The number of prompts to get satisfactory results

Thank you for your impressive work! It is so helpful and elegant! I am recently trying to apply this work on relevant downstream tasks. I note that you get an error rate of **17.38** on CIFAR100 dataset, which is pretty cool. I tried to use 3 prompts as a start point, but failed to get result good enough to beat the one training on the original dataset. I wonder how many prompts should be used to attain comparable results to your experiments (on CIFAR100)😉. Much thanks!