aug_mixup — Within-batch convex mixup of spectra¶
Group: Augmentation · C ABI: n4m_augmentation_mixup_*
Description¶
Batch-wise mixup augmentation.
Parameters¶
Name |
Type |
Default |
|---|---|---|
|
|
|
|
|
|
|
|
|
API and bindings¶
C ABI (ABI 2): n4m_augmentation_mixup_apply · n4m_augmentation_mixup_create · n4m_augmentation_mixup_destroy. Use the linked public header for the exact signature, configuration, and result handles.
Python (verified public re-export):
from n4m.augmentation.mixup import MixupAugmenter
Source signature: MixupAugmenter(alpha: float = 0.2, rng: Optional[PCG64] = None, seed: int = 0).
R: no current source-verified entry point was found for this catalog method.
MATLAB / Octave: no current source-verified entry point was found for this catalog method.
Explanations¶
Bibliographic source¶
Zhang, Cisse, Dauphin & Lopez-Paz (2018), mixup: Beyond Empirical Risk Minimization, ICLR, arXiv:1710.09412 (https://arxiv.org/abs/1710.09412).
Mathematical principle¶
A random permutation \(\pi\) pairs rows and independent weights are drawn as \(\lambda_i\sim\operatorname{Beta}(\alpha,\alpha)\). The operator returns \(X'_i=\lambda_iX_i+(1-\lambda_i)X_{\pi(i)}\). alpha below one favors near-endpoint mixtures; larger values concentrate around one half.
Appropriate uses¶
Regularizing models by filling linear neighborhoods between observed spectra.
Limits and validation¶
This transformer returns only mixed X. It does not mix target labels, although the original mixup method requires the same convex combination of y; callers must preserve label consistency themselves. Pairing is restricted to the batch.
Implementation¶
Python role API n4m.augmentation.mixup.MixupAugmenter; ABI 2 family n4m_augmentation_mixup_{create,apply,destroy}.
The ABI-2 implementation is the n4m_augmentation_mixup_* lifecycle in libn4m.
Sources and provenance¶
https://github.com/GBeurier/nirs4all-methods/blob/main/cpp/src/core/augmentation/mixup/mixup.h; https://arxiv.org/abs/1710.09412
See also: methods index · interactive dashboard