Base64-to-Tensor

This package enables the conversion of Base64-encoded images to tensor objects using pure JavaScript, compatible with TensorFlow.js.

Installation

Install the package using npm:

1npm install base64-to-tensor --save

Prerequisites

Ensure @tensorflow/tfjs-core is installed alongside a valid TensorFlow backend. Choose between the synchronous package jpeg-js for full blocking sync or the asynchronous package sharp for non-blocking async operations:

1# For synchronous operations:
2npm install @tensorflow/tfjs-core jpeg-js
3
4# For asynchronous operations:
5npm install @tensorflow/tfjs-core sharp

Usage

Refer to the convert.test.ts file for example usage. Below are snippets demonstrating both synchronous and asynchronous conversions:

1import { convert, convertAsync } from "base64-to-tensor";
2import { setBackend } from "@tensorflow/tfjs-core";
3import "@tensorflow/tfjs-backend-wasm";
4
5await setBackend("wasm");
6
7// Synchronous conversion (jpeg-js)
8const tensorSync = convert(mybase64); // Ensure mybase64 is a valid JPEG
9
10// Asynchronous conversion (sharp)
11const tensorAsync = await convertAsync(mybase64); // Enhanced performance
12
13// Example tensor output
14{
15  kept: false,
16  isDisposedInternal: false,
17  shape: [189, 300, 3],
18  dtype: "int32",
19  size: 170100,
20  strides: [900, 3],
21  dataId: { id: 1 },
22  id: 1,
23  rankType: "3",
24}

Benefits

Using pure JavaScript for image conversion to tensors offers several advantages:

1. Reduced Size and Portability: Eliminates the need for cairo or other native image development converters.
2. Increased Speed: Performs calculations on-the-fly without external call overheads.
3. Worker Thread Compatibility: Facilitates the use of TensorFlow WASM backends within API services, enhancing performance and scalability.

Benchmarks

Performance Benchmarks

Performance benchmarks conducted on a Mac M1 (64GB RAM):

Test Case	Characters	Size	Synchronous	Asynchronous
JPEG	26,791	26.16 KB	100 ms	50 ms