Image based ML models trained on TeachableMachine V2 show 'NaN' prediction on Chrome and Firefox (Windows OS v 7)

[pmov]

Image based models trained on TeachableMachine V2 do not work on Chrome and Firefox (Windows OS v 7)The predictions show "NaN" only and no numeric prediction is made and the models are trained.No error come at the time of training.What could be the problem?

[adamalbu]

I have the same problem.

[NandiniSrirambhatla]

Even I have the same problem

[irealva]

Hi, we've heard complaints about this on the Pose models as well. Would you be able to provide the specific computer model and browser you were using? We believe it is a hardware specific problem so need to know what specific models you are using.

We've heard this reported from HP Elitebook 840 G-1 using all browsers for now.

[Aendir]

Hi, I am facing this same issue, but on a different platform.
I have an Android app, which creates a Chrome Webview. From inside the webview, upon calling model.predict(), I keep getting NaN as the probability value for all classes.

When I try this from a website that holds the exact same JavaScript code as the webview, it works perfectly fine.
Additionally, when I try to use TensorflowJS to get the classification of an image using the pretrained mobileNet model, it works fine in both the website and the webview.

But if I attempt to use TensorflowJS and the tensorflow model (attribute within the Teachable Machine Model), to get the classification for a TM Image model, I keep getting "NaN".

Model being Used: https://teachablemachine.withgoogle.com/models/LDsdxzd1H/

System Info:
Device: OnePlus 6
Android 10
Gradle version: 4.0.1
Cordova version: 5.3.0-dev
Chrome: 87.0.4280.101

[Aendir]

Hello,
I was able to resolve the issue I was facing.
The problem was that when the model was being loaded, my code was treating the contents of the weights.bin file as a string (When loading from within the webview).
Once I fixed this and treated the contents of weights.bin as an arraybuffer, the prediction started working fine.

Hope this helps someone.

[uhmin99]

I have same problem in pose detection. NaN% problem happens even before I import the model. After I train the model and test it on Teachable machine website, it keep giving me NaN% for every value.

System Info:
Device: LG gram
Processor : Intel i7-4500U
RAM : 8GB
OS : Windows 10 Pro

If I use F12 to see the error message, it says

index.bundle.js:2214 Model disposed so nothing to render with hook
index.bundle.js:5624 Class Equestrian: 25 samples
index.bundle.js:5624 Class Wrong: 21 samples
index.bundle.js:5624 Epochs: 50
index.bundle.js:5624 Learning rate: 0.0001
index.bundle.js:5624 Batch size: 16

[adamalbu]

My computer info (microsoft exel .csv file):
computor info.zip