TensorFlow gets its own PyTorch Geometric

What it does

tf_geometric implements graph neural networks for TensorFlow using a message-passing backend rather than dense matrix operations. It provides Keras-like OOP layers—GCN, GAT, GraphSAGE, and others—plus a lower-level functional API for custom mapper and reducer logic. The library covers node classification, graph classification, link prediction, and distributed training, and includes built-in loaders for Cora, PPI, and TU datasets.

The interesting bit

The library’s most unusual trait is its straddling of TensorFlow 1.x and 2.x at a time when most new ML code ignores the former. It also exposes raw message-passing primitives—identity_mapper, sum_reducer, sum_updater—beneath the friendly layer API, letting you drop into map-reduce-style graph operations when pre-built layers are too rigid.

Key highlights

• Implements a broad catalog of models: GAT, GCN, GraphSAGE, APPNP, GIN, DiffPool, SAGPooling, and others.
• Supports batched graphs via BatchGraph with pooling, top-k node selection, and graph-level readout.
• Provides distributed training examples and sparse node feature handling.
• Ships with dataset loaders for standard benchmarks (Cora, PPI, TU).

Caveats

• The README claims message passing is “way efficient” compared with dense-matrix implementations and friendlier than sparse-matrix ones, but offers no benchmarks or numbers to back up the performance angle.
• Explicit support for TensorFlow 1.x alongside 2.x is a notable anachronism; the README does not clarify how deeply the legacy path is maintained or whether it imposes constraints on the TF2 API.
• At 509 stars, the project is a niche alternative, and its documentation focuses on demos rather than design rationale or comparative analysis.

Verdict

Worth a look if you need native graph attention, pooling, or graph classification layers inside TensorFlow. If you are free to choose your framework, the README itself points to PyTorch Geometric as the original reference.