Prediction are different in tensorflow.js from the web one

[mehdish3022]

I've trained a model on multiple classes and it worked very well when I tested it on the web but when I imported my model to a local tensorflowjs the predictions were completely wrong.

[sryu1]

I'm guessing it'll be because the js snippet is outdated, if possible, can you just check if it works if you use the keras snippet? You need python though, just to check that the model is fine...

[mehdish3022]

It does work with keras tensorflow but I want to use it with tensorflowjs

[sryu1]

Ok, then it's a problem with the tfjs snippet, I'll try looking into it, however, if you find a way to make it work, it'll be appreciated :)

[mehdish3022]

I converted the model into a tensorflowjs graph model and it kinda works but the predictions are wrong and sometimes inefficient

[sryu1]

It's the problem with the tensorflowjs code snippet, not the model itself, since the keras model is actually originally a tfjs model but converted to keras.

[mehdish3022]

How do I solve this problem?

[sryu1]

Could you try using the snippet from my repo here? https://github.com/sryu1/teachablemachine-community/blob/master/snippets/markdown/image/tensorflowjs

[mehdish3022]

Could you try using the snippet from my repo here? https://github.com/sryu1/teachablemachine-community/blob/master/snippets/markdown/image/tensorflowjs

it works now