This repository provides native TensorFlow execution in backend JavaScript applications under the Node.js runtime, accelerated by the TensorFlow C binary under the hood. It provides the same API as TensorFlow.js.

This package will work on Linux, Windows, and Mac platforms where TensorFlow is supported.

TensorFlow.js for Node currently supports the following platforms:

• Mac OS X CPU (10.12.6 Siera or higher)
• Linux CPU (Ubuntu 14.04 or higher)
• Linux GPU (Ubuntu 14.04 or higher and Cuda 10.0 w/ CUDNN v7) (see installation instructions)
• Windows CPU (Win 7 or higher)
• Windows GPU (Win 7 or higher and Cuda 10.0 w/ CUDNN v7) (see installation instructions)

For GPU support, tfjs-node-gpu@1.2.4 or later requires the following NVIDIA® software installed on your system:

Other Linux variants might also work but this project matches core TensorFlow installation requirements.

npm install @tensorflow/tfjs-node
(or)
yarn add @tensorflow/tfjs-node

npm install @tensorflow/tfjs-node-gpu
(or)
yarn add @tensorflow/tfjs-node-gpu

Windows & OSX build support for node-gyp requires Python 2.7. Be sure to have this version before installing @tensorflow/tfjs-node or @tensorflow/tfjs-node-gpu. Machines with Python 3.x will not install the bindings properly.

For more troubleshooting on Windows, check out WINDOWS_TROUBLESHOOTING.md.

If you do not have Xcode setup on your machine, please run the following commands:

$ xcode-select --install

After that operation completes, re-run yarn add or npm install for the @tensorflow/tfjs-node package.

You only need to include @tensorflow/tfjs-node or @tensorflow/tfjs-node-gpu in the package.json file, since those packages ship with @tensorflow/tfjs already.

To use this package on Raspberry Pi, you need to rebuild the node native addon with the following command after you installed the package:

$ npm rebuild @tensorflow/tfjs-node --build-from-source

Before executing any TensorFlow.js code, import the node package:

// Load the binding
import * as tf from '@tensorflow/tfjs-node';

// Or if running with GPU:
import * as tf from '@tensorflow/tfjs-node-gpu';

Note: you do not need to add the @tensorflow/tfjs package to your dependencies or import it directly.

# Download and install JS dependencies, including libtensorflow 1.8.
yarn

# Run TFJS tests against Node.js backend:
yarn test

# Switch to GPU for local development:
yarn enable-gpu

See the tfjs-examples repository for training the MNIST dataset using the Node.js bindings.

To get the most optimal TensorFlow build that can take advantage of your specific hardware (AVX512, MKL-DNN), you can build the libtensorflow library from source:

• Install bazel
• Checkout the main tensorflow repo and follow the instructions in here with one difference: instead of building the pip package, build libtensorflow:

./configure
bazel build --config=opt --config=monolithic //tensorflow/tools/lib_package:libtensorflow

The build might take a while and will produce a bazel-bin/tensorflow/tools/lib_package/libtensorflow.tar.gz file, which should be unpacked and replace the files in deps folder of tfjs-node repo:

cp bazel-bin/tensorflow/tools/lib_package/libtensorflow.tar.gz ~/myproject/node_modules/@tensorflow/tfjs-node/deps
cd path-to-my-project/node_modules/@tensorflow/tfjs-node/deps
tar -xf libtensorflow.tar.gz

If you want to publish an addon library with your own libtensorflow binary, you can host the custom libtensorflow binary and optional pre-compiled node addon module on the cloud service you choose, and add a custom-binary.json file in scripts folder with the following information:

{
  "tf-lib": "url-to-download-customized-binary",
  "addon": {
    "host": "host-of-pre-compiled-addon",
    "remote_path": "remote-path-of-pre-compiled-addon",
    "package_name": "file-name-of-pre-compile-addon"
  }
}

The installation scripts will automatically catch this file and use the custom libtensorflow binary and addon. If addon is not provided, the installation script will compile addon from source.