FOMO Training Fails Due to Output Layer != Output Features

Report bugs
#1

Question/Issue: FOMO Training Fails Due to Object Detection Output Layer Count not equal to Create Impulse Output Feature Count

Project ID: 113205

Context/Use case:Login - Edge Impulse

In Impulse Design > Create Impulse the Output has 2 classes.
In Impulse Design > Object Detection the Output has 3 classes.

This is causing a divide by zero error and the Training fails.

#2

Hi @MMarcial,

Did you have 3 classes before in your dataset? It looks like existing features had 3 different ones. I just recreated your Image and Object Detection blocks and now the number of classes are correct.

Let me know if that works,

Aurelien

#3

Yes, the project once had 3 labels in the dataset. I deleted one of the labels and the project output classes got misaligned between the Create Image block and Object Detection block.

I was hoping

  • there was a button I did not know about that would reconcile the Create Image block and Object Detection block if they get misaligned or
  • there was a quick Keras edit I could make.

So if one removes a complete group of labels, then they must remember to recreate the Image and Object Detection blocks.

1 Like
#4

Thanks for your feedback @MMarcial, I opened an internal ticket on this topic.

Aurelien

1 Like