Let's see https://github.com/kshmelkov/incremental_detectors/blob/master/network.py#L59-L87

Say self.num_classes = 15, and self.subnet.num_classes = 10, so we use logits = self.logits_for_distillation[:, :cached_classes+1] for distillation.

But wait, [:, :cached_classes+1]? It is 0-6 now, rather than 0-11. Should we use 0-11 here?

def compute_distillation_crossentropy_loss(self):
        cached_classes = self.subnet.num_classes
        ...

def compute_distillation_bbox_loss(self):
        cached_classes = self.subnet.num_classes
        ...