Implementation:Tensorflow Tfjs CPU ResizeBilinearGrad Kernel

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

Overview

The ResizeBilinearGrad kernel computes the gradient of bilinear image resizing for the CPU backend. During backpropagation, it distributes each gradient value from the resized output (dy) back to the four source pixels in the original image that contributed to it, weighted by the same bilinear interpolation coefficients used in the forward pass. The reference implementation follows the TensorFlow C++ kernel.

Code Reference

Source Location

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

Signature

export function resizeBilinearGrad(args: {
  inputs: ResizeBilinearGradInputs,
  backend: MathBackendCPU,
  attrs: ResizeBilinearGradAttrs
}): TensorInfo

Imports

import {KernelConfig, KernelFunc, ResizeBilinearGrad, ResizeBilinearGradAttrs, ResizeBilinearGradInputs, TensorInfo, TypedArray, util} from '@tensorflow/tfjs-core';
import {MathBackendCPU} from '../backend_cpu';
import {assertNotComplex} from '../cpu_util';

Kernel Config

export const resizeBilinearGradConfig: KernelConfig = {
  kernelName: ResizeBilinearGrad,
  backendName: 'cpu',
  kernelFunc: resizeBilinearGrad 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 float32 dtype and shape [batch, xWidth, xHeight, depth], containing the accumulated gradients for the original image pixels.

Algorithm

1. Computes effective input (xHeight, xWidth) and output (yHeight, yWidth) sizes with alignCorners adjustments.
2. Calculates heightScale and widthScale ratios.
3. For each output pixel in dy at position [b, r, c]:
   1. Computes the corresponding fractional source position (dxR, dxC).
   2. Determines floor and ceil indices for both row and column, clamped to valid bounds.
   3. Computes four interpolation weights: inverseDxRLerp * inverseDxCLerp, inverseDxRLerp * dxCLerp, dxRLerp * inverseDxCLerp, dxRLerp * dxCLerp.
   4. For each depth channel, distributes the dy value to the four corresponding input positions weighted by these coefficients (accumulating via +=).

Usage Example

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

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.resizeBilinear(img, [4, 4]).sum());
const [dx] = grads([image]);
dx.print();

Related Pages

• Principle:Tensorflow_Tfjs_CPU_Kernel_Registration