Reader and writer for the TensorFlow Record file format

Usage no npm install needed!

<script type="module">
  import tfrecordStream from '';


TensorFlow record (.tfrecord) File I/O for Node

Build Status NPM Version

The TFRecord format is briefly documented here, and described as the recommended format for feeding data into TenosorFlow here and here.

This library facilitates producing data in the TFRecord format directly in node.js. The library is not "official" - it is not part of TensorFlow, and it is not maintained by the TensorFlow team.


This module uses ES2017's async / await, so it requires node.js 7.6 or above.

While this module will presumably be used to interoperate with TensorFlow, it does not require a working TensorFlow installation.


The example below covers recommended API usage.

import { Reader, Writer } from 'tfrecord'

async function writeDemo() {
  const builder = tfrecord.createBuilder();
  builder.setInteger('answer', 42);
  builder.setFloat('pi', 3.14);
  builder.setBinary('name', new Uint8Array([65, 66, 67]));
  const example = builder.releaseExample();

  const writer = await Writer.createFromStream(fs.createWriteStream('data.tfrecord'));
  await writer.writeExample(example);
  await writer.close();

async function readDemo() {
  const reader = await Reader.createFromStream(fs.createReadStream('data.tfrecord'));
  let example;
  while (example = await reader.readExample()) {
    console.log('%j', example.toJSON());
  // The reader auto-closes after it reaches the end of the file.

async function demo() {
  await writeDemo();
  await readDemo();

let _ = demo();

The module also exposes the following low-level APIs:

  • tfrecord.RecordReader, tfrecord.RecordWriter - read/write files in the TensorFlow-flavored RecordIO format
  • tfrecord.Example - TensorFlow's Example protobuf, as compiled by protobuf.js
  • tfrecord.protos - the classes generated by compiling TensorFlow's protobuf definitions

The low-level APIs are exposed to make it easier to start working on an advanced use cases. While no current plan involves breaking these APIs, they might break more often than the high-level APIs.


Run the following command to populate the pre-generated files. These files are distributed in the npm package, but not checked into the git repository.


The test data can be regenerated by the following command, which requires a working TensorFlow installation on Python 3.

python3 scripts/

The test data is in the repository so we don't have to spend the time to install TensorFlow on Travis for every run.