Hi, when training downstream tasks, are the weights of the Vision encoder frozen, or does it participate in full fine-tuning training?