Skip to content
Snippets Groups Projects
README.md 1.44 KiB
Newer Older
David Mueller's avatar
David Mueller committed
# LGE-CNN: Lattice Gauge Equivariant Convolutional Neural Networks
David Mueller's avatar
David Mueller committed
(formerly known as "cloughmore")
David Mueller's avatar
David Mueller committed

Andreas Ipp's avatar
Andreas Ipp committed
This is the code repository for the paper "Lattice Gauge Equivariant Convolutional Neural Networks" ([arXiv:2012:12901](https://arxiv.org/abs/2012.12901))
David Mueller's avatar
David Mueller committed
by M. Favoni, A. Ipp, D. I. Müller and D. Schuh.  Our code includes a basic SU(2) Yang-Mills code that is used to
generate datasets found in the package `lge_cnn.ym`. The machine learning code (model classes, layers, datasets classes)
is in `lge_cnn.nn`.

## Conda environment

```shell
conda env create -f environment.yml
```

## Generating datasets

Datasets can be generated using the ``generate_dataset.py`` script in the ``scripts`` folder.

To generate the datasets used in the paper, run the shell scripts
```shell
David Mueller's avatar
David Mueller committed
./scripts/generate_2d.sh
./scripts/generate_4d.sh
David Mueller's avatar
David Mueller committed
```

## Training

Training scripts are in the ``scripts`` directory. 

To train all 1+1D models used in the paper, run the shell scripts
```shell
David Mueller's avatar
David Mueller committed
./scripts/train_D2_trW_1.sh
./scripts/train_D2_trW_1_base.sh
./scripts/train_D2_trW_1x2.sh
./scripts/train_D2_trW_1x2_base.sh
./scripts/train_D2_trW_2.sh
./scripts/train_D2_trW_2_base.sh
./scripts/train_D2_trW_4.sh
./scripts/train_D2_trW_4_base.sh
David Mueller's avatar
David Mueller committed
```

For 3+1D models, run
```shell
David Mueller's avatar
David Mueller committed
./scripts/train_D4_trW_2.sh
./scripts/train_D4_trW_4.sh
./scripts/train_D4_QP.sh
David Mueller's avatar
David Mueller committed
```

## Testing

Tests and plot data is generated using the  ``Test all results.ipynb`` Jupyter notebook in the ``notebooks`` directory.