Skip to content

callbacks.md

Latest commit

 

History

History
23 lines (17 loc) · 917 Bytes

callbacks.md

File metadata and controls

23 lines (17 loc) · 917 Bytes

ROCV Callbacks

PatchTSTTrainer provides a callback hook that fires after rolling-origin cross-validation (ROCV) slices are generated and before training begins. Callbacks receive (seed, fold_idx, train_mask, val_mask, cfg) and run once for every fold.

from LGHackerton.models.patchtst.trainer import PatchTSTTrainer

def log_fold(seed, fold_idx, tr_mask, va_mask, cfg):
    print(f"fold {fold_idx} -> {tr_mask.sum()} train / {va_mask.sum()} val")

PatchTSTTrainer.register_rocv_callback(log_fold)

Failures inside the callback are isolated from training: the trainer converts exceptions into warnings so that logging issues do not interrupt model fitting.

The public callback API intentionally avoids exposing private helpers. Older versions without register_rocv_callback required wrapping the private _make_rocv_slices function; train.py still supports this as a legacy fallback.