Implementation:Microsoft Onnxruntime Ort Tensor Nodejs

Field	Value
Implementation Name	Ort_Tensor_Nodejs
Overview	Construction of typed multi-dimensional tensors from JavaScript TypedArrays for ONNX Runtime inference.
Type	API Doc
Language	JavaScript
Domains	ML_Inference, JavaScript_Integration
Source Repository	microsoft/onnxruntime
Last Updated	2026-02-10

Overview

Construction of typed multi-dimensional tensors from JavaScript TypedArrays for ONNX Runtime inference. The ort.Tensor constructor wraps a TypedArray or string array with type and shape metadata to create the fundamental data container used for both inputs and outputs of ONNX model inference.

API

new ort.Tensor(type?: string, data: TypedArray | string[], dims?: number[])

Import

const ort = require('onnxruntime-node');
// or
const { Tensor } = require('onnxruntime-node');

Source Code Reference

Repository: microsoft/onnxruntime
Primary Source: samples/nodejs/02_create-tensor/index.js:L9-37

Key Parameters

Parameter	Type	Required	Description
type	string	No	Tensor element type: 'float32', 'float64', 'int32', 'uint8', 'bool', 'string', etc. Can be omitted if inferable from data.
data	TypedArray / string[]	Yes	The underlying data buffer. Must match the declared type.
dims	number[]	No	Tensor shape as an array of dimension sizes. Use [] for scalar, omit for 1-D tensor.

I/O Contract

Direction	Name	Type	Description
Input	type	string (optional)	Element type identifier for the tensor data
Input	data	TypedArray / string[]	Raw data buffer with numerical or string values
Input	dims	number[] (optional)	Shape dimensions; product must equal data length
Output	tensor	ort.Tensor	Tensor object with .data, .dims, .type, and .size properties

Usage Examples

Float32 Tensor with Explicit Type

const ort = require('onnxruntime-node');

// Create a [2, 3, 4] float32 tensor (24 elements total)
const buffer = new Float32Array(24);
buffer[0] = 0.1; // fill buffer data
const tensor = new ort.Tensor('float32', buffer, [2, 3, 4]);
console.log(tensor.type);  // 'float32'
console.log(tensor.dims);  // [2, 3, 4]
console.log(tensor.size);  // 24

Type Inferred from TypedArray

const ort = require('onnxruntime-node');

// Type 'float32' is inferred from Float32Array
const tensor = new ort.Tensor(new Float32Array(24), [2, 3, 4]);
console.log(tensor.type);  // 'float32'

Boolean Tensor

const ort = require('onnxruntime-node');

// Boolean tensor requires explicit type since both 'bool' and 'uint8' use Uint8Array
const boolData = new Uint8Array(2);
boolData[0] = 1;  // true
boolData[1] = 0;  // false
const tensor = new ort.Tensor('bool', boolData, [1, 2]);

Scalar Tensor

const ort = require('onnxruntime-node');

// Scalar tensor with empty dims array
const scalar = new ort.Tensor(new Float64Array(1), []);
scalar.data[0] = 1.0;  // setting data after creation is allowed
console.log(scalar.dims);  // []
console.log(scalar.size);  // 1

1-D Tensor (Dims Omitted)

const ort = require('onnxruntime-node');

// When dims are omitted, a 1-D tensor is created
const tensor = new ort.Tensor(new Float32Array(100));
console.log(tensor.dims);  // [100]

String Tensor

const ort = require('onnxruntime-node');

// String tensors use a JavaScript string array as data
const strTensor = new ort.Tensor('string', ['hello', 'world'], [1, 2]);
// Type 'string' can also be inferred
const strTensor2 = new ort.Tensor(['hello', 'world'], [1, 2]);

Key Details

The constructor validates that data size matches the product of dimensions.
Type/data mismatches (e.g., 'float64' type with Float32Array data) throw an error.
Negative dimension values are not allowed and will throw an error.
The .data property of the resulting tensor provides direct access to the underlying TypedArray.
Setting values via .data[index] after construction is permitted.

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment