fix odd input size issue

[oguzhanilter]

For some input images, the 'x.shape // 8' creates a different size than the tensor that is created by the network itself.

I believe it also solves the closed issue #13

[17648145240]

对于某些输入图像，“x.shape // 8”创建的大小与网络本身创建的张量不同。

我相信它也解决了已关闭的问题#13

Have you implemented the author's original code? If it reproduces, can you share the compressed package with me, thank you very much! My email: [email protected]

[pstemporowski]

Hi, have you tested it? Does it impact any performance of the model?