build_thread_env
Env vars pinning XLA (Chronax) CPU parallelism to threads.
NOTE: intra_op_parallelism_threads is a TensorFlow session option and may be ignored by XLA; the effective CPU pin rests on OMP_NUM_THREADS + xla_cpu_multi_thread_eigen. The actual Chronax thread pin is validated empirically during the Task 14 calibration, not assumed here.
| Parameter | Type | Default | Description |
|---|---|---|---|
| threads | int | - | (undocumented) |
Returns: dict[str, str] (undocumented)
dataset_kind
(No prose summary provided in docstring.)
| Parameter | Type | Default | Description |
|---|---|---|---|
| spec | dict | - | (undocumented) |
Returns: str (undocumented)
resolve_horizon
Per-dataset h / input_size override experiment defaults when present.
| Parameter | Type | Default | Description |
|---|---|---|---|
| spec | dict | - | (undocumented) |
| exp | dict | - | (undocumented) |
Returns: tuple[int, int] (undocumented)
load_dataset
Load a wide Chronax-style CSV according to dataset kind.
| Parameter | Type | Default | Description |
|---|---|---|---|
| spec | dict | - | (undocumented) |
Returns: dict with keys:
* kind: str
* y: float32 array — shape (T,) univariate/covariate or (T, N) multivariate
* X: float32 array | None — shape (T, F) for covariate, else None
* y_cols: list[str] | None — multivariate channel names
* hist_exog_cols: list[str] | None — covariate exog names
load_dataset_y
Backward-compatible univariate loader (tests / callers).
| Parameter | Type | Default | Description |
|---|---|---|---|
| spec | dict | - | (undocumented) |
Returns: np.ndarray (undocumented)
nf_subprocess_code
Build the python -c source run inside .venv-nf for one NF seed.
torch.set_num_threads pins CPU threads; the neuralforecast import is BEFORE t0 so the timer measures only fit+predict (fair timing). Loss is MAE.
Dataset kinds: * univariate — (unique_id, ds, y); n_series=1 when required * multivariate — one unique_id per y_col; n_series=N when required * covariate — single series + hist_exog_list / futr_exog_list when supported
| Parameter | Type | Default | Description |
|---|---|---|---|
| nf_name | str | - | (undocumented) |
| spec | dict | - | (undocumented) |
| h | int | - | (undocumented) |
| input_size | int | - | (undocumented) |
| params | dict | - | (undocumented) |
| seed | int | - | (undocumented) |
| threads | int | - | (undocumented) |
Returns: str (undocumented)
run_chronax_seed
Fit+predict one Chronax seed; return metric row with an error field.
| Parameter | Type | Default | Description |
|---|---|---|---|
| cls | type | - | (undocumented) |
| data | dict | - | (undocumented) |
| h | int | - | (undocumented) |
| input_size | int | - | (undocumented) |
| chronax_params | dict | - | (undocumented) |
| seed | int | - | (undocumented) |
Returns: dict (undocumented)
run_nixtla_seed
Spawn a fresh .venv-nf process for one NF seed; return a metric row.
Fresh process per seed = clean per-seed RNG + matches how baselines were captured (spec §7, §14).
| Parameter | Type | Default | Description |
|---|---|---|---|
| nf_name | str | - | (undocumented) |
| spec | dict | - | (undocumented) |
| h | int | - | (undocumented) |
| input_size | int | - | (undocumented) |
| nf_params | dict | - | (undocumented) |
| seed | int | - | (undocumented) |
| threads | int | - | (undocumented) |
Returns: dict (undocumented)
emit_row
(No prose summary provided in docstring.)
| Parameter | Type | Default | Description |
|---|---|---|---|
| library | str | - | (undocumented) |
| dataset | str | - | (undocumented) |
| model | str | - | (undocumented) |
| seed | int | - | (undocumented) |
| iter_idx | int | - | (undocumented) |
| warmup_seeds | int | - | (undocumented) |
| row | dict | - | (undocumented) |
Returns: None (undocumented)