model-zoo-pretraining
Select, instantiate, inspect, and troubleshoot DGL-LifeSci model-zoo architectures, pretrained models, readouts, molecule embeddings, and link prediction heads.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- model-zoo-pretraining
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Select, instantiate, inspect, and troubleshoot DGL-LifeSci model-zoo architectures, pretrained models, readouts, molecule embeddings, and link prediction heads.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/constructor-recipes.md7.1 KB
- 参考文档references/model-catalog.md12.7 KB
- 参考文档references/troubleshooting.md9.0 KB
- 脚本scripts/inspect_model_constructors.py10.0 KB
03 SKILL.md 正文
Model Zoo and Pretraining
Use this sub-skill when an agent needs to choose or instantiate dgllife.model architectures, inspect constructor compatibility, load pretrained models, produce molecule embeddings from pretrained GIN encoders, or wire link-prediction heads.
Use This For
- Selecting among
dgllife.model.gnn,dgllife.model.readout,dgllife.model.model_zoo, anddgllife.model.load_pretrainedentry points. - Matching graph feature tensors such as
g.ndata['h'],g.ndata['hv'],g.edata['e'], or pretrained categorical lists to constructor parameters. - Diagnosing constructor signatures, missing root exports, zero-in-degree errors, and pretrained checkpoint download failures.
- Adapting molecule embedding or link-prediction patterns without running large dataset downloads.
Route Elsewhere
- For SMILES parsing, featurizers, graph construction, label masks, splitters, and dataset field names, use
../molecule-data-prep/SKILL.md. - For property prediction training loops, metrics, masking, and model evaluation workflows, use
../property-prediction/SKILL.mdafter choosing the model here. - For WLN reaction-center/ranking workflows and reaction-specific featurization, use
../reaction-prediction/SKILL.md; this sub-skill only covers shared constructors and pretrained names.
References
- Start with references/model-catalog.md for constructor families, feature dimensions, output expectations, pretrained names, molecule embedding, and link prediction usage.
- Use references/constructor-recipes.md for tiny CPU instantiation snippets and shape reasoning without data downloads.
- Use references/troubleshooting.md for import/install, optional dependency, data/config, CLI/API misuse, pretrained download, zero-in-degree, and version/export caveats.
- Run scripts/inspect_model_constructors.py to print installed constructor signatures and optionally instantiate tiny CPU-safe constructors.
Fast Workflow
- Identify whether the user needs a GNN encoder, graph-level predictor, readout, pretrained model, molecule embedding encoder, or link-prediction head.
- Check feature ownership: node and edge feature construction belongs to
molecule-data-prep; this sub-skill only maps those tensor widths and field names into constructors. - Read
references/model-catalog.mdfor the closest constructor family and any pretrained feature contract. - Validate the installed API with
python scripts/inspect_model_constructors.py --constructors GCNPredictor,GATPredictor,MPNNPredictorbefore writing code that must run across DGL-LifeSci versions. - If pretrained checkpoints are needed, provide a no-network fallback plan because
load_pretrained()downloads checkpoint files on first use.
Version Caveats
The verified installed package was dgllife 0.3.1 with DGL 1.1.3 and CPU Torch. The source includes GATv2Predictor, but one installed inspection found no GATv2Predictor attribute at dgllife.model; prefer GATv2 as the stable encoder and inspect submodule fallback availability before promising root-level GATv2Predictor imports.
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).