ProbeDerivativesKind

class t3toolbox.backend.fitting.ProbeDerivativesKind

Bases: _WwPpSample, ProbeOutputKind, _DerivativesKind

Vector-valued derivative probing: the residual is a list of d order-leading arrays. sample = (ww, pp). Probe has both a mode and an order axis, so weight is the full omega[mode, order] matrix (a row = per-order, a column = per-mode, a matrix = both).

name = 'probe_derivatives'

chunk_size: t3toolbox.backend.common.typ.Optional[int] = 100

Methods

precompute(frame, s)	The reusable frame sweep at frame for sample (computed once per local model).
forward(v, s, frame, sweep)	The bare Jacobian J v (no gauge).
transpose(r, s, frame, sweep)	The bare transpose J^T r, summed over the sample stack W; raw (dU, dG).
point_forward(x_cores, s)	The point operation S(x) -- what the residual is measured against.
take(sample, data, idx)	Gather a flat W subset of (sample, data) -- the default minibatch draw's layout hook.