Skip to contents

Whole-window model op: fn over the raw padded chunk. Halo-consuming.

PatchNode.Rd

Created by lazy_patch(). The stencil shape for kernels whose spatial context is far beyond a shift-list focal (CNN inference, e.g. OmniCloudMask): fn(xpad) receives the parent's value carrying at least radius halo cells per side (a (C, H, W) cube, traced or plain) and returns a SIZE-PRESERVING result on the same spatial dims, deriving every size from the input shape (chunk shapes vary); the evaluator crops the radius ring afterwards, exactly as focal shifts consume theirs. Output rank: 2 when out_bands == 0 (the band axis is consumed), else 3 with a leading band axis of length out_bands.

Usage

PatchNode(
  id = integer(0),
  parents = integer(0),
  grid = GridSpec(),
  role = character(0),
  fn = function() NULL,
  radius = integer(0),
  out_bands = integer(0),
  dtype = character(0),
  kernel_id = character(0),
  bytes_px = integer(0),
  flops_px = integer(0)
)

Arguments

id

Integer node id (assigned by graph_add()).

parents

Integer ids of parent nodes (length 1).

grid

Output GridSpec of this node.

role

Optional semantic role tag (e.g. "mask", set by mask()). Pure metadata: never read by the planner or executors; surfaced by draw() and plan_view().

fn

The model body fn(xpad).

radius

Halo consumed per side, pixels.

out_bands

Output band count; 0 drops the band axis.

dtype

Length-0 (parent's dtype) or length-1 dtype override.

kernel_id

Content hash standing in for fn in kernel signatures.

bytes_px

Resident working-set estimate, bytes per core pixel.

flops_px

Compute estimate, flops per core pixel.

Value

A PatchNode.

Details

kernel_id is the content identity of the closed-over model (weights

  • configuration): it stands in for fn in kernel signatures, so fn itself is never serialised or hashed (a model closure can carry tens of MB). Two nodes with equal kernel_id are treated as the same kernel, so kernel_id must change whenever the model content changes. bytes_px / flops_px are planner cost hints per CORE pixel; they price chunk sizing, memory admission, and placement, which cannot introspect an arbitrary model closure.