The approach we used was to ensure that there were enough stride convolutions such that the final layer would have a grid size of 1.
確定! 回上一頁