| English | 中文 |
timexer_workflow_reference is a non-executable catalog reference. Run reviewed upstream code in its own environment; AirMosaic has no TimeXer training/inference adapter.
This sequence reference is part of the parent machine-learning capability. Read the matching human-facing method before adapting the code.
Use upstream tracking for the reviewed source commit and dependencies. Keep its source outside this checkout, referenced by TIMEXER_SOURCE_DIR. Use an independent Python/PyTorch environment.
| Setting | Required decision |
|---|---|
| Target and time | Target column, sampling frequency, forecast origin and horizon |
| Input | Ordered regular-time CSV with a date column and numeric variables; one entity per sequence |
| Roles | MS for multivariate history with one target, M for multivariate outputs |
| Windows | seq_len, pred_len, patch_len; preserve sufficient history and target-patch divisibility |
| Model/budget | Dimensions, heads, layers, batch size, epochs, workers and device |
| Storage | Fixed input version and fresh external project result directory |
Usage and command example contains the complete upstream run.py call and dependency/input checks. Begin with enough synthetic history. Run from a new project result folder so relative checkpoint, prediction and log files remain there.
The tracked model consumes historical inputs. Known-future covariates, mixed frequencies and multiple entities need a data/model adaptation. The tracked benchmark prints test errors during training; an independent final-test protocol needs explicit handling.
Check forecast origins, horizons, shapes, finite values, target units and inverse scaling. Compare persistence/seasonal baselines and inspect errors by horizon/entity. Preserve checkpoint identity and the exact training environment.
--is_training 0 is labeled testing, not history-only deployment inference. Implement and verify scaler reuse and a history-only input contract before using new unlabeled batches. Source-specific seed/device/checkpoint details are in the usage reference.
The forecasting method explains the research design. Tensor details remain in the technical reference; asset storage governs study outputs.