Implementation:Tensorflow Tfjs CPU ResizeNearestNeighborGrad Kernel

Knowledge Sources	Tensorflow_Tfjs
Domains	Deep_Learning, CPU_Backend
Last Updated	2026-02-10 06:00 GMT

Overview

The ResizeNearestNeighborGrad kernel computes the gradient of nearest-neighbor image resizing for the CPU backend. It determines which input pixels contributed to each output pixel in the forward pass and accumulates the upstream gradient values back to those input positions.

Code Reference

Source Location

tfjs-backend-cpu/src/kernels/ResizeNearestNeighborGrad.ts (GitHub)

Signature

export function resizeNearestNeighborGrad(args: {
  inputs: ResizeNearestNeighborGradInputs,
  backend: MathBackendCPU,
  attrs: ResizeNearestNeighborGradAttrs
}): TensorInfo

Imports

import {KernelConfig, KernelFunc, ResizeNearestNeighborGrad, ResizeNearestNeighborGradAttrs, ResizeNearestNeighborGradInputs, TensorInfo, TypedArray, util} from '@tensorflow/tfjs-core';
import {MathBackendCPU} from '../backend_cpu';
import {assertNotComplex} from '../cpu_util';

Kernel Config

export const resizeNearestNeighborGradConfig: KernelConfig = {
  kernelName: ResizeNearestNeighborGrad,
  backendName: 'cpu',
  kernelFunc: resizeNearestNeighborGrad as unknown as KernelFunc
};

I/O Contract

Inputs

Name	Type	Description
dy	TensorInfo	Upstream gradient of shape [batch, yHeight, yWidth, depth]
images	TensorInfo	Original input images of shape [batch, xHeight, xWidth, depth]

Attributes

Name	Type	Description
alignCorners	boolean	If true, aligns the corner pixels of input and output

Output

Returns a TensorInfo with the same shape and dtype as images, containing accumulated gradients.

Algorithm

1. Computes effective input and output sizes with alignCorners adjustments.
2. Calculates forward and inverse height/width scaling ratios.
3. Computes search window dimensions: winHeight = ceil(invHeightScale) * 2 + 2 and winWidth = ceil(invWidthScale) * 2 + 2.
4. For each input pixel [b, r, c, d]:
   1. Determines a search window in dy centered around the expected contributing output pixels.
   2. For each dy pixel in the search window:
      1. Computes the nearest-neighbor source row/column using Math.round (if alignCorners) or Math.floor.
      2. If the source row and column match the current input position, accumulates the dy value.
   3. Stores the accumulated gradient in the output.

Usage Example

This kernel is invoked automatically during backpropagation through tf.image.resizeNearestNeighbor:

import * as tf from '@tensorflow/tfjs-core';
import '@tensorflow/tfjs-backend-cpu';

const image = tf.tensor4d([1, 2, 3, 4], [1, 2, 2, 1]);
const grads = tf.grads(
    (img) => tf.image.resizeNearestNeighbor(img, [4, 4]).sum());
const [dx] = grads([image]);
dx.print();

Related Pages

• Principle:Tensorflow_Tfjs_CPU_Kernel_Registration