Physics Data for Machine Learning (pd4ml)

This repository contains datasets and model for machine learning from the publication "Shared Data and Algorithms for Deep Learning in Fundamental Physics" (arXiv:2107.00656)

You can install this package as a python module with pip via:

pip install git+https://github.com/erum-data-idt/pd4ml

# or just git clone & 'pip install .' in this folder

The essential function is the load function to load the training and testing datasets. The datasets features "X" are returned as a list of numpy arrays. The labels are returend directly as a numpy array.

from pd4ml import Spinodal   # or any other dataset (see below) 

# loading training data into RAM (downloads dataset first time)
X_train, y_train  = Spinodal.load('train', path='./datasets')

# loading test data into RAM (downloads dataset first time)
X_test, y_test = Spinodal.load('test', path = './datasets')

Here a subfolder ./datasets is created. The datasets take up a total disk space of about 2.4 GB. For loading the training datasets a free RAM of at at least 5 GB is necessary (depending on the dataset).

Included datasets at the moment with the tags:

1: TopTagging, 2: Spinodal, 3: EOSL, 4: Airshower, 5: Belle

An description of the datasets can be printed via the function:

Spinodal.print_description()

Show all available datasets:

import pd4ml

for dataset in pd4ml.Dataset.datasets_register:
    print(dataset.name)

An additionally load_data function performs some basic preprocessing steps as well as allows the return of an adjecancy matrix:

from pd4ml import Spinodal   # or any other dataset
x_train, y_train = Spinodal.load_data('train', path = './datasets', graph = True)

x_train is dictionary with the contents features and adj_matrix. If no adjecancy matrix is required, one may set graph = False.

Some example plots can be found in the notebooks in the example folder.

---

Creating a model:

In the folder models multiple model implementations can be found. Each can be imported in the main.py script and run on the specified datasets. If you'd like to contribute a model, feel free to implement it using the template.py.

---