About visual prompt training

[bugmaker4]

Thanks for your great job! I have a question: I have finished text prompt training and get a good result and use this pre_trained weights to train visual prompt. But during visual prompt training, mAP can not be improved after 19. Is there a plateau in early-stage training？I do not know whether it is converge towards easy instances in early-stage training. I notice that you say it needs 100000 iterations to train. How does the change in mAP occur? Thanks！