Zhaoyi-Yan

@youarenotaloneor Maybe using transformer MHSA to replace shift operation.

You should train a model with random mask. Please set 'mask_type='random' and fixed_mask=false'. Besides, you should also change https://github.com/Zhaoyi-Yan/Shift-Net/blob/master/train.lua#L102-L103 to make the random mask continuous area fit the area of...

https://github.com/Zhaoyi-Yan/Shift-Net/blob/master/test.lua#L106-L108 , as said here, If you have a 'mask' picture, you should set 'fixed_mask=true and mask_type='random' '. The 'mask' picture should be saved as 'v_mask.png'. If you want to...

Hi, you only need to set `opt.dataroot` to the folder of `training set`(when training) or `testing set`(when testing).

In fact, it actually continues training. It loads `latest` model then continues training with `epoch 1`. If you change `which_epoch=2`, then it continues training with `epoch 3`.

In the very begining, I just select several scenes from Places2, besides, you can refer to this issue https://github.com/Zhaoyi-Yan/Shift-Net/issues/12 . For now, I have no much resource to train models...

Email this guy : [email protected]

Hi, Shelley ```python # Adding eps=1e-4 is important here. mm = m.gt(mask_thred/(1.*patch_size**2 + 1e-4)).long() ``` In fact, it equals to ``` # using torch.ge instead of torch.gt mm = m.ge(mask_thred/(1.*patch_size**2)).long()...

Wow, it is truly a bug. However, I am afraid such small masks deverse no shift at all. In fact, the information flow of convolution from the outside to the...

Emmm, in fact, several reasons. 1. Shift-Net is started in early 2017. At that time, no works mention directly how to obtain the mask region in feature map. We have...