open-source-resources.md

DeepMind Open Source Resources

This document provides a comprehensive index of all datasets, code repositories, and resources associated with DeepMind Research projects. This centralizes access to improve reproducibility and research collaboration.

📊 Datasets

Computer Vision & Image Processing

• Perceiver IO - General architecture for structured inputs & outputs
  • Paper: Perceiver IO: A General Architecture for Structured Inputs & Outputs
  • Code: perceiver/
  • Datasets: Multi-modal examples included

Reinforcement Learning

• RL Unplugged - Benchmarks for offline reinforcement learning
  • Paper: RL Unplugged: A Suite of Benchmarks for Offline Reinforcement Learning
  • Code: rl_unplugged/
  • Datasets: Multiple domains (Atari, Control Suite, etc.)

Physics & Simulation

• MeshGraphNets - Learning mesh-based simulation with graph networks

  • Paper: Learning Mesh-Based Simulation with Graph Networks
  • Code: meshgraphnets/
  • Datasets: CFD (cylinder_flow), cloth simulation (flag_simple)
  • Download: bash meshgraphnets/download_dataset.sh <dataset_name> <output_dir>

• Learning to Simulate - Graph network-based physics simulators

  • Paper: Learning to Simulate Complex Physics with Graph Networks
  • Code: learning_to_simulate/
  • Datasets: Various physics simulations (Water, Sand, etc.)

Natural Language Processing

• WikiGraphs - Wikipedia-Knowledge Graph paired dataset
  • Paper: WikiGraphs: A Wikipedia Text - Knowledge Graph Paired Dataset
  • Code: wikigraphs/
  • Datasets: Wikipedia articles with knowledge graphs

Self-Supervised Learning

• BYOL - Bootstrap Your Own Latent
  • Paper: Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning
  • Code: byol/
  • Datasets: ImageNet pre-training pipeline

Graph Neural Networks

• Graph Matching Networks - Learning similarity of graph structured objects
  • Paper: Graph Matching Networks for Learning the Similarity of Graph Structured Objects
  • Code: graph_matching_networks/
  • Datasets: Various graph matching tasks

🧬 Computational Biology

• AlphaFold CASP13 - Protein structure prediction
  • Paper: Improved protein structure prediction using potentials from deep learning
  • Code: alphafold_casp13/
  • Datasets: CASP13 protein structures

🎮 Game AI & Multi-Agent

• OpenSpiel Integration - Collection of environments for research in general reinforcement learning
  • External Repository: github.com/deepmind/open_spiel

🌡️ Climate & Weather

• Nowcasting - Precipitation nowcasting using deep generative models
  • Paper: Skilful precipitation nowcasting using deep generative models of radar
  • Code: nowcasting/
  • Datasets: UK precipitation radar data

📚 How to Use This Index

For Researchers:

1. Find your domain of interest above
2. Follow the paper link to understand the methodology
3. Navigate to the code directory for implementation details
4. Download datasets using provided scripts or links

For Dataset Access:

Most datasets are downloaded programmatically:

# Example: MeshGraphNets CFD data
bash meshgraphnets/download_dataset.sh cylinder_flow ./data/

# Example: RL Unplugged datasets
python -m rl_unplugged.load_dataset --task=cartpole_swingup

For Code Usage:

Each project directory contains:

• README.md - Project-specific documentation
• requirements.txt - Python dependencies
• run.sh or similar - Execution scripts
• Example notebooks where applicable

🔗 External Resources

Official DeepMind Repositories:

• DeepMind Lab - 3D learning environments
• OpenSpiel - Multi-agent reinforcement learning
• dm-haiku - JAX neural network library
• Acme - Reinforcement learning research framework

DeepMind Official Links:

• Research Publications
• Open Source Datasets
• Open Source Code
• Open Source Environments

🐛 Reporting Issues

If you find broken links or missing resources:

1. Check if the resource has moved to a dedicated repository
2. Open an issue in this repository with the "documentation" label
3. Include the specific resource and expected location

📄 License

All code and datasets are subject to their individual licenses as specified in each project directory. Most DeepMind research code is released under the Apache 2.0 License.

---

Last updated: August 2025 This document is maintained by the community. Contributions welcome!