Sorry to bother you, but I have some questions about the test set of your model in few-shot scenarios. I noticed that you keep a balanced distribution of labels when testing the few-shot performance, and the len of the test set is the same as COCO val2014. However, after selecting 16 test classes, just about 4k out of 40k images remained in the test phase, so I wonder whether it is valid to set the length of the test set the same as the whole COCO val2014 because it may reuse a lot of images, or I misunderstood your codes :) . I am looking forward to your kindly reply.
Sorry to bother you, but I have some questions about the test set of your model in few-shot scenarios. I noticed that you keep a balanced distribution of labels when testing the few-shot performance, and the len of the test set is the same as COCO val2014. However, after selecting 16 test classes, just about 4k out of 40k images remained in the test phase, so I wonder whether it is valid to set the length of the test set the same as the whole COCO val2014 because it may reuse a lot of images, or I misunderstood your codes :) . I am looking forward to your kindly reply.