---

name: data-and-datasets description: "Create, validate, split, transform, and package PyTorch Geometric Data, HeteroData, Batch, Dataset, and InMemoryDataset objects with safe local fixtures." disable-model-invocation: true metadata: disco-role: operating license: MIT

Data and Datasets

Use this sub-skill when the task is about graph containers, hetero graph containers, batching graph objects, local dataset classes, transforms, dataset splitting, or synthetic fixtures in PyTorch Geometric 2.9.

Start Here

• Read data formats to construct and validate Data, HeteroData, and Batch objects, including expected tensor shapes and metadata.
• Read datasets and transforms to implement Dataset or InMemoryDataset, choose transform vs pre_transform, create local layouts, and split nodes or links.
• Read troubleshooting when validation fails, masks are missing, transforms run in the wrong order, or processed dataset caches look stale.
• Run validate_basic_graph.py for a safe synthetic smoke check of homogeneous and heterogeneous graph containers.

Quick Commands

python scripts/validate_basic_graph.py --fixture homogeneous
python scripts/validate_basic_graph.py --fixture hetero
python scripts/validate_basic_graph.py --fixture invalid --expect-invalid

Each command prints a JSON summary and uses only tiny in-memory tensors plus installed torch and torch_geometric APIs. No downloads, GPUs, network access, credentials, or destructive writes are required.

Route Boundaries

• For mini-batch loaders, neighbor sampling, and link sampling loaders, use ../loaders-and-sampling/SKILL.md.
• For model layers, MessagePassing, convolutions, pooling, and training loops, use ../gnn-modeling/SKILL.md.
• For advanced heterogeneous modeling, HeteroConv, to_hetero, and hetero-specific training recipes, use ../heterogeneous-graphs/SKILL.md.
• For distributed graph stores, remote backends, partitioning, and hardware scaling, use ../scalable-distributed/SKILL.md.

Safe Defaults

• Build graph fixtures with torch_geometric.data.Data(x=..., edge_index=..., y=...) or HeteroData() and call .validate(raise_on_error=True) before passing them to loaders or models.
• Set data.num_nodes explicitly when node features are absent or isolated nodes matter.
• Use torch.long COO edge_index tensors shaped [2, num_edges] with zero-based node ids.
• Prefer InMemoryDataset only when all processed graphs fit in CPU memory; otherwise implement Dataset with one processed file per graph.
• Keep runtime examples local and synthetic unless the user explicitly requests a dataset download.