Model always predicting 1 class

Initially, my model had an accuracy of 99.6% on the data, and it was trained last week. I came back today and in the middle of a run, when the model was up to 80%, the accuracy suddenly dropped down to 20%, and from that point on the confusion matrix shows it’s only predicting one class. That one class had the most data points, so I disabled others to make all classes equal in terms of the number of data points, but that hasn’t helped the issue. I’ve increased the learning rate and changed the number of layers and neurons, but I’m unsure what else to do.

Is there any way to get back the version of the model I had before? I’d forgotten to save it in the versioning and I have access to a downloaded C++ build, but I can’t seem to find how to get that back into Edge Impulse.

My project ID is 318543. Thanks for any help!

This is what the confusion matrix looks like now:

Before, it looked like this:


I’ve just fixed the problem - I disabled and enabled all my data and reran feature generation, which appeared to fix the issue.

1 Like

Hello @c24ar,

Thanks for sharing your solution.