Value and gradient of a scalar LazyRaster loss wrt a focal kernel.

loss must be a scalar pipeline (global sum or mean reduction) containing wrt, a focal_kernel() raster whose weights are the parameters. Executes chunk by chunk (gradients compose by linearity). Nodata is handled by a mask-multiply rewrite: nodata cells are zero-substituted in the inputs, a validity mask is carried through the pipeline, and the loss reduces over valid cells only, so gradients are never poisoned by NaN.

Usage

lazy_value_and_grad(loss, wrt, weights = NULL)

Arguments

loss

Scalar LazyRaster (reduced over x and y).

wrt

The focal_kernel() LazyRaster to differentiate against.

weights

Optional kernel matrix overriding the weights stored in wrt (used by optimisation loops to avoid rebuilding graphs).

Value

list(value = <scalar>, grad = <kernel-shaped matrix>).