DriftEngineDriftTexture · function
halfPrecisionErrorBound
How far the half-precision evaluation of this network, at this input, can sit from the single-precision one.
Explained in DriftTexture.
function halfPrecisionErrorBound(shape: NetworkShape, weights: Float32Array, input: Float32Array): numberimport { halfPrecisionErrorBound } from '@driftengine/texture';Parameters
| Parameter | Type | Description |
|---|---|---|
shape | NetworkShape | |
weights | Float32Array | |
input | Float32Array |
In depth
A bound per network rather than one tolerance for all, because a sampled tolerance is a claim
about the corpus that sampled it — 2,400 networks put the worst relative error at 0.00213, the
device's own corpus reached 0.0039, and 40,000 reached 0.0066. This carries the error instead:
the weights' and the input's own rounding exactly, then half a unit in the last place for every
product, every partial sum and the bias, through each layer — a rectifier moves no error, since
it is one-Lipschitz. It holds whichever form a device computes: every operation rounded, or
each multiply-add contracted into a single rounding, which WGSL permits and this project's
development machine does. It assumes no value overflows, which activationBound is for.