@tensorflow-models/tasks
TypeScript icon, indicating that this package has built-in type declarations

0.0.1-alpha.8 • Public • Published

TFJS Task API

WORK IN PROGRESS

TFJS Task API provides an unified experience for running task-specific models on the Web. It is designed with ease-of-use in mind, aiming to improve usability for JS developers without ML knowledge. It has the following features:

  • Easy-to-discover models

    Models from different runtime systems (e.g. TFJS, TFLite, MediaPipe, etc) are grouped by popular ML tasks, such as sentiment detection, image classification, pose detection, etc.

  • Clean and powerful APIs

    Different tasks come with different API interfaces that are the most intuitive to use for that particular task. Models under the same task share the same API, making it easy to explore. Inference can be done within just 3 lines of code.

  • Simple installation

    You only need to import this package (<20K in size) to start using the API without needing to worry about other dependencies, such as model packages, runtimes, backends, etc. They will be dynamically loaded on demand without duplication.

The following table summarizes all the supported tasks and their models:

Task Model Supported runtimes · Docs · Resources
Image Classification
Identify images into predefined classes.
Demo
Mobilenet
TFJS   · API doc
TFLite · API doc
Custom model
Object Detection
Localize and identify multiple objects in a single image.
Demo
COCO-SSD
TFJS   · API doc
TFLite · API doc
Custom model
Image Segmentation
Predict associated class for each pixel of an image.
Demo
Deeplab
TFJS   · API doc
TFLite · API doc
Custom model
Sentiment Detection
Detect pre-defined sentiments in a given paragraph of text.
Demo
Toxicity
TFJS   · API doc
Movie review
TFLite · API doc
NL Classification
Identify texts into predefined classes.
Demo
Custom model
Question & Answer
Answer questions based on the content of a given passage.
Demo
BertQA
TFJS   · API doc
TFLite · API doc

(The initial version only supports the web browser environment. NodeJS support is coming soon)

Usage

Import the package

This package is all you need. The packages required by different models will be loaded on demand automatically.

Via NPM

// Import @tensorflow-models/tasks.
import * as tfTask from '@tensorflow-models/tasks';

Via a script tag

<!-- Import @tensorflow-models/tasks -->
<script src="https://cdn.jsdelivr.net/npm/@tensorflow-models/tasks"></script>

Load model and run inference

The code snippet below shows how to load various models for the Image Classification task:

import * as tfTask from '@tensorflow-models/tasks';

// Load the TFJS mobilenet model.
const model1 = await tfTask.ImageClassification.MobileNet.TFJS.load({
  backend: 'wasm'});

// Load the TFLite mobilenet model.
const model2 = await tfTask.ImageClassification.MobileNet.TFLite.load();

// Load a custom image classification TFLite model.
const model3 = await tfTask.ImageClassification.CustomModel.TFLite.load({
  model: 'url/to/your/bird_classifier.tflite'});

Since all these models are for the Image Classification task, they will have the same task model type: ImageClassifier in this case. Each task model's predict inference method has an unique and easy-to-use API interface. For example, in ImageClassifier, the method takes an image-like element and returns the predicted classes:

const result = model1.predict(document.querySelector(img)!);
console.log(result.classes);

TFLite custom model compatibility

TFLite is supported by the @tensorflow/tfjs-tflite package that is built on top of the TFLite Task Library and WebAssembly. As a result, all TFLite custom models should comply with the metadata requirements of the corresonding task in the TFLite task library. Check out the "model compatibility requirements" section of the official task library page. For example, the requirements of ImageClassifier can be found here.

See an example of how to use TFLite custom model in the Load model and run inference section above.

Advanced Topics

Performance

For TFJS models, the choice of backend affects the performance the most. For most cases, the WebGL backend (default) is usually the fastest.

For TFLite models, we use WebAssembly under the hood. It uses XNNPACK to accelerate model inference. To achieve the best performance, use a browser that supports "WebAssembly SIMD" and "WebAssembly threads". In Chrome, these can be enabled in chrome://flags/. The task API will automatically choose the best WASM module to load and set the number of threads for best performance based on the current browser environment.

As of March 2021, XNNPACK works best for non-quantized TFLite models. Quantized models can still be used, but XNNPACK only supports ADD, CONV_2D, DEPTHWISE_CONV_2D, and FULLY_CONNECTED ops for models with quantization-aware training using TF MOT.

Development

Building

$ yarn
$ yarn build

Testing

$ yarn test

Deployment

$ yarn build-npm
# (TODO): publish

Readme

Keywords

none

Package Sidebar

Install

npm i @tensorflow-models/tasks

Weekly Downloads

205

Version

0.0.1-alpha.8

License

Apache-2.0

Unpacked Size

1.98 MB

Total Files

161

Last publish

Collaborators

  • fengwuyao
  • linchan
  • caisq
  • pyu10055
  • annxingyuan
  • mattsoulanille
  • linazhao128
  • jinjingforever