---

name: pytorch-geometric-temporal description: "Routes PyTorch Geometric Temporal tasks for temporal graph signals, built-in datasets, recurrent and attention graph neural layers, index-batching, and optional Dask-DDP workflows." disable-model-invocation: true metadata: disco-role: operating license: MIT

PyTorch Geometric Temporal

Use this repo skill when a task involves the torch-geometric-temporal / torch_geometric_temporal package: temporal graph signal iterators, built-in spatiotemporal datasets, temporal graph neural network layers, memory-efficient index-batching, or optional distributed Dask-DDP training patterns.

This skill is self-contained. Do not require the original repository checkout for runtime use; use the bundled references and scripts below.

Start here

1. Check installation and versions. Read install and compatibility, then run check_environment.py if you need an import/backend smoke check.
2. Choose the route. Use the sub-skill map below. Open the nearest sub-skill before writing task-specific code.
3. Validate with synthetic smoke first. Prefer the bundled no-download scripts before adapting network-backed dataset examples or long training loops.
4. Treat real dataset loaders as side-effectful. Many loader constructors download remote files; read the dataset-loader route before constructing them.

Sub-skill map

User task	Read
Build a temporal signal from arrays, slice snapshots, use temporal_signal_split, add optional attributes, or debug Data/Batch/HeteroData outputs	temporal-signals
Choose a built-in dataset loader, understand get_dataset signatures, plan raw_data_dir, or avoid constructor-time downloads	dataset-loaders
Use recurrent temporal graph layers such as DCRNN, GConvGRU, TGCN, A3TGCN, AGCRN, MPNNLSTM, or manage hidden states	recurrent-layers
Use attention or heterogeneous layers such as STConv, ASTGCN, MSTGCN, GMAN, MTGNN, AAGCN, DNNTSP, or HeteroGCLSTM	attention-and-hetero-layers
Use IndexDataset, get_index_dataset, index=True, allGPU, 5-tuple/7-tuple loader returns, or optional Dask-DDP	index-batching

Package facts to remember

• Public distribution name: torch-geometric-temporal / installed metadata torch_geometric_temporal.
• Import package: torch_geometric_temporal.
• Base install expects compatible PyTorch and PyTorch Geometric first or through dependency resolution.
• Public extras: torch-geometric-temporal[index] for index-batching data dependencies; torch-geometric-temporal[ddp] for Dask-DDP-oriented dependencies.
• The inspected source metadata reports distribution version 0.56.2, while torch_geometric_temporal.__version__ reports 0.54.0; use distribution metadata for package install/version comparisons and note the in-package constant mismatch when debugging.
• The package exposes no main CLI. Workflows are Python API-first.

Shared references and scripts

• Repository provenance: source commit, package versions, evidence paths, and refresh cues.
• Router metadata: structured scenario placement for managed repo-skill import tooling.
• Install and compatibility: install commands, optional extras, PyTorch/PyG/CUDA/DDP notes, and import checks.
• Model and data map: compact catalog of signal classes, loaders, model families, sub-skill owners, and validation paths.
• Cross-cutting troubleshooting: install/import, optional dependencies, backend, downloads, and version mismatch recovery.
• check_environment.py: safe import/version/backend smoke check with JSON output.

Safe validation ladder

Run checks in this order when diagnosing user code:

1. python scripts/check_environment.py --json from this skill root or with an absolute script path.
2. The relevant sub-skill smoke script, for example sub-skills/temporal-signals/scripts/signal_iterator_smoke.py --mode all --json.
3. A tiny user-shaped synthetic case using the selected sub-skill references.
4. Real dataset loader construction or original benchmark-scale behavior only after the user accepts network/cache/runtime costs.

Avoid using this skill when

• The task is only generic PyTorch Geometric static graph modeling with no temporal signal, dataset loader, or PGT class involved.
• The user asks for general time-series forecasting without graph structure.
• The user is editing the package source as a maintainer; use a repository-maintenance workflow rather than this operating skill.