i think this is doing the same thing you’re talking about but a little differently. you can index into the last linear layer, set it to what you want, then train on the new output classes
i think this is doing the same thing you’re talking about but a little differently. you can index into the last linear layer, set it to what you want, then train on the new output classes