I am using YOLO v8 ultrlytics, pretrained weights on COCO dataset. It is trained on 80 classes. Now I want to add some more classes in my trained model, without losing previous one. For example I have 4 new classes. So I want to my model to detect 84 classes, without re-training of already trained 80 classes.
I am not sure, either it is called incremental learning or something, but that is what I wanna achieve. Can any body please give me some roadmap to do so, if it is possible to do so…
I don’t have a clue but would imagine that you’d merely have to modify the classification layer.
either add another one beside the existing one (two-headed?), or extend the existing one. keep the parameters of the existing 80 classes, and everything else of course.
I can’t say what the training should look like. you didn’t say whether your new classes overlap the existing ones. you’ll have to consider examples (training data) that is new class but not existing class, existing class but not new class, new AND existing class, or no class.