weird results

I tried
python run.py attribute_manipulation -mp checkpoints/CelebA -sp checkpoints/CelebA/test/Smiling --filter_target_attr NOTSmiling -s 1 --branch_idx 0 --n_ref 5 -bs 8 --test_folder examples/aligned

and am getting weird results that lack identity preservation which is shown in the paper

vs

[yingcong]

Could you send me the original image? Let me test on my side.

@yingcong here is the original image

@yingcong were you able to test it?

[yingcong]

My result is similar. I have double-checked this released code with my previous ones, which is used in the paper, but too messy to release.
It seems the optimizer of the decoder matters.
My previous implementation is
self.optim_decoder = torch.optim.Adam(self.decoder.parameters(), lr=1e-4)
while in this verion it is
self.optim_decoder = torch.optim.Adam(self.decoder.parameters(), lr=1e-4,betas=[0.5, 0.999])
After changing it back, the result becomes

Note that the model hasn't converged yet. I am still training it and will update a new one in the dropbox later.

@yingcong ok yeah your results seem to also have the identity preservation problem, looking forward to the updated model to test out.

[yingcong]

@ak9250 I have updated the checkpoints. The results seem better than the previous one. The model keeps training and I'll have a check if the results are further improved after cvpr.

@yingcong thanks, I have tried the updated checkpoints. The results do seem better now although identity is still a problem

[Feiyu-Zhang]

@ak9250 I have updated the checkpoints. The results seem better than the previous one. The model keeps training and I'll have a check if the results are further improved after cvpr.

@ak9250 @yingcong Which is the latest pretrained model? In code ./network/model.py line:115 model_path = 'https://www.dropbox.com/s/8lwmwfs42w5oioi/homo-decoder-8a84d0ce.pth?dl=1' or link in Readme.
Besides, pretrained decoder model's key name in Readme is not same as homo-decoder-8a84d0ce.pth's key name.

[yingcong]

@Feiyu-Zhang The pretrained model is stored in the link of readme, which is https://www.dropbox.com/sh/31dki21jqaifzjj/AACPH11iwBK-38rFy8oKqeraa?dl=0

The homo-decoder-8a84d0ce.pth is trained by simply inverting pretrained VGG feature, which is used to accelerating the training of any attributes.

[yingcong]

@ak9250 It seems that the problem comes from the alignment. If the model is tested on celeba-aligned dataset, then this is not a problem as all images are aligned. As the face alignment tool is not so perfect, it may sometimes lead to less appealing results due to the alignment issue.

I have added small random shift during training, making the model less sensitive to the alignment issue.