Ace

@ACEaaaaa I think this is due to matconvnet's version. (You can check `/matconvnet-1.0-betaxx/matlab/+dagnn/Loss.m`, whether there is property of `ignoreAverage = false` --- For example, beta20 without this property)~

after I test the astra create_sino. I think it's better to set the default RayTransform cpu version with proj_type='linear' rather than 'line' (However, I find it's strange of parallel_beam_geometry in...

It's may convenient change [parallel_beam_geometry](https://github.com/odlgroup/odl/blob/6a0ace8986514e24d2b1dd24c5650046213d2d00/odl/tomo/geometry/parallel.py#L1471) parameters (det_shape is inconvenient to user to change "detector space" --- due to int form, the detector space can not be 1.0 in most cases)...

The author also provide it on BaiduYun: http://pan.baidu.com/s/1o7MVg6I passward: 4imi

yes, this is load the pretrained vgg features (pretrained from imagenet --- here using the pytorch pretrained model)

It's my fault， please change line34 in dataset.py（zenCompose to Compose）

please see #9 ，and if you want to split by yourself. Follow the step: 1. read all image name (5000 image name) --- can use `os.listdir` 2. shuffle the 0-4999...

Thank you!, You are right, I will modifty it. I reference the meausre from https://github.com/zhimingluo/NLDF/blob/master/evaluate.py . And this paper use beta=sqrt(0.3)

Which the pytorch version and torchvision version you use? maybe you can use 链接: https://pan.baidu.com/s/1XbNjFJm5KmLJTFvU7uL4IQ 提取码: h3er this link to download vgg_weights

我这边Ubuntu下面没有这个问题。你看看是否是因为一些路径的问题呢？比如[Line 32](https://github.com/AceCoooool/DSS-pytorch/blob/159b6cf79e26f105b04df52c9ab054c409111bed/solver.py#L32)里面的路径在windows下应该为\而不是/