Tensor must have a shape comprised of positive integers but got shape [1917,]

[mingsoohea]

I followed cloud.annotations.ai to train a custom model for tensorflow.js. After cloning this project and copy the model_web to the public folder, I run npm install then npm run start.

The local page is showed with an error when detectFromVideoFrame: Error: Tensor must have a shape comprised of positive integers but got shape [1917,]. at g (index.esm.js:17) at index.esm.js:17 at Array.forEach (<anonymous>) at Q (index.esm.js:17) at kn (index.esm.js:17) at Object.Dn [as tensor2d] (index.esm.js:17) at index.esm.js:500 at index.esm.js:17 at t.scopedRun (index.esm.js:17) at t.tidy (index.esm.js:17) at Object.Ke [as tidy] (index.esm.js:17) at runObjectDetectionPrediction (index.esm.js:499) at async Object.detect (index.esm.js:556)

After digging a bit into the source code, I found this strange:

As you see, the scores and boxes have been swapped together. Is it intended?

[mingsoohea]

@bourdakos1

[bourdakos1]

This normally means you didn’t have enough training data

[mingsoohea]

I used 200 images to annotate on cloud.annotations.ai. How many images should I use at least? From the last screenshot, result[1].shape only have 3 components [1, 1917, 1] but the code tries to get result[1].shape[3] Is it make sense?

[bourdakos1]

200 hundred images might be enough, but it more depends on how many labels and how many objects per label

[mingsoohea]

1 label for all 1 object per image only.

[bourdakos1]

Try adding a second label

[Roger92]

Did it help to add more images or an additional lable? Because i have the same error :D

[sebaspv]

I'm having the same error. Did adding more images help? I have 3 labels and 200 images for each of them

[menuRivera]

Does anybody have a working model_web so I can make some tests?