| English | 中文 |
Scientific choices belong to methods and its data guide. This reference provides concrete software operations and their actual verification scope.
The direct example uses Python 3.11+, PyTorch 2.6+ and this repository checkout. Select a trusted environment whose python resolves to those dependencies. PyTorch CPU execution was checked on Python 3.14.4 and PyTorch 2.13.0+cpu on 2026-10-30. That is an environment receipt, not a requirement to upgrade other environments. Other versions/platforms need their own checks.
If dependencies are missing, prepare a separate environment using the official PyTorch selector under the task’s installation authorization. The helper never installs packages, requests credentials, downloads data/weights or selects a GPU. It needs no Lightning or Hugging Face dependency for these examples.
| Software reference | Useful operation | Adaptation boundary |
|---|---|---|
| PyTorch Lightning skill | Separate data, training, validation, checkpoint and prediction stages | Its classification template is not a numeric regression model; change the data contract, head, loss and metrics before use |
| Transformers skill | Inspect model versions, local weights and supported inference/fine-tuning APIs | Text tokenization and language-model Trainer examples do not define an atmospheric numeric input contract |
| TimeXer operations | Use the actual upstream environment, CLI and input conventions | Custom splits, training-time test access, scaler persistence and history-only inference require adaptation |
The local helper was independently written using the first reference’s separation of responsibilities, not copied from its classification scripts. These source links are references reviewed on 2026-10-30, not an automatic installation or update service.
Use the complete configuration and one declared case:
| Case | Generated input, including batch axis B |
Encoder units |
|---|---|---|
tabular |
static: [B, 4] |
Four numeric features |
sequence |
history: [B, 8, 3] |
Eight historical steps |
grid |
grid: [B, 1, 4, 4] |
Four non-overlapping spatial patches |
fusion |
All three inputs above | Separate projections and identity/position embeddings, concatenated for common self-attention |
No dataset or checkpoint input argument exists. A known synthetic scalar target combines all three internally generated sources; single-input cases receive only their declared source. The examples check software mechanics, not a scientific ranking of model families.
The JSON explicitly controls samples, seed, d_model, heads, layers, epochs, batch_size, learning_rate, weight_decay and dropout. It requires purpose=synthetic_software_demo. Counts govern independent generated cohorts, not a split prescription for real data. The helper rejects unknown fields, incompatible head dimensions, nonfinite settings and budgets beyond its small limits. This JSON is not a task accepted by either existing regression tool.
Commands run from the repository root after selecting the intended Python environment. Create three separate existing roots outside Git, then set AIRMOSAIC_ASSETS_DIR, AIRMOSAIC_CONFIG_DIR, AIRMOSAIC_AUTH_DIR following storage setup. This helper reads only the explicitly selected non-secret JSON under config/examples/ or the personal configuration root; it does not load a profile or inspect authentication files.
Choose an unused external results directory. This PowerShell example uses the project’s minute naming convention:
$amCase = 'fusion'
$amTaskLabel = (Get-Date -Format yyyyMMddHHmm) + '_' + $amCase
$amOutput = Join-Path $env:AIRMOSAIC_ASSETS_DIR ('results/transformer_software_demo/' + $amTaskLabel)
python skills/model/shared/machine_learning_model/scripts/python/run_transformer_demo.py --case $amCase --config-file config/examples/machine-learning-model/python/transformer.synthetic.json --out-dir $amOutput --dry-run
Expected: can_execute=true, blockers=[], external_input_supported=false, output_created=false. Preview validates paths, JSON and installed package metadata without importing Torch or generating tensors. It does not prove that a Torch installation can initialize. Missing Torch gives a dependency blocker instead of an installation attempt.
After preview, execute the same selected case and output path:
python skills/model/shared/machine_learning_model/scripts/python/run_transformer_demo.py --case $amCase --config-file config/examples/machine-learning-model/python/transformer.synthetic.json --out-dir $amOutput
To try another case, choose its name and a fresh output path before repeating preview/execution. Changing the case alone while reusing an existing output directory fails; there is no overwrite switch. Outputs must be beneath external asset results, with at least a task folder and a fresh run folder. Linked locations, Git ancestors and reserved authentication/record names are rejected. Direct execution provides these checks, not OS isolation or the runner’s minimal child environment.
| Artifact | What to inspect |
|---|---|
configuration.json, input-contract.json |
Actual case, counts, shapes, seed and train-only preprocessing |
events.jsonl |
Ordered, time-stamped epochs, validation criterion, checkpoint selection and later test evaluation |
metrics.json, predictions.csv |
Finite keyed predictions, independent RMSE/MAE/R² and training-mean reference |
checkpoint.pt |
Model parameters and training-fitted preprocessing; only this newly generated file is reloaded |
checks.json, status.json |
Finite nonzero gradients, metric agreement, reload equality, CPU/package versions and scientific_acceptance=false |
artifacts.json |
Relative file paths and checksums; absent on an incomplete run |
The implementation fits preprocessing on training inputs, trains on training labels, chooses a checkpoint by validation MSE, and generates/evaluates the test cohort only after that choice. It reloads its own checkpoint with weights_only=True. It does not open arbitrary saved models. Existing metric helpers are reused; network scores may be worse than the mean reference and must still be reported.
Events carry actual UTC times with an offset and record software actions/checks only. They do not capture a host agent conversation, a study’s generation rationale or a retrospective scientific review; those follow the workflow handoff contract when applicable.
Keep partial outputs and use a fresh destination after investigating a failure. Typical codes are missing_separated_roots, demo_configuration_root, demo_budget_exceeded, demo_output_exists and demo_unsafe_output; missing_torch is a preview blocker. Do not move data into an authentication root or relabel a real task to make it fit this example.
Real tables, observations, raster files, text, missing-modality inputs, irregular windows, GPU/distributed training and arbitrary pretrained weights are unsupported by this helper. It has no run-tool/MCP registration and no scientific plan-approval gate, because its sole accepted task is this internal synthetic demonstration. New scientific use still follows planning and review, with an actual loader, explicit sample keys, availability/split rules and a reviewed execution adapter.
Keep generic code in skills and scientific design in methods; workflows may refer to these cases without copying the training logic. Project results retain external assets/results/<project_id>/ organization, with task-specific internals. No example output is automatically promoted into shared data or committed.