fetch issue running pre-trained USE model in Node.js

Had an issue with fetch when loading a pre-trained TensorFlow.js model in a Node.js app.

System information

• Mac OS Mojave - 10.14.1
• Node.js - 10.15.1
• relevant npm packages:
  • @tensorflow-models/universal-sentence-encoder@1.0.1
  • @tensorflow/tfjs@1.0.0
  • @tensorflow/tfjs-node@1.0.1
  • express@4.16.4
  • node-fetch@2.3.0

Issue

I’m using the pre-trained USE model in a simple Express server endpoint—it basically just returns the text embedding for an input string.

The relevant source (which lives in a separate file called by the endpoint):

// use tensorflow js node backend
require('@tensorflow/tfjs-node');
const universalSentenceEncoder = require('@tensorflow-models/universal-sentence-encoder');

/**
 * Encode a variable length string as a sentence embedding via the Universal Sentence Encoder.
 *
 * @param {*} text - the input string
 */
exports.embed = async function embed(text) {
  const result = universalSentenceEncoder.load().then(model => model.embed(text));
  return result;
};

This appears to be related to #489, but my issue remained.

Calling the universalSentenceEncoder.load() function causes a ReferenceError: fetch is not defined error.

This appears to be the offending line: https://github.com/tensorflow/tfjs-models/blob/master/universal-sentence-encoder/src/index.ts#L49

It seems odd to me that tfjs-node requires the node-fetch package and yet still required a separate package to fix this.

The workaround I am using is overriding the global fetch with a line like this at the top of the file:

global.fetch = require('node-fetch');

[annxingyuan]

@alexellis yes good point - thanks for pointing this out! We’ll patch this soon. Meanwhile you can override global fetch as a bandaid, but hopefully not for too much longer 😃

[alexellis]

It’d be ideal if all examples worked on Node as well as in the browser.