multimodal-automl
Use AutoGluon MultiModalPredictor for text, image, tabular+text/image, document, NER, semantic matching, object detection, semantic segmentation, feature extraction, zero-shot, and deployment workflows.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- multimodal-automl
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Use AutoGluon MultiModalPredictor for text, image, tabular+text/image, document, NER, semantic matching, object detection, semantic segmentation, feature extraction, zero-shot, and deployment workflows.
- disable-model-invocation
- true
02 资源文件 8
- 参考文档references/api-reference.md8.8 KB
- 参考文档references/data-formats.md6.9 KB
- 参考文档references/deployment.md5.0 KB
- 参考文档references/object-detection.md5.5 KB
- 参考文档references/troubleshooting.md6.5 KB
- 参考文档references/workflows.md8.5 KB
- 脚本scripts/inspect_multimodal_inputs.py19.6 KB
- 脚本scripts/multimodal_smoke.py8.0 KB
03 SKILL.md 正文
MultiModal AutoML
Use this sub-skill when the user needs AutoGluon AutoMM or MultiModalPredictor for text, image, mixed tabular+text/image, document, NER, semantic matching, object detection, semantic segmentation, feature extraction, zero-shot inference, or multimodal deployment/export tasks.
Route structured-only supervised learning to ../tabular-ml/ and forecasting to ../time-series-forecasting/. Do not treat external dataset downloads, competition submissions, or raw repository CI examples as runtime requirements; distill the needed data/API pattern and work with the user's local files.
Start Here
- For constructor and method signatures, problem types, metrics, and model config knobs, read
references/api-reference.md. - For end-to-end task recipes, including semantic matching and save/load, read
references/workflows.md. - For DataFrame, image, document, NER, matching, segmentation, and local path expectations, read
references/data-formats.md. - For COCO/VOC/DataFrame object detection handling, read
references/object-detection.md. - For ONNX, TensorRT, embedding extraction, and offline packaging, read
references/deployment.md. - For dependency, backend, schema, and checkpoint failure modes, read
references/troubleshooting.md.
Safe Bundled Helpers
- Use
scripts/inspect_multimodal_inputs.py --helpto validate local CSV/JSON data, image/document paths, matching identifiers, COCO metadata, or VOC-style annotations without training or downloading. - Use
scripts/multimodal_smoke.py --helpto check installed AutoGluon multimodal imports, API signatures, problem types, and optional backend versions without training by default.
Routing Guide
| User request | Read first | Key API |
|---|---|---|
| Text/image/mixed classification or regression | references/workflows.md |
MultiModalPredictor(label=..., problem_type=...) |
| Document image/PDF classification | references/data-formats.md |
column_types={...: "document"} or inferred document columns |
| NER/entity extraction | references/data-formats.md |
problem_type="ner", column_types={text_col: "text_ner"} |
| Text/image/image-text similarity or retrieval | references/workflows.md |
query=..., response=..., match_label=..., candidate_data, id_mappings |
| Object detection | references/object-detection.md |
problem_type="object_detection", COCO/VOC/DataFrame data |
| Semantic segmentation | references/data-formats.md |
problem_type="semantic_segmentation", image and mask columns |
| Feature extraction or zero-shot image classification | references/workflows.md |
problem_type="feature_extraction" or "zero_shot_image_classification" |
| ONNX/TensorRT/export/offline deployment | references/deployment.md |
save, load, export_onnx, optimize_for_inference |
Default Agent Approach
- Identify the task family and route away if it is structured-only tabular or time-series.
- Confirm data shape before code: label/query/response columns, local image/document paths, object-detection annotation format, segmentation mask availability, and optional backend constraints.
- Prefer CPU-safe, no-download validation first with the bundled scripts; only propose training/export after local data and dependencies are plausible.
- Keep
path/save_pathexplicit for real training, use shorttime_limitonly for smoke examples, and set backend hyperparameters intentionally for CPU/GPU. - For matching/retrieval, distinguish row-pair prediction from candidate search:
predict(data, candidate_data=...),evaluate(query_data=..., response_data=...), or embedding-based search. - For deployment, call out pickle trust for
load, optionalstandalone=Truepackaging, and ONNXRuntime/TensorRT dependency requirements.
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).