A fixed-size image goes in and a fixed list of scores comes out. That size is built into the network, but sentences come in any length.
Every network so far took a fixed-size input. 224×224×3 in, ten scores out. The architecture itself hard-codes those numbers.
