| English | 中文 |
Use architecture and data methods to design a study. Use the software guide here to select an environment, configure, invoke and check actual operations.
| Input route | Computer-use case | Current support |
|---|---|---|
| Numeric table | --case tabular: feature units |
Implemented direct CPU synthetic helper |
| Numeric sequence | --case sequence: historical units |
Implemented direct CPU synthetic helper; not a real forecasting adapter |
| Numeric grid | --case grid: spatial patches |
Implemented direct CPU synthetic helper |
| Combined numeric inputs | --case fusion: static, history and grid units |
Implemented direct CPU synthetic helper |
| External sequence model | TimeXer | Reviewed upstream commands; manual adaptation, no AirMosaic runner |
| Text/pretrained multimodal models | External framework references in the software guide | Requires a separate data/model implementation; no text training case here |
The four CPU cases generate all inputs internally and never accept an external dataset. They demonstrate tensor preparation, branch projection, attention, fitting, held-out scoring and model reload. They are not reproductions of FT-Transformer, PatchTST or TimeXer. The fusion case concatenates projected units for shared self-attention; it does not implement TimeXer’s cross-attention.
The entry is run_transformer_demo.py; the reusable toy model is transformer_demo.py. Existing R/Python scientific tools and task contracts are unchanged. The helper is not registered with run-tool or MCP, and does not inherit runner environment isolation or scientific plan enforcement. Follow the guide’s external-output checks and use a reviewed new-study implementation for real data.