Finetune SSD model while keeping model weights for one class

Hi there,
I need to re-train an SSD model pre-trained on COCO in order to make it recognize persons (which is already one of the COCO classes) and another custom object (a specific mug).

I have the labelled images and the .rec files for the “mug” object. I was wondering if it’s possible to fine-tune the model only on the new class, and to keep the original network weights for “person”.

In the .lst files relative to the new “mug” class, the object is identified by the ID 0, so I was thinking maybe I should do something like:

net = gcv.model_zoo.get_model("ssd_512_mobilenet1.0_coco", pretrained=True)
classes = ["mug", "person"]
net.reset_class(classes=['mug', person'], reuse_weights={'person': 'person'})

but it doesn’t seem to be working, I’m getting NaN for the accuracy on “person” when evaluating the training.