SetRandomValue

class SetRandomValue(sigma=0.05, clip=0.05, keys=('energy',))

Bases: object

Overwrite one or more keys with clipped Gaussian noise.

Replaces each listed key in place with freshly sampled Gaussian noise (std sigma, clamped to [-clip, clip]) of the same shape – the original values are discarded entirely (unlike the additive/multiplicative jitter transforms). Useful as an ablation that destroys the information in a feature while preserving its shape. Always applies (no probability gate). Requires each listed key to be present (raises ValueError otherwise). Registered as SetRandomValue – use this string as the type in a transform=[...] config list.

Parameters:

• sigma (float) – standard deviation of the replacement noise. Defaults to 0.05.

• clip (float) – magnitude the noise is clamped to (must be > 0). Defaults to 0.05.

• keys (str | Sequence[str]) – keys to overwrite. A bare string is wrapped into a single-element tuple. Defaults to ("energy",).

Example

>>> import numpy as np
>>> from pimm.datasets.transform import SetRandomValue
>>> np.random.seed(0)
>>> data = {"energy": np.array([[100.], [100.], [100.]], dtype="f4")}
>>> out = SetRandomValue(sigma=0.05, keys=("energy",))(data)
>>> out["energy"].shape  # same shape, original values discarded
(3, 1)
>>> bool(np.abs(out["energy"]).max() <= 0.05)  # overwritten with clipped noise
True