Graph Decoder
Graph neural network decoder for the rotated surface code.
Includes the source code for the GNN decoder, as well as scripts to run the code on a cluster.
Getting started
Follow these steps to run the code on a cluster.
- Clone the repository with
git clone https://github.com/LangeMoritz/GNN_decoder - Install the required packages in
requirements.txt, for example using a virtual environment:
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt
- Install PyTorch by following the instructions found here.
- Install PyTorch Geometric by following the instructions found here.
Repository structure
src/contains the "source code" for the project, split into a number of modules:GNN_Decoder.pyA class for creating a decoder object with methods for training a GNN decoder and running decoding simulations.__init__.pyturnssrc/into an importable package.gnn_models.pyContains PyTorch geometric graph neural network decoder models.graph_representation.pyFunctions for converting syndromes to graphs.
├── src
│ ├── GNN_Decoder.py
│ ├── __init__.py
│ ├── gnn_models.py
│ ├── graph_representation.py
│ └── rotated_surface_code.py
models/contains trained models corresponding to fig. 3 (circuit_level_noise/), fig. 5 and 6 (repetition_code/) and fig. 7 (perfect_stabilizers/)results/contains model and training history from training runs as .pt files (each run generates one file upon finishing). This directory should exist in your working directory.job_outputs/contains the standard output files from runs. This directory should exist in your working directory.buffer_training.pyis the python script used to run training with a data buffer, replacing part of the buffer with new data after a fixed number of training iterations.run_buffer_training.shis the shell run script used to start is jobs forbuffer_training.py.
.gitignorelists files and directories in the git repository to be ignored in commits.requirements.txtlists the required python packages. See the Getting Started section above.