Jitter

rf.Jitter adds bounded noise to visible continuous inputs during training. Evaluation and prediction use unperturbed values. It changes neither stored inputs nor reconstruction targets.

import relflow as rf

measurement = rf.Number(
    jitter=rf.Jitter(add=0.1, multiply=0.02, normalize=False),
)

Here additive noise uses the measurement’s source units, while multiplication models proportional uncertainty. Jitter is regularization; validate its magnitude against the distinctions the task needs to preserve.

Options

Option	Default	Meaning
add	0.0	Maximum absolute additive displacement.
multiply	0.0	Maximum multiplicative deviation from one.
normalize	True	Use the normalized boundary when the field has one.

Both magnitudes must be finite and nonnegative. The configuration is frozen; construct a new value to change it. rf.Jitter() disables noise and consumes no random numbers.

Each enabled component uses independent triangular noise: the difference of two uniform draws on [0, 1]. Small perturbations are more likely than extremes. With both components enabled:

\[ J(v) = v[1 + m(U_1-U_2)] + a(U_3-U_4) \]

The additive range is [-add, add]; the scale range is [1 - multiply, 1 + multiply]. No application-domain clipping is performed, and multiply > 1 can reverse a sign.

Where it acts

Field	Continuous representation
Number	Standardized values by default; raw values before normalization when normalize=False.
Vector	Supplied coordinates before projection; no learned normalizer.
DateParts	Sine/cosine pairs, without moving the source timestamp or renormalizing the pair.
Text	Pooled frozen-encoder embedding before relflow projection.

Only Number has separate raw and standardized boundaries. For the other types, both settings of normalize use the same coordinate space. Choose add in that space’s units.

Noise applies only to finite coordinates of valued training inputs. Nulls, padding, learned-mask representations, skipped values, and target embeddings remain unchanged. Number’s normalizer also learns from pristine values.

Torch supplies the random draws. Reproducing them requires the same seed, device, batch geometry, and execution order. Jitter is not a persisted data transformation.

For custom tensorfields, jitter.apply(inputs, mask) owns the tensor operation; the consumer owns training-mode gating, coordinate eligibility, normalization placement, and pristine targets. See Custom tensorfields.