pyhgf.model.hybrid.linear_adapter#

pyhgf.model.hybrid.linear_adapter(linear)[source]#

Build a frozen linear map around an eqx.nn.Linear’s parameters.

Forward, x @ weight.T (+ bias) on the last axis. Backward, the error is multiplied back through the weight matrix (error @ weight) — the exact input error of a linear map, no derivation needed. The bias, being added, passes the error through untouched.