Implementation:Tensorflow Tfjs CPU Conv3DBackpropInputV2 Kernel

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

Overview

The Conv3DBackpropInputV2 kernel computes the gradient of a 3D convolution with respect to the input tensor on the CPU backend. It extends the 2D backpropagation pattern into three spatial dimensions (depth, height, width), iterating over the upstream gradient and transposed filter to accumulate the input gradient. This kernel is used during backpropagation through conv3d layers.

Code Reference

Source Location

tfjs-backend-cpu/src/kernels/Conv3DBackpropInputV2.ts (GitHub)

Signature

export function conv3DBackpropInputV2(args: {
  inputs: Conv3DBackpropInputV2Inputs,
  backend: MathBackendCPU,
  attrs: Conv3DBackpropInputV2Attrs
}): TensorInfo

Imports

import {backend_util, Conv3DBackpropInputV2, Conv3DBackpropInputV2Attrs, Conv3DBackpropInputV2Inputs, KernelConfig, KernelFunc, TensorBuffer, TensorInfo, TypedArray, util} from '@tensorflow/tfjs-core';
import {MathBackendCPU} from '../backend_cpu';
import {assertNotComplex} from '../cpu_util';

Kernel Registration

export const conv3DBackpropInputV2Config: KernelConfig = {
  kernelName: Conv3DBackpropInputV2,
  backendName: 'cpu',
  kernelFunc: conv3DBackpropInputV2 as unknown as KernelFunc
};

I/O Contract

Parameter	Type	Description
inputs.dy	TensorInfo	Upstream gradient tensor from the forward conv3d output
inputs.filter	TensorInfo	The convolution filter of shape [filterDepth, filterHeight, filterWidth, inChannels, outChannels]
attrs.inputShape	[number, number, number, number, number]	Shape of the original input tensor
attrs.strides	[number, number, number]	Stride of the convolution
attrs.pad	'same' | number	Padding mode
Return	TensorInfo	The gradient tensor dx with shape matching inputShape

Implementation Details

• Asserts that dy is not a complex tensor.
• Computes 3D convolution metadata via backend_util.computeConv3DInfo with dilation fixed at 1.
• Allocates a TensorBuffer for the output gradient dx.
• Precomputes stride arrays for dx, dy, and the filter for efficient flat-index computation.
• Computes padding offsets: frontPad = filterDepth - 1 - padInfo.front, topPad = filterHeight - 1 - padInfo.top, leftPad = filterWidth - 1 - padInfo.left.
• Uses seven nested loops (batch, inChannels, inDepth, inHeight, inWidth, plus output spatial dimensions) to accumulate dot products.
• Filter indices are reversed (e.g., filterDepth - 1 - wF) to implement the transposed convolution.
• Each output gradient position contributes to the input gradient weighted by the corresponding filter value across all output channels.

Usage Examples

import {conv3DBackpropInputV2} from './Conv3DBackpropInputV2';

const dx = conv3DBackpropInputV2({
  inputs: {dy: dyTensor, filter: filterTensor},
  backend: cpuBackend,
  attrs: {
    inputShape: [1, 8, 8, 8, 3],
    strides: [1, 1, 1],
    pad: 'same'
  }
});

Related Pages

• Principle:Tensorflow_Tfjs_CPU_Kernel_Registration