Hao Jiang

1. I don't understand either. 2. Maybe you don't use mean() instead of sum() when you calculate the final loss (Flattening Criterion), which in paper loss divided by total pixel...

300 images seems not enough for well training, especially there is no significant domain features like face in your training data. I suggest to use at least thousands images and...

You can use [pSp](https://github.com/eladrich/pixel2style2pixel) to train an image encoder that project image into stylegan2 Z+ latent space or W+ latent space.

I might figured out what happened. In Fig.2, extracted Cxh1xw1 feature map was sent to C-group-convolution to get Cx1x1 representation, and two Cx1x1 represetations were sent to calculate inner product...

Thanks for your early reply, I will try to train model following your advice.

Change "from datasets_USR import ImageDataset" to "from datasets_ISTD import ImageDataset" in "train_Mask-ShadowGAN" and have a try again.

> Maybe you should check "dataroot" option in train_Mask-ShadowGAN is correct like "xxx/ISTD/", and also change path of "self.files_B" in ImageDataset from "train_C" to "train_C_fixed" or something else if you...

> Thank you for your reply.Do you want to make changes like this?  However,the results did not change.  The dataroot option should be your own ISTD dataset's path,...

>  Hello,I'm the partner of Lee, I'm wondering why this situation still happens. I have changed my data path., which is a path without shadow image. Should I add...

>  I have changed the path, but it showed this information, I don't understand what it means, maybe I can never run this code > > > > ...