Getting error message in console: Initialization of backend webgpu failed

[jht9629-nyu]

Issue Type

Runtime Error, Unexpected Behavior

ml5.js Version

v.1.2.0 Latest

Development Environment

p5js editror

Browser & Operating System

Chrome Macos Version 131.0.6778.69 (Official Build) (arm64)

Issue Description

From documentation:
https://docs.ml5js.org/#/reference/bodypose

run the example sketch:
https://editor.p5js.org/ml5/sketches/hMN9GdrO3
Following error is displayed in console:
Initialization of backend webgpu failed
index.js:259 TypeError: t.requestAdapterInfo is not a function
at Object.factory (pose-detection.esm.js:17:97544)
at async b.ready (engine.js:93:13)
at async fG.loadModel (index.js:256:7)

Minimal Reproducible Example

No response

Code of Conduct

• I agree to follow this project's Code of Conduct

[jht9629-nyu]

Today the error is no longer been displayed, but this may be indicative of a bigger issue where sub components of ml5 library are changing that are not part of the npm library delivery

[TimChen3345]

I am running into the same issue:
https://editor.p5js.org/tc3902/sketches/7h4WyQQe7

Here is the error I got in console:

Initialization of backend webgpu failed
TypeError: t.requestAdapterInfo is not a function
at Object.factory (https://unpkg.com/ml5@1/dist/ml5.min.js:18352:117148)
at async b.ready (https://unpkg.com/ml5@1/dist/ml5.min.js:102:916)
at async fG.loadModel (https://unpkg.com/ml5@1/dist/ml5.min.js:18359:1917)

I am running in Google Chrome Version 131.0.6778.86 (Official Build) (64-bit)

Processor: AMD Ryzen 9 7845HX with Radeon Graphics 3.00 GHz
Edition Windows 11 Home
Version 23H2
Installed on ‎8/‎16/‎2024
OS build 22631.4460
Experience Windows Feature Experience Pack 1000.22700.1047.0

[shiffman]

I've been noticing this issue recently as well, it seems to come and go, it's very strange! I wonder if it's something related to chrome/webgpu update? It resolves if you force a non-webgpu backend, for example:

ml5.setBackend('webgl');

[TimChen3345]

It works! Thanks!

[jht9629-nyu]

Error "Initialization of backend webgpu failed"
is back!
Run this sketch:
https://editor.p5js.org/ml5/sketches/hMN9GdrO3
On Chrome Version 131.0.6778.86 (Official Build) (arm64
macOS 14.7.1

[shiffman]

Strange! For me, I get the "error" but the sketch still runs. @jht9629-nyu if you add ml5.setBackend('webgl') before you load the model does that resolve it for you? (It does for me.)

function preload() {
  // Load the bodyPose model
  ml5.setBackend("webgl");
  bodyPose = ml5.bodyPose();
}

@ziyuan-linn do you have any ideas about this? It seems something has changed with webgpu?

[jht9629-nyu]

Error message does not appear to affect behavior,
my sketches still work with it.
Using ml5.setBackend("webgl") as @shiffman recommends makes the error message go away.

[ziyuan-linn]

@shiffman When webgpu backend fails, tfjs automatically rolls back to webgl. I tried logging the backend after the warning:

The tfjs webgpu backend called the requestAdapterInfo function. This function is deprecated and non-standard, so Chrome might have removed it during a recent update.

It looks like this issue has already been addressed by tfjs in tensorflow/tfjs#8392, which is released in tfjs-v4.22.0. I think bumping our dependency versions should fix the issue.

This might also be a good opportunity to look into bringing back Dependabot, which automatically checks for dependency updates.

[shiffman]

Thank you @ziyuan-linn! I re-enabled dependabot in the repo as well as updated tensorflow.js version in #232. I haven't had a chance to test it yet though. Take a look and we can do a quick release for just this update if need be!