Why not use the masked transformers directly in the first two stages?

[xwan0527]

Why use convolutions instead? Since upsampling is already employed to obtain the mask matrix, it seems like transformers could also be used.