Preprocessing and augmentation

Preprocessors

Preprocessing (baseline removal, band-pass filtering) is expressed as composable configs. The high-level helpers apply a standard chain to whole segments:

from torch_ecg._preprocessors import PreprocManager, preprocess_multi_lead_signal

filtered = preprocess_multi_lead_signal(
    raw_sig,
    fs=500,
    bl_win=[0.2, 0.6],      # median-filter baseline windows (s)
    band_fs=[0.5, 45.0],    # band-pass (Hz)
    filter_type="fir",
)

Each step can also be used standalone via its PreProcessor wrapper (function + hyper-parameters), or composed explicitly through PreprocManager, which is how the datasets apply them.

Augmenters

Augmenters are nn.Modules operating on batched (batch, lead, time) signals together with their labels (some augmenters, e.g. lead flips, permute labels too):

import torch
from torch_ecg.augmenters import AugmenterManager, BaselineWanderAugmenter, RandomFlip, RandomRenormalize

augs = [
    BaselineWanderAugmenter(fs=500, bw=0.5, fs_bw=100),
    RandomFlip(),
    RandomRenormalize(),
]
am = AugmenterManager(*augs, random=False)
aug_sig, aug_label = am(sig, label)     # (B, L, T), (B, C) in; same shapes out

With random=True the manager picks a random augmenter per batch (useful for aggressive augmentation budgets). The available augmenters are listed in torch_ecg.augmenters (baseline wander, flips, stretch-compress, cutout, noise, …).

Determinism

Reproducible runs need exactly one seeding point at the entry of the program plus per-worker reseeding for DataLoader workers:

from torch_ecg.utils import seed_everything, worker_init_fn
from torch.utils.data import DataLoader

seed_everything(42)        # python / numpy / torch / DEFAULTS.RNG, all at once
dl = DataLoader(ds, batch_size=32, worker_init_fn=worker_init_fn, ...)

worker_init_fn seeds each worker from the base seed and the worker id, so augmentation inside workers is reproducible across runs. This is the pattern every benchmark uses; with a fixed seed, a run is fully determined.