Browser Error on Tensorflow pose model Detection

TensorFlow General Discussion
,
1

Hi,
I am receiving this error in the chrome browser
pose-detection:17 Uncaught (in promise) Error: WebGPU readSync is only available for CPU-resident tensors. at Yt.readSync (pose-detection:17:86437) at e.value (engine.js:1221:25) at e.value (tensor.js:411:30) at y (tfjs-converter:17:6090) at tfjs-converter:17:283731 at tfjs-converter:17:283897 at engine.js:469:20 at e.value (engine.js:480:19) at e.value (engine.js:467:17) at ko (tfjs-core:17:95606)

I am not using arraySync
or dataSync anywhere in my code that I know of.
I have tried to set the backend. Same error.

This is for a retrained movenet model.( google collab)

I am using it in the browser. It was working and giving an output and then it started giving this error with no other errors.
the retrained model is tfjs converted saved model to a graph model, though the error is pointing to the model from the pose detection file.
I am using a tf 1D tensor.

I am a beginner and don’t exactly understand everything.
Any help appreciated.
Thanks

3

@J_Conrs,

Welcome to the Tensorflow Forum!

Could you please share with code snippet and steps to reproduce the above error from our end?

Thank you!

5

let output = await tf.tensor1d(x, 'float32'); const dt = output.expandDims() * const po = await model.predict(dt).squeeze(); * or const po = await model.predict(dt);
Nothing is being output once it hits this line *.

I have found this but don’t know if they solved it, what the fix is or if it is relevant:

https://github.com/tensorflow/tfjs/issues/7274

If it is an error in TFJS because of how the model is trained how can I get round it, I am just trying to do a classification task.
Thanks

7

I can confirm I’ve been seeing this error since May 20th, no code changes. I’ll try reverting to an older patch version of TFJS but it’s happening for us too.