Is there a way to make a 2 steps classification training?

I’m training a model to classify 14 different plants using resnet101, but 3 of them are decreasing the accuracy of my model. I’ve trained a model especific for classify these 3 plants and then i got a good precision on it. Is there a way to use this new “3 problematic plants model” in the original 14 plants classification model in order to get a bigger accuracy?

You could try to create a staged classifier, where the first level would classify each sample into the “easy” and “hard” classes, and call the corresponding model.
However, you would of course have to retrain this model and would have to compare it to the standard approach of a single model.

I would guess that a single model with some tweaks (regularization, more aggressive data augmentation) would work better.

1 Like

Thanks! I’ll try the first solution and i will continue to do my best in the process of regularization and data augmentation, even though i still haven’t gotten the results i expected.