Questions about table 5

[kaikai23]

Hi,

In your paper table 5, the (G,G,G,G) uses the numbers (79.8%) from PVT paper, which uses absolution positional encoding. However, I suppose the other model variants listed in this table use CPE, so they are not directly comparable. Should the accuracy of (G,G,G,G) with CPE be 81.2% as shown in table 1?

In general, I am interested in knowing if there is a benifit of using global attention in the early layers.

Thanks.