Pre-trained neural nets for the browser, no PhD required

What it does This repository hosts a collection of pre-trained machine learning models ported to TensorFlow.js, spanning image classification, object detection, pose and face tracking, speech recognition, text embedding, and toxicity scoring. Each model is packaged as a standalone module served from NPM and unpkg, intended to be imported directly into browser or Node projects.

The interesting bit The maintainers deliberately hide tensors from the public APIs so non-machine-learning experts can use them. The collection is explicitly framed as composable building blocks rather than a monolithic framework, with a KNN classifier utility included for transfer learning experiments.

Key highlights

• Broad coverage: MobileNet, COCO-SSD, DeepLab v3, hand/pose/face detection, portrait depth, speech commands, universal sentence encoder, and text toxicity scoring
• Browser-first deployment with live demos available for most (though not all) models
• KNN classifier utility supports transfer learning without custom training pipelines
• Each model lives in its own directory with isolated API docs and demo source

Caveats

• The top-level README is strictly a catalog; model-specific behavior, input requirements, and performance details require diving into individual sub-directory READMEs
• Several model entries (COCO-SSD, DeepLab v3, Universal Sentence Encoder) have blank live demo links in the index

Verdict A solid starting point if you need off-the-shelf ML primitives for a web app. Less useful if you want a single, deeply integrated toolkit or unified documentation.