Questions about the value of k in the paper

[sleepypineapple]

In the experimental section, ViT-B16 is used as the image encoder.
Generally, ViT-B16 divides the image into 196 patches and adds a [CLS] token, so for the patch embedding matrix of x_i, N=196. However, when using topk() operation for sparsity, the value of k is [50, 250], which is unreasonable.
Did I overlook some details? Hope to get your reply.