experiment.ipynb to run a QSVR on multiple train/test splits and/or analize the results of this experiments. Saved results that can be analized without using quantum resources can be found in results/good_qsvr/chain_str_10.
Use code/QSVR_combination.ipynb to train and/or evaluate the combination technique that allows to use more training points on a QSVR. Saved results that can be analized without using quantum resources can be found in results/good_qsvr/combination.
Use code/saved_QSVR_true_vs_predicted_loss.ipynb to load a saved pre-trained QSVR and plot its predicted loss vs true loss for all the reconstruction techines. Some pre-trained QSVRs can be found in results/good_qsvr/chain_str_10.
code/SVR_training.ipynb can be used to train and store a performance predictor to be used later on later optimization scripts as both code/distributed_search.py andc ode/optimize.py need to load pretrained scalers and a regressor.
code/generate_data.py can be used to generate a CSV file of learning curves for a given model. To use a model other than the default one edit code/trial.py.
The plots stored in results/ can be generated with the notebook code/plots.ipynb.

In order to be able to run the scripts that use QSVR class you need to have the code/qa_summer folder content not publicly available here.
Information on how to install and configure ocean (D-wave software) can be found here

"The code in this project is licensed under Apache License 2.0 license."