Implementation:Tensorflow Tfjs CPU SparseFillEmptyRows Impl

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

Overview

The sparseFillEmptyRowsImpl function fills empty rows in a sparse tensor with a default value. Given a 2D sparse tensor represented by indices and values, it ensures every row (in the first dimension up to denseShape[0]) has at least one entry. Empty rows are filled with a single entry at column 0 containing the provided default value. It also returns metadata needed for backpropagation.

Code Reference

Source Location

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

Signature

export function sparseFillEmptyRowsImpl(
    indices: TypedArray, indicesShape: number[], indicesDType: DataType,
    values: TypedArray, valuesDType: DataType, denseShape: TypedArray,
    defaultValue: number):
    [TypedArray, number[], TypedArray, boolean[], number[]]

Imports

import {backend_util, DataType, TypedArray, util} from '@tensorflow/tfjs-core';

I/O Contract

Inputs

Name	Type	Description
indices	TypedArray	Flat 2D indices of shape [N, rank] for the sparse tensor
indicesShape	number[]	Shape of the indices tensor ([N, rank])
indicesDType	DataType	Data type of the indices
values	TypedArray	Sparse tensor values of length N
valuesDType	DataType	Data type of the values
denseShape	TypedArray	Dense shape of the sparse tensor
defaultValue	number	Value used to fill empty rows

Output

Returns a tuple of five elements:

• outputIndices (TypedArray): New indices with filled rows included.
• outputIndicesShape (number[]): Shape of the output indices ([M, rank]).
• outputValues (TypedArray): New values with default values for filled rows.
• emptyRowIndicator (boolean[]): Boolean array of length denseRows indicating which rows were empty.
• reverseIndexMap (number[]): Maps each original index to its position in the output, needed for gradient computation.

Algorithm

1. Edge case: If denseRows === 0, returns empty arrays (with validation that indicesCount is also 0).
2. CSR offset computation: Counts elements per row. Validates row indices are non-negative and within bounds.
3. Fast path: If all rows are already filled and rows are ordered, returns the original indices/values unchanged with an identity reverse index map.
4. Slow path:
   1. Computes cumulative CSR offsets to determine output positions.
   2. Allocates output arrays sized to fullIndicesCount (sum of max(1, count) per row).
   3. Copies existing entries to their correct positions in the output.
   4. For each empty row, inserts a single entry with the row index and default value.

Usage Example

import {sparseFillEmptyRowsImpl} from './SparseFillEmptyRows_impl';

const indices = new Int32Array([0, 1, 2, 3]);  // shape [2, 2], rows 0 and 2
const values = new Float32Array([10, 20]);
const denseShape = new Int32Array([4, 5]);

const [outIdx, outIdxShape, outVals, emptyRows, reverseMap] =
    sparseFillEmptyRowsImpl(
        indices, [2, 2], 'int32',
        values, 'float32',
        denseShape, 0);
// emptyRows: [false, true, false, true]

Related Pages

• Principle:Tensorflow_Tfjs_CPU_Kernel_Registration