---

name: graph-nets description: "Use DeepMind Graph Nets for GraphsTuple data, TensorFlow graph utilities, and Sonnet graph-network modules." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Graph Nets

Use this repo skill when a task uses DeepMind Graph Nets (graph_nets) to build graph neural networks in TensorFlow/Sonnet, manipulate GraphsTuple graph batches, convert NetworkX or NumPy graph data, or debug Graph Nets module and utility errors.

Read first

• Repository provenance: read before deciding whether this skill is current for a checkout or package version.
• Compatibility: choose TensorFlow 1/Sonnet 1 versus TensorFlow 2/Sonnet 2, and avoid incompatible NetworkX/TensorFlow stacks.
• Troubleshooting: cross-cutting install/import, dependency, GPU, and version symptoms.
• Demo recipes: distilled guidance from the tutorial, sorting, shortest-path, and physics notebooks without depending on the original notebooks.
• Install check: safe JSON diagnostic for an installed graph_nets runtime.

Install and import check

Graph Nets does not declare TensorFlow as a package requirement, so choose a TensorFlow/Sonnet pair explicitly:

# TF2/Sonnet2-style workflows verified for this skill.
python -m pip install "graph_nets" "tensorflow>=2.2,<2.3" "dm-sonnet>=2,<3" "networkx<3" "protobuf<3.20" "numpy<1.20"

# Legacy TF1/Sonnet1-style workflows verified for this skill.
python -m pip install "graph_nets" "tensorflow>=1.15,<2" "dm-sonnet<2" "tensorflow_probability<0.9" "networkx<2.7" "protobuf<3.20" "numpy<1.20"

Then check:

python - <<'PY'
import graph_nets as gn
from graph_nets import graphs, utils_np, utils_tf, blocks, modules
print("graph_nets import ok")
PY

Use the bundled diagnostic when a user reports a version or import problem:

python path/to/graph-nets/scripts/check_graph_nets_install.py --pretty

Sub-skill routing

Task signal	Read
Create, validate, batch, slice, or round-trip GraphsTuple data from dictionaries, NumPy arrays, or NetworkX graphs	graph-data
Use utils_tf for tensor GraphsTuples, placeholders/feed dicts, concat, repeat, fully-connected graphs, padding/masks, tf.function signatures, or TF1/TF2 execution choices	tensorflow-ops
Build or debug EdgeBlock, NodeBlock, GlobalBlock, GraphNetwork, GraphIndependent, InteractionNetwork, RelationNetwork, DeepSets, CommNet, SelfAttention, or demo model factories	graph-models

Operating checklist

1. Choose the runtime stack first. The TF1 stack exposes tf.Session and top-level tf.placeholder; the verified TF2 stack uses eager tensors and tf.function signatures and does not expose those top-level TF1 symbols.
2. Normalize graph data before modeling. Use the graph-data sub-skill to enforce GraphsTuple field invariants, data dictionary key consistency, and NetworkX feature conventions.
3. Move into TensorFlow utilities only after data is valid. Use the TensorFlow ops sub-skill for tensor conversion, padding, masks, session fetchability, or tf.function input signatures.
4. Assemble modules with explicit field assumptions. Use graph-models when choosing use_* flags, reducers, Sonnet factories, or high-level message-passing architecture.
5. Validate with tiny smoke checks before long demos. Full notebooks are interactive/training-scale; prefer the bundled scripts and selected native unit tests before attempting notebook reproduction.

Avoid this skill when

• The task is about JAX graph networks; prefer Jraph or other JAX-specific guidance.
• The task is about PyTorch Geometric, DGL, or generic NetworkX algorithms rather than Graph Nets GraphsTuple/TensorFlow/Sonnet APIs.
• The task only edits this repository's packaging or contribution process and does not use Graph Nets as a library; use general Python repository maintenance guidance.
• The task needs a modern TensorFlow/Keras GNN library with maintained TF 2.x APIs; Graph Nets is legacy and may require pinned dependencies.