question about three training stages

[freeman-1995]

I have some questions about three training stages.

1. I find you train novel class together with base class. However, the feature of base class far greater than novel class. So it suffers a severe label imbalance during low shot training. why it still can work?
2. During base training, I find you use 389 class out of 1000 class, however you train a classifier with 1000 output, it is so weired.
3. When low shot training finished, I got a linear module, but its output is combine novel class with base class, so if I use imagenet1k for base training and use flowers102 for few-shot(one image per class) novel class training, so I can get a model with a 1102 classifier? It's very weird.