Jingyun Liang

As clarified in the paper, `z_k` denotes the corresponding latent variable of kernel `k`.

Section 5.2 doesn't specify the scale factors. Did you use the same settings (e.g. kernel width, rotation and noise) for both scale factor 2 and 4?

Which cuda and pytorch versions do you use? I currently use `CUDA11.3+Pytorch1.12`.

We didn't test it on Win10. A linux system should be fine.

`Range()` only receives `int`. You should use `int(your float number)` to turn it to `int`.

The test accuracy is 0.9889. Here is the checkpoint in the 5th epoch ( I just run it with negligible modification, extract it first and resume from this epoch). [model_4.pth.zip](https://github.com/shzygmyx/Matrix-Capsules-pytorch/files/1835678/model_4.pth.zip)

I fixed the bug in DDP by directly hacking the class before creating my own instance as ``` def hacked__len__(self, *arg, **kwarg): return DATASET_SIZE wds.WebDataset.__len__ = hacked__len__ wds.WebLoader.__len__ = hacked__len__...

Problem solved. The authors use Common_Yellowthroat as the 0 class, which is different to most of us: use Common_Yellowthroat as the 200 class. However, I still don't achieve reported accuracy....

The link is still available now. Maybe you should turn to a VPN tool like Lantern. For convenience, I uploaded the resource to [BaiduYun](https://pan.baidu.com/s/1ONQ16J96xqFtTvYnmmK7rQ).

I don't have train.prototxt neither. Even for testing, I failed to reproduce.