deep-gcns-torch
Use the DeepGCNs PyTorch operating guide for graph convolution layers, dynamic and dilated KNN blocks, point-cloud classification or segmentation, PPI, OGB/DeeperGCN, and reversible memory-efficient GNN workflows.
- License 许可
- MIT
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- deep-gcns-torch
- license
- MIT
- metadata
- { "disco-role": "operating" }
- description
- Use the DeepGCNs PyTorch operating guide for graph convolution layers, dynamic and dilated KNN blocks, point-cloud classification or segmentation, PPI, OGB/DeeperGCN, and reversible memory-efficient GNN workflows.
- disable-model-invocation
- true
02 资源文件 6
- 参考文档references/api-overview.md2.5 KB
- 参考文档references/repo-provenance.md1.8 KB
- 参考文档references/repo-routing-metadata.json509 B
- 参考文档references/troubleshooting.md3.5 KB
- 脚本scripts/check_env.py2.5 KB
- 脚本scripts/checkpoint_roundtrip.py1.7 KB
03 SKILL.md 正文
DeepGCNs Torch
Use this router when a task names DeepGCNs, DeeperGCN, deep_gcns_torch,
gcn_lib, GENConv, dynamic KNN graph blocks, ModelNet40/S3DIS/PartNet point
clouds, PPI, OGB graph benchmarks, RevGNN, or RevGAT.
Operating boundary
This is a self-contained guide distilled from a historical PyTorch repository. It does not download datasets, fetch external checkpoints, launch long training, start distributed jobs, or open visualization windows. First stage caller-owned data and checkpoints, then follow the relevant workflow. Read repository provenance when deciding whether a source checkout has drifted, and read troubleshooting when installation or version behavior is unclear.
Minimal environment check
The source repository has no pyproject.toml, setup.py, or console entry
point. Use a coherent PyTorch/PyG stack rather than the obsolete repository
installer. At minimum, install PyTorch, PyTorch Geometric, matching
torch-scatter and torch-cluster wheels, and add ogb, h5py, scikit-learn,
or tensorboard only for the selected workflows. Run the bundled checks before
using a real model:
python scripts/check_env.py
python scripts/checkpoint_roundtrip.py
For a CUDA task, use python scripts/check_env.py --cuda; CUDA availability is
not proof of dataset-scale memory or benchmark reproduction. Match every
compiled PyG extension to the installed PyTorch and CPU/CUDA build. Do not run
the old one-shot installer: it mutates environments and pins obsolete CUDA
10.2-era packages.
Route by the user's task
- Layer or architecture API — sparse/dense
GraphConv,DynConv,GENConv, aggregators, KNN/dilation, blocks, or reversible primitives: read graph-layers. - Point clouds — ModelNet40, S3DIS dense/sparse, PartNet, segmentation layouts, point-cloud checkpoints, or optional OBJ visualization: read point-cloud-workflows.
- OGB/DeeperGCN —
ogbn-*,ogbg-*,ogbl-collab, graph pooling, partitions, atom/bond features, RevGNN, or DGL RevGAT: read ogb-workflows. - PPI — PPI multilabel node classification, F1 metrics, or PPI checkpoint alignment: read ppi-workflows.
If a request crosses routes, start here, state the data layout and backend,
then read each owning sub-skill. Generic layer semantics belong to
graph-layers; task-specific data and metrics remain with their workflow.
Fast safety gates
- Identify CPU versus CUDA, PyG extension compatibility, data root, checkpoint availability, and expected metric before running anything.
- Use the owning sub-skill's bundled synthetic helper first. These helpers are offline, deterministic, and do not import the original checkout.
- Never treat a CPU shape smoke as proof of GPU memory, a historical benchmark number, or a dataset split result.
- Keep the source task's working-directory import quirks and external download requirements explicit; do not turn them into hidden runtime dependencies.
- For a changed source revision or package API, compare provenance and use a refresh workflow rather than silently applying stale claims.
Shared references and checks
- API overview summarizes shared tensor, module, and utility contracts before choosing a focused route.
- Troubleshooting covers cross-cutting install, import, data, checkpoint, backend, and version failures.
- Environment check is a safe dependency/backend probe.
- Checkpoint roundtrip verifies basic PyTorch checkpoint serialization without touching caller data.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).