Return the canonical horizon grid for a dataset name.

Split a single-stream (test, labels) chronologically with a gap. Args: test: (T, C) features. labels: (T,) binary labels.

Binary label surface y(t, Delta_t) from time-to-event values. y(t, Delta_t_k) = 1 iff the next event arrives within Delta_t_k steps. Args:

Compute the discrete-hazard CDF probability surface. Discrete-hazard parameterization (monotone in Delta_t by construction): lamb

Build event-prediction entities: label the last timestep as positive.

Pool all (t, Delta_t) cells; report AUPRC, AUROC, best-F1. Args: p_surface: (N, K) predicted probabilities in [0, 1]. y_surface:

Hard-copy encoder -> target_encoder every ``sync_interval_steps``. Only fires under ``target_mode == 'periodic_sync'``. Call once per

Fraction of consecutive (k, k+1) pairs where p decreased.

Per-horizon AUROC + AUPRC + prevalence.

Self-supervised forward pass. Returns the **raw** (unnormalized) predictor output and target embedding. The regularizer requires raw

Pretraining loss: L1 alignment plus a variance-covariance regularizer. L = (1 - alpha) * ||normalize(h_pred) - normalize(h_target)||_1

Fit mean/std on ``train`` and apply to all arrays.

Copy matching weights from encoder into target encoder.

Mean per-horizon AUROC across horizons with 0 < prevalence < 1. The paper's primary metric. Robust to drifting horizon base rates that can ma

Sinusoidal positional encoding for arbitrary positions. Args: positions: 1D tensor of position indices, shape (N,). d: embedding

Positive-weighted BCE on probabilities (post-sigmoid). Args: p: (B, K) probabilities in (0, 1) - typically the discrete-hazard CDF.

Locate the C-MAPSS directory using ``HEPA_DATA_DIR``.

Return h_t for analysis/probing.

Generate the probability surface on a test loader.

Finetune the predictor and event head with positive-weighted BCE. ``early_stop=True`` (default) stops on the validation loss and restores the

Return the per-dataset context window length (Table L).

Return the normalization mode for a dataset ('none' for C-MAPSS, else 'revin').

Global per-channel z-score for a dataset bundle (norm_mode='none' path). Fits a single per-channel mean/std on the concatenated pretrain sequence

Load a C-MAPSS subset, id-splitting the run-to-failure TRAIN engines. The event-prediction framing requires run-to-failure streams so that the

Dispatch a dataset name to its loader.

Load PSM (CSV or NPY) and split chronologically.

Load SMAP as a single concatenated stream + chronological split.

Pretrain HEPA with L1 alignment plus a variance-covariance regularizer. Loss: ``(1 - alpha) * L1(normalize(h_pred), normalize(h_target)) + al

Persist a probability surface plus aligned labels to .npz.

Seed Python, NumPy, and PyTorch (CPU + CUDA) RNGs.

VICReg variance + covariance terms on a (B, D) batch. Variance hinge encourages each feature's standard deviation to be at least 1; covarianc

Mean squared error between predicted probability and binary label.

Right-pad context and target to batch max lengths, build masks.

Expected Calibration Error (ECE) over equal-width probability bins.

Predict future embedding conditioned on horizon Delta_t. Args: h: (B, d_model) context embedding. delta_t: (B,) float

Normalize each (batch, channel) timeseries; return stats for denorm. Args: x: (B, T, C) input. mask: optional (B, T)

Map (B, T, C) -> (B, N_tokens, d). Pads time dim if not divisible.

Encode context to h_t. Args: x: (B, T, C) raw observations. mask: optional (B, T) bool, True = padding. Retu

Encode target interval to a single pooled embedding. Args: x: (B, T, C) target observations x(t : t+Delta_t]. mask: o

Return logits with the trailing feature dim collapsed. Accepts (B, d) -> (B,) or (B, K, d) -> (B, K).

Return the per-dataset context window length (Table L).

Resolve the directory for a dataset, creating parents as needed.

Return the canonical horizon grid for a dataset name.

Load a surface saved by `save_surface`. Returns (p_surface, y_surface, horizons, t_index).

After sync_interval_steps, encoder weights are copied to target.

Sanity check: the default architecture is roughly the paper's 2.16M.