errors executing async models loaded via tf.loadGraphModel.

TensorFlow.js version

"@tensorflow-models/coco-ssd": "^1.0.0",
"@tensorflow-models/mobilenet": "^1.0.0",
"@tensorflow-models/universal-sentence-encoder": "1.0.1",
"@tensorflow/tfjs": "1.0.0",
"@tensorflow/tfjs-node": "1.0.1",

Error also occurs on tfjs 1.0.2

Node version

Node 11.11

Describe the problem or feature request

The following error occurs when making a prediction with either coco-ssd or universal-sentence-encoder (which are both models with control flow ops). But doesn’t happen with mobilenet.

node:59415) UnhandledPromiseRejectionWarning: TypeError: Cannot read property 'id' of undefined
    at /Users/yassogba/projects/intent-classifier/node_modules/@tensorflow/tfjs-converter/dist/src/executor/graph_executor.js:295:99
    at Array.map (<anonymous>)
    at GraphExecutor.<anonymous> (/Users/yassogba/projects/intent-classifier/node_modules/@tensorflow/tfjs-converter/dist/src/executor/graph_executor.js:295:58)
    at step (/Users/yassogba/projects/intent-classifier/node_modules/@tensorflow/tfjs-converter/dist/src/executor/graph_executor.js:56:23)
    at Object.next (/Users/yassogba/projects/intent-classifier/node_modules/@tensorflow/tfjs-converter/dist/src/executor/graph_executor.js:37:53)
    at fulfilled (/Users/yassogba/projects/intent-classifier/node_modules/@tensorflow/tfjs-converter/dist/src/executor/graph_executor.js:28:58)
    at processTicksAndRejections (internal/process/next_tick.js:81:5)

From doing some console logging it looks like the tensors in the graph are undefined during executeAsync.

Note that I only get this error when using the node backend (tfjs-node), if i use the vanilla cpu backend (tfjs) it works.

Code to reproduce the bug / link to feature request

// const tf = require('@tensorflow/tfjs'); // using this import works
const tf = require('@tensorflow/tfjs-node');
const cocossd = require('@tensorflow-models/coco-ssd');
const mob = require('@tensorflow-models/mobilenet');
global.fetch = require('node-fetch');
const useLoader = require('@tensorflow-models/universal-sentence-encoder');

// This works
const model = await mob.load();
const res = await model.classify(tf.randomNormal([224, 224, 3]));
console.log(res);

// This fails.
const cocomodel = await cocossd.load();
const cocoa = await cocomodel.detect(tf.randomNormal([224, 224, 3]));
console.log(cocoa)
  
// this fails
console.time('Loading Universal Sentence Encoder');
const use = await useLoader.load();
console.timeEnd('Loading Universal Sentence Encoder');

[tafsiri]

@pyu10055 This repo code demonstrates it https://github.com/tafsiri/use-text-classifier/blob/master/training/train.js, update the package.json to remove the other models. However on your point about that I thought all the models specify tfjs as a peerDependency, so they won’t install their own copy.

Also see https://github.com/tensorflow/tfjs/issues/1454

[tafsiri]

No longer an issue on latest tfjs.