Liu Lei

Ok, i solve the problem. The reason is that basemodel i used resnet101, but the pretrained model is used BNInception.

We don't have any pretrained models trained with MS-SSIM loss.

z is the hyperprior information, which is used to estimate the distribution of the latent representation. The details about z could be found in the [1]. [1] [Learned Image Compression...

New link is here~ https://bhpan.buaa.edu.cn/link/AA3B813D5EB9DB44FEADE19C8C20EAEEA1

We follow the [official tensorflow code](https://github.com/ZhengxueCheng/Learned-Image-Compression-with-GMM-and-Attention), and use the attention at `models/attention.py`.

This problem i also interested when i wrote this code. The base code is copy from [here](https://github.com/liujiaheng/compression), and i also tested torch.distributions.normal.Normal and Laplace in the code. The results shows...

The new link is here~ https://bhpan.buaa.edu.cn:443/link/38850D3409A7334A5A5B5316371218C1

我认为主要是为了让指针有回退能力，比如第二个字符左边右边都符合第三个字符的要求，第三个字符指针先指向左边字符，但是左边的字符周围没有第四个字符，这样就需要将指针回退到第二个字符，然后再将指针指向右边的第三个字符，再进行判断。本人初学者，拙见，望指教。

同问。。