transcranial/keras-js

Converts Keras models trained in any backend (TensorFlow, CNTK, etc.) to browser-compatible format with WebGL 2-accelerated inference on the client side. Implements layer-by-layer operations in JavaScript, supporting convolutional networks, LSTMs, and attention-based architectures demonstrated across image classification, generative models, and NLP tasks. Also enables CPU-mode execution in Node.js for server-side inference without GPU dependencies.

4,968 stars and 1,366 monthly downloads. No commits in the last 6 months. Available on npm.