---

name: temporal-signals description: "Guides construction, slicing, splitting, iteration, batching, and debugging of PyTorch Geometric Temporal signal iterators." disable-model-invocation: true metadata: disco-role: operating license: MIT

Temporal Signals

Use this sub-skill when a task is about building or debugging PyTorch Geometric Temporal snapshot iterators from in-memory arrays, not about downloading datasets or training models.

Keep this file as the router. Read the bundled references for constructor signatures, shape contracts, recipes, and failure recovery.

Trigger phrases

Load this sub-skill for requests such as:

• "create a StaticGraphTemporalSignal/DynamicGraphTemporalSignal/DynamicGraphStaticSignal"
• "iterate temporal graph snapshots" or "what does dataset[0] return?"
• "split a temporal signal into train and test snapshots"
• "slice a temporal iterator without shuffling"
• "add optional per-snapshot attributes with **kwargs"
• "use temporal batches with a batch vector or batch_dict"
• "construct heterogeneous author/paper temporal snapshots"
• "debug Temporal dimension inconsistency. or missing hetero keys"

Covered APIs

This sub-skill covers:

• Homogeneous iterators: StaticGraphTemporalSignal, DynamicGraphTemporalSignal, DynamicGraphStaticSignal.
• Homogeneous batch iterators: StaticGraphTemporalSignalBatch, DynamicGraphTemporalSignalBatch, DynamicGraphStaticSignalBatch.
• Heterogeneous iterators: StaticHeteroGraphTemporalSignal, DynamicHeteroGraphTemporalSignal, DynamicHeteroGraphStaticSignal.
• Heterogeneous batch iterators: StaticHeteroGraphTemporalSignalBatch, DynamicHeteroGraphTemporalSignalBatch, DynamicHeteroGraphStaticSignalBatch.
• temporal_signal_split, integer snapshot access, slice access, iterator reset behavior, snapshot_count, and optional temporal attributes passed through **kwargs.

Route elsewhere

• Real benchmark loader choice, downloads, cache directories, or get_dataset side effects: use dataset-loaders.
• IndexDataset, get_index_dataset, sequence-to-sequence index batches, allGPU, world_size, or Dask-DDP: use index-batching.
• Model forward calls, training loops, recurrent hidden state, or forecasting heads after snapshots are produced: use recurrent-layers or attention-and-hetero-layers.

First workflow

1. Classify graph and feature dynamics. Static edges plus temporal features use StaticGraphTemporalSignal; temporal edges plus temporal features use DynamicGraphTemporalSignal; temporal edges plus one fixed feature matrix use DynamicGraphStaticSignal.
2. Choose homogeneous vs heterogeneous. Homogeneous snapshots use arrays and return PyG Data or Batch; heterogeneous snapshots use node/relation dictionaries and return HeteroData or a hetero Batch.
3. Choose ordinary vs batch variant. Use batch variants only when each temporal snapshot already contains multiple disjoint graphs encoded by a batch vector or batch_dict.
4. Validate temporal lengths before construction. Every temporal sequence must have the same snapshot count for its class. Prefer iterator.snapshot_count over len(iterator) because __len__ is not implemented consistently across all iterator classes.
5. Inspect one snapshot. Check object type and shapes: edge_index, edge_attr, x, y, and optionally batch or per-type attributes.
6. Split or slice only after validating shapes. temporal_signal_split preserves chronological order and calls slice access; it does not shuffle, copy deeply, or validate the ratio for you.
7. Hand off after snapshots. Once the iterator yields correct snapshots, route model selection/training to a model sub-skill.

Runtime references and helper

• API reference: signatures, temporal length contracts, snapshot object types, shapes, slicing, and split behavior.
• Workflow recipes: self-contained static, dynamic, dynamic-static, heterogeneous, batched, slicing, and train/test split examples.
• Troubleshooting: mismatched lengths, edge shapes, hetero key mismatches, optional attribute alignment, and split/slice surprises.
• Synthetic smoke script: no-download validator for ordinary, batch, hetero, dynamic-static, and split cases.