How can we do multi labels classification with Edge Impulse?
Hi @naveen, you can train a model that outputs non-categorical outputs (of all sorts) through the Expert mode in the Keras block, but we don’t provide any proper integration in the Studio for it - so Live Classification and Model Testing will probably not work as expected, and we don’t have a nice way of setting multiple output labels on a single data sample either.
Thanks @janjongboom for quick reply! I am thinking how to pass multiple labels to the ingestion service or is there any way to preprocess the training data before passing to the model?
@naveen Not at this point, both ingestion and data acquisition expect a single label per sample at the moment.
There’s a bit of a hack you could do that might approximate this sort of usage for some problems:
Imagine you have classes A and B. You want to determine the following states:
- Input has characteristics of class A but not B
- Input has characteristics of class B but not A
- Input has characteristics of both classes A and B
- Input has characteristics of neither class A or B
You can achieve this in Edge Impulse by organizing your data with one label for each of the above states. If there are more classes, you can add more “mixed” labels (e.g. “input has characteristics of classes A and B, but not C”).
This might be helpful for some problems, as long as there is sufficient training data for each label.